[
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "POLARIS",
  "year": 2026,
  "venue": "SoK paper (IACR ePrint 2026/935)",
  "authors": "Al Badawi, Alexandru, Polyakov, Vaikuntanathan",
  "title": "SoK: Private LLM Inference using Approximate Homomorphic Encryption",
  "paper": "https://eprint.iacr.org/2026/935",
  "code": null,
  "family": "CKKS · Model-preserving (N-PRE)",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE",
  "scale": "BERT-Tiny, BERT-Mini",
  "status": "Active SOTA",
  "rel": "Reference CKKS implementation combining the SoK's recommended model-preserving building blocks.",
  "note": "Open-source PoC reference framework; model-preserving, GPU accelerated; 0% accuracy loss on SST-2 subset.",
  "hardware": "A100 GPU",
  "library": "OpenFHE + FIDESlib",
  "packing": "L-ROW",
  "latency": "18 s (BERT-Tiny), 83 s (BERT-Mini), unamortized",
  "open_source": "promised at publication",
  "sok_ref": "paper",
  "id": 0
 },
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "THOR",
  "year": 2025,
  "venue": "IACR ePrint / CCS 2025",
  "authors": "Moon, Omarov, Yoo, Cho et al. (Desilo)",
  "title": "THOR: Secure transformer inference with homomorphic encryption",
  "paper": "https://eprint.iacr.org/2024/1881",
  "code": "https://github.com/Desilo/thor",
  "family": "CKKS · Model-preserving (N-PRE)",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE",
  "scale": "BERT-Base",
  "status": "Active SOTA",
  "rel": "Model-preserving CKKS evaluation of BERT-Base using normalize-and-square Softmax, Goldschmidt division and aSOR.",
  "note": "First model-preserving BERT-Base under CKKS; only 5-6 orders of magnitude slower than FP16 GPU cleartext.",
  "hardware": "A100 GPU",
  "library": "Liberate.FHE",
  "packing": "L-COL",
  "latency": "602 s / 128-token sequence",
  "open_source": "yes",
  "sok_ref": 16,
  "id": 1
 },
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "CEP / EncryptedLLM (GPU-accelerated FHE GPT-2)",
  "year": 2024,
  "venue": "NeurIPS 2024 workshop; ICML 2025 (EncryptedLLM)",
  "authors": "de Castro, Polychroniadou, Escudero",
  "title": "EncryptedLLM: Privacy-Preserving Large Language Model Inference via GPU-Accelerated Fully Homomorphic Encryption",
  "paper": "https://proceedings.mlr.press/v267/de-castro25a.html",
  "code": null,
  "family": "CKKS · Model-preserving (N-PRE)",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE",
  "scale": "GPT-2",
  "status": "Active SOTA",
  "rel": "First open GPU-accelerated CKKS bootstrapping backend for OpenFHE; model-preserving GPT-2 decoding.",
  "note": "~150x speedup over CPU OpenFHE; adopts NEXUS arg max.",
  "hardware": "A100 GPU",
  "library": "Custom OpenFHE GPU backend",
  "packing": "L-ROW",
  "latency": "~100 s per generated token",
  "open_source": "no",
  "sok_ref": 17,
  "id": 2
 },
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "CACHEMIR",
  "year": 2026,
  "venue": "arXiv 2602.11470",
  "authors": "Yu, Zhou, Chen, Soto et al.",
  "title": "Cachemir: Fully Homomorphic Encrypted Inference of Generative Large Language Model with KV Cache",
  "paper": "https://arxiv.org/abs/2602.11470",
  "code": null,
  "family": "CKKS · Model-preserving (N-PRE)",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE",
  "scale": "Llama-3-8B",
  "status": "Active SOTA",
  "rel": "Graph-based bootstrapping placement (extends Orion) and cyclic encodings for multi-head attention.",
  "note": "Model-preserving Llama-3-8B decoding with KV cache; about 4 orders of magnitude slower than cleartext.",
  "hardware": "A100 GPU",
  "library": "Lattigo + Phantom",
  "packing": "L-CYC",
  "latency": "97 s time-per-output-token",
  "open_source": "no",
  "sok_ref": 74,
  "id": 3
 },
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "CERIUM",
  "year": 2025,
  "venue": "arXiv 2512.11269",
  "authors": "Jayashankar, Kim, Sullivan, Zheng et al.",
  "title": "A scalable multi-GPU framework for encrypted large-model inference",
  "paper": "https://arxiv.org/abs/2512.11269",
  "code": null,
  "family": "CKKS · Model-preserving (N-PRE)",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE",
  "scale": "Llama-3-8B, BERT-Base",
  "status": "Active SOTA",
  "rel": "First end-to-end multi-GPU CKKS compiler with kernel fusion; ~14 KB ciphertext memory per model parameter.",
  "note": "Best-known runtimes for model-preserving Llama-3-8B; bootstrapping under 10 ms on multi-GPU cluster.",
  "hardware": "1-8x A100/H100/B200 GPUs",
  "library": "Custom multi-GPU CKKS compiler",
  "packing": "compressed sparse plaintext encoding",
  "latency": "134 s prefill (128 tokens, 8x B200); 66 s BERT-Base on A100",
  "open_source": "planned",
  "sok_ref": 18,
  "id": 4
 },
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "SplitPrompt",
  "year": 2026,
  "venue": "arXiv 2601.18511",
  "authors": "Park, Park, Park, Ahn et al.",
  "title": "Scaling up FHE-based Privacy-Preserving ML: Higher Throughput, Longer Inputs for LLama-3-8B (v1: Scaling up Privacy-Preserving ML: A CKKS Implementation of Llama-2-7B)",
  "paper": "https://arxiv.org/abs/2601.18511",
  "code": null,
  "family": "CKKS · Model-preserving, hybrid public-private prompt split (N-PRE†)",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE (prompt partitioning)",
  "scale": "Llama-2-7B, Llama-3-8B",
  "status": "Active SOTA",
  "rel": "Relaxed threat model: only sensitive tail of the prompt is encrypted, rest processed in the clear.",
  "note": "Normalize-and-square Softmax; Chebyshev SiLU; shows practicality of hybrid prompt partitioning.",
  "hardware": "8x RTX 4090 GPUs",
  "library": "HEaaN2",
  "packing": "L-ALG",
  "latency": "33-34 s per output token; 85 s summarization with 4096 tokens (128 encrypted)",
  "open_source": "no",
  "sok_ref": 78,
  "id": 5
 },
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "ARION",
  "year": 2025,
  "venue": "IACR ePrint 2025/2271",
  "authors": "(see paper)",
  "title": "ARION: Attention-optimized transformer inference on encrypted data",
  "paper": "https://eprint.iacr.org/2025/2271",
  "code": null,
  "family": "CKKS · Model-modifying (N-MOD)",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE",
  "scale": "BERT-Tiny, BERT-Base",
  "status": "Active SOTA",
  "rel": "CPU-only Lattigo implementation with column/diagonal packing.",
  "note": "Strong amortized throughput on CPU; requires architectural modifications.",
  "hardware": "64-core CPU",
  "library": "Lattigo",
  "packing": "L-COL",
  "latency": "9.3 s (Tiny) / 225 s (Base) amortized per sequence, batch 256",
  "open_source": "yes",
  "sok_ref": 13,
  "id": 6
 },
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "ELLMO",
  "year": 2026,
  "venue": "IACR ePrint 2026/198",
  "authors": "Guzelhan, Daksha, Agulló Domingo, Jonatan, Kim et al.",
  "title": "ELLMo: Packing- and Depth-Aware Encrypted Transformer Inference",
  "paper": "https://eprint.iacr.org/2026/198",
  "code": null,
  "family": "CKKS · Model-modifying (N-MOD)",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE",
  "scale": "BERT-Tiny",
  "status": "Active SOTA",
  "rel": "Pre-computed plaintext transforms to minimize depth; delayed LayerNorm; statistical max estimation.",
  "note": "Fastest BERT-Tiny entry in the SoK; sequence length not reported.",
  "hardware": "H200 GPU",
  "library": "OpenFHE + FIDESlib",
  "packing": "L-ALG",
  "latency": "8 s unamortized",
  "open_source": "partial",
  "sok_ref": 77,
  "id": 7
 },
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "RL (Rovida & Leporati)",
  "year": 2024,
  "venue": "IWSPA 2024 (ACM)",
  "authors": "Rovida, Leporati",
  "title": "Transformer-based language models and homomorphic encryption: An intersection with BERT-tiny",
  "paper": "https://doi.org/10.1145/3643651.3659893",
  "code": "https://github.com/narger-ef/EncryptedTransformer",
  "family": "CKKS · Model-modifying (N-MOD)",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE",
  "scale": "BERT-Tiny",
  "status": "Foundational",
  "rel": "Early open-source CKKS BERT-Tiny with precomputed LayerNorm statistics.",
  "note": "Educational baseline; accuracy loss >1.5% / retraining needed.",
  "hardware": "Apple M1 Pro CPU",
  "library": "OpenFHE",
  "packing": "L-ROW",
  "latency": "~672 s unamortized (32 tokens)",
  "open_source": "yes",
  "sok_ref": 11,
  "id": 8
 },
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "TRICYCLE",
  "year": 2025,
  "venue": "IACR ePrint 2025/1200",
  "authors": "(see paper)",
  "title": "TRICYCLE: Private transformer inference with tricyclic encodings",
  "paper": "https://eprint.iacr.org/2025/1200",
  "code": null,
  "family": "CKKS · Model-modifying (N-MOD)",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE",
  "scale": "BERT-Tiny",
  "status": "Active SOTA",
  "rel": "Tricyclic encodings process multiple attention heads per ciphertext; statistical max for Softmax.",
  "note": "CPU-only; closed source.",
  "hardware": "128-core CPU",
  "library": "OpenFHE",
  "packing": "L-CYC",
  "latency": "322 s unamortized (27 tokens)",
  "open_source": "no",
  "sok_ref": 73,
  "id": 9
 },
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "EFLA (Encryption-Friendly LLM Architecture)",
  "year": 2025,
  "venue": "ICLR 2025",
  "authors": "Rho, Kim, Lee et al. (CryptoLab / SNU)",
  "title": "Encryption-friendly LLM architecture",
  "paper": "https://arxiv.org/abs/2410.02486",
  "code": "https://github.com/CryptoLabInc/EFLA",
  "family": "CKKS · Model-modifying (N-MOD)",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE",
  "scale": "BERT-like (fine-tuning + inference)",
  "status": "Active SOTA",
  "rel": "Gaussian-kernel attention replaces Softmax; LoRA fine-tuning under CKKS.",
  "note": "Also addresses encrypted fine-tuning, not just inference.",
  "hardware": "RTX 4090 GPU",
  "library": "HEaaN",
  "packing": "L-ROW",
  "latency": "27 s per layer",
  "open_source": "yes",
  "sok_ref": 80,
  "id": 10
 },
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "PRITRAN",
  "year": 2026,
  "venue": "CGO 2026",
  "authors": "(see paper)",
  "title": "PRITRAN: Privacy-preserving inference for transformer-based language models under FHE",
  "paper": "https://doi.org/10.1109/cgo68049.2026.11395232",
  "code": null,
  "family": "CKKS · Model-modifying (N-MOD)",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE",
  "scale": "BERT-Tiny, BERT-Mini",
  "status": "Niche",
  "rel": "Row packing with precomputed LayerNorm statistics.",
  "note": "No utility metric reported.",
  "hardware": "64-core CPU",
  "library": "OpenFHE",
  "packing": "L-ROW",
  "latency": "398 s (Tiny) / ~2000 s (Mini) unamortized",
  "open_source": "no",
  "sok_ref": 12,
  "id": 11
 },
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "EUSTON",
  "year": 2026,
  "venue": "IACR ePrint 2026/046",
  "authors": "Gao, Fu, Liu, Liu, Luo et al.",
  "title": "Euston: Efficient and User-Friendly Secure Transformer Inference with Non-Interactivity",
  "paper": "https://eprint.iacr.org/2026/046",
  "code": null,
  "family": "CKKS · Model-modifying (N-MOD)",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE",
  "scale": "BERT-Base, GPT-2-1.5B, Llama-3-8B",
  "status": "Active SOTA",
  "rel": "SVD/precomputed transforms to reduce depth; NEXUS arg max.",
  "note": "Scales to Llama-3-8B classification with modest accuracy loss.",
  "hardware": "RTX 6000 GPU",
  "library": "SEAL + Phantom",
  "packing": "L-ALG",
  "latency": "46 s (BERT-Base) / 136 s (GPT-2-1.5B) / 484 s (Llama-3-8B) amortized, batch 32",
  "open_source": "yes",
  "sok_ref": 19,
  "id": 12
 },
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "Powerformer",
  "year": 2025,
  "venue": "ACL 2025 (long)",
  "authors": "Park, Lee, Lee",
  "title": "Powerformer: Efficient and high-accuracy privacy-preserving language model with homomorphic encryption",
  "paper": "https://aclanthology.org/2025.acl-long.543",
  "code": null,
  "family": "CKKS · Model-modifying (N-MOD)",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE",
  "scale": "BERT-Base",
  "status": "Active SOTA",
  "rel": "Batch rectifier power max replaces Softmax; 90% reduction in Softmax cost.",
  "note": "Requires retraining/distillation.",
  "hardware": "A100 GPU",
  "library": "Liberate.FHE",
  "packing": "L-COL",
  "latency": "345 s unamortized (128 tokens)",
  "open_source": "yes",
  "sok_ref": 71,
  "id": 13
 },
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "MOAI",
  "year": 2026,
  "venue": "ICLR 2026; IACR ePrint 2025/991",
  "authors": "Zhang, Wang, Sim, Huang, Zhong, Wang",
  "title": "MOAI: Module-optimizing architecture for non-interactive secure transformer inference",
  "paper": "https://eprint.iacr.org/2025/991",
  "code": null,
  "family": "CKKS · Model-modifying (N-MOD)",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE",
  "scale": "BERT-Base, Llama-3-8B",
  "status": "Active SOTA",
  "rel": "Bootstrapping inside Softmax before division halves required levels (20 to 10).",
  "note": "Depth-minimization placement; constant max values from plaintext profiling.",
  "hardware": "H200 / A100 GPU",
  "library": "SEAL + Phantom",
  "packing": "L-COL",
  "latency": "141 s BERT-Base amortized; 225 s Llama-3-8B decode (8 tokens)",
  "open_source": "yes",
  "sok_ref": 14,
  "id": 14
 },
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "NEXUS",
  "year": 2025,
  "venue": "NDSS 2025",
  "authors": "Zhang, Yang, He et al.",
  "title": "Secure transformer inference made non-interactive",
  "paper": "https://eprint.iacr.org/2024/136",
  "code": "https://github.com/zju-abclab/NEXUS",
  "family": "CKKS · Model-modifying (N-MOD)",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE",
  "scale": "BERT-Base, Llama-3-8B",
  "status": "Active SOTA",
  "rel": "Introduced O(log V) SIMD-folding QuickMax arg max adopted by CEP, MOAI, EUSTON.",
  "note": "First end-to-end non-interactive CKKS transformer; widely used baseline.",
  "hardware": "4x A100 GPUs",
  "library": "SEAL + Phantom",
  "packing": "L-CYC",
  "latency": "37 s BERT-Base amortized (batch 32); 52 s/token Llama-3-8B",
  "open_source": "yes",
  "sok_ref": 15,
  "id": 15
 },
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "STIP",
  "year": 2026,
  "venue": "IACR ePrint 2026/174",
  "authors": "Wang, Chen, Gao, Wang, Liu et al.",
  "title": "STIP: Efficient and secure non-interactive transformer inference via compact packing",
  "paper": "https://eprint.iacr.org/2026/174",
  "code": null,
  "family": "CKKS · Model-modifying (N-MOD)",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE",
  "scale": "BERT-Base, GPT-2-1.5B, Llama-3-8B",
  "status": "Active SOTA",
  "rel": "Uses both real and imaginary slots; Gaussian-kernel attention; delayed LayerNorm.",
  "note": "Compact packing; model modifications required.",
  "hardware": "A100 GPU",
  "library": "SEAL + Phantom",
  "packing": "L-CYC (real + imaginary slots)",
  "latency": "64 s (BERT-Base) / 142 s (GPT-2-1.5B) / 407 s (Llama-3-8B) amortized",
  "open_source": "partial",
  "sok_ref": 20,
  "id": 16
 },
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "ENSI",
  "year": 2025,
  "venue": "arXiv",
  "authors": "He, Wang, Gao, Luo, Liu, Fu",
  "title": "ENSI: Efficient non-interactive secure inference for large language models",
  "paper": "https://arxiv.org/abs/2509.09424",
  "code": null,
  "family": "CKKS · Model-modifying (N-MOD)",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE",
  "scale": "Llama-3-700M (BitNet b1.58)",
  "status": "Active SOTA",
  "rel": "BitNet ternary weights replace multiplications with additions; sigmoid attention.",
  "note": "Model-modifying; demonstrates 1.58-bit LLMs under FHE.",
  "hardware": "A100 GPU",
  "library": "OpenFHE + Phantom",
  "packing": "L-ALG (ternary weights)",
  "latency": "958 s TTFT (batch 32, 2048 tokens)",
  "open_source": "yes",
  "sok_ref": 75,
  "id": 17
 },
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "Polyformer",
  "year": 2024,
  "venue": "arXiv",
  "authors": "Zimerman, Baruch, Drucker et al. (IBM)",
  "title": "Converting transformers to polynomial form for secure inference over homomorphic encryption",
  "paper": "https://arxiv.org/abs/2311.08610",
  "code": null,
  "family": "CKKS · Model-modifying (N-MOD)",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE",
  "scale": "BERT-6L",
  "status": "Niche",
  "rel": "100% polynomial transformer via scaled-sigma attention and range-minimization training.",
  "note": "Requires retraining; HELayers tile tensors.",
  "hardware": "A100 GPU",
  "library": "HELayers",
  "packing": "tile tensors",
  "latency": "305 s (128 tokens next-word prediction)",
  "open_source": "no",
  "sok_ref": 82,
  "id": 18
 },
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "SecretVoice 1.0",
  "year": 2026,
  "venue": "(see paper)",
  "authors": "(see paper)",
  "title": "SecretVoice1.0: Non-interactive secure transformer inference",
  "paper": null,
  "code": null,
  "family": "CKKS · Model-modifying (N-MOD)",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE",
  "scale": "Original Transformer encoder-decoder",
  "status": "Niche",
  "rel": "Fuses linear layers into StC-first bootstrapping.",
  "note": "Operation fusion trades memory for latency; no utility metric.",
  "hardware": "1-core CPU",
  "library": "Microsoft SEAL",
  "packing": "L-ROW",
  "latency": "1488 s per encoder block (32 tokens)",
  "open_source": "no",
  "sok_ref": 69,
  "id": 19
 },
 {
  "group": "0. CKKS LLM frameworks surveyed in the SoK",
  "name": "RZW (encrypted LoRA split model)",
  "year": 2025,
  "venue": "arXiv 2025",
  "authors": "Zhang, Zheng, Bao",
  "title": "Practical secure inference algorithm for a fine-tuned large language model based on fully homomorphic encryption",
  "paper": "https://arxiv.org/abs/2501.01672",
  "code": null,
  "family": "CKKS · Hybrid model partitioning",
  "arith": "Approximate real (CKKS)",
  "interaction": "non-interactive FHE on LoRA modules only",
  "scale": "fine-tuned LLM (LoRA adapters)",
  "status": "Niche",
  "rel": "Only encrypted LoRA adapters are evaluated server-side; base model runs at client in the clear.",
  "note": "Protects proprietary fine-tuning rather than full prompt privacy.",
  "hardware": "(see paper)",
  "library": "(see paper)",
  "packing": "n/a",
  "latency": "1.61 s/token (reported)",
  "open_source": "no",
  "sok_ref": 90,
  "id": 20
 },
 {
  "group": "2. CKKS core and variants",
  "name": "CKKS (HEAAN)",
  "year": 2017,
  "venue": "ASIACRYPT 2017",
  "authors": "Cheon, Kim, Kim, Song",
  "title": "Homomorphic Encryption for Arithmetic of Approximate Numbers",
  "paper": "https://eprint.iacr.org/2016/421",
  "code": "https://github.com/snucrypto/HEAAN",
  "family": "RLWE, leveled + bootstrapping",
  "arith": "Approximate fixed-point real/complex, SIMD",
  "interaction": "Non-interactive",
  "scale": "Llama-3-8B (via frameworks below)",
  "status": "Foundational",
  "rel": "Baseline scheme of the SoK.",
  "note": "De facto scheme for private LLM inference. Native SIMD over N/2 slots, but non-linear functions need polynomial approximation.",
  "id": 21
 },
 {
  "group": "2. CKKS core and variants",
  "name": "RNS-CKKS (full RNS variant)",
  "year": 2018,
  "venue": "SAC 2018",
  "authors": "Cheon, Han, Kim, Kim, Song",
  "title": "A Full RNS Variant of Approximate Homomorphic Encryption",
  "paper": "https://eprint.iacr.org/2018/931",
  "code": null,
  "family": "RLWE, RNS/CRT",
  "arith": "Approximate real/complex",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Foundational",
  "rel": "The form of CKKS implemented in every modern library.",
  "note": "Removes multi-precision arithmetic by using a chain of word-sized primes. Enables GPU/ASIC acceleration.",
  "id": 22
 },
 {
  "group": "2. CKKS core and variants",
  "name": "Conjugate-invariant CKKS",
  "year": 2018,
  "venue": "ICISC 2018",
  "authors": "Kim, Song",
  "title": "Approximate Homomorphic Encryption over the Conjugate-invariant Ring",
  "paper": "https://eprint.iacr.org/2018/952",
  "code": null,
  "family": "RLWE over real subring",
  "arith": "Approximate real-only, N slots",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Niche",
  "rel": "Real-only variant that doubles usable real slots.",
  "note": "Packs N real values instead of N/2 complex ones. Bootstrapping is more expensive, so it is rarely used.",
  "id": 23
 },
 {
  "group": "2. CKKS core and variants",
  "name": "Reduced-error CKKS (FLEXIBLEAUTO)",
  "year": 2022,
  "venue": "CT-RSA 2022",
  "authors": "Kim, Papadimitriou, Polyakov",
  "title": "Approximate Homomorphic Encryption with Reduced Approximation Error",
  "paper": "https://eprint.iacr.org/2020/1118",
  "code": "https://github.com/openfheorg/openfhe-development",
  "family": "RNS-CKKS with exact rescaling",
  "arith": "Approximate real/complex, higher precision",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Widely used",
  "rel": "Default precision-management mode in OpenFHE.",
  "note": "Removes rescaling error from prime approximation. Gains several bits of precision at the same parameters.",
  "id": 24
 },
 {
  "group": "2. CKKS core and variants",
  "name": "IND-CPA-D attack on CKKS",
  "year": 2021,
  "venue": "EUROCRYPT 2021",
  "authors": "Li, Micciancio",
  "title": "On the Security of Homomorphic Encryption on Approximate Numbers",
  "paper": "https://eprint.iacr.org/2020/1533",
  "code": null,
  "family": "Security analysis",
  "arith": "-",
  "interaction": "-",
  "scale": "-",
  "status": "Foundational",
  "rel": "Shows CKKS decryptions leak the secret key if shared.",
  "note": "Motivates the SoK's assumption that the client never shares decrypted outputs with the server.",
  "id": 25
 },
 {
  "group": "2. CKKS core and variants",
  "name": "IND-CPA-D secure CKKS (noise flooding / DP)",
  "year": 2022,
  "venue": "CRYPTO 2022",
  "authors": "Li, Micciancio, Schultz, Sorrell",
  "title": "Securing Approximate Homomorphic Encryption Using Differential Privacy",
  "paper": "https://eprint.iacr.org/2022/816",
  "code": "https://github.com/openfheorg/openfhe-development",
  "family": "CKKS + noise flooding",
  "arith": "Approximate real/complex",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Active",
  "rel": "Hardened CKKS decryption for threshold or shared-output settings.",
  "note": "Implemented as NOISE_FLOODING_DECRYPT in OpenFHE. Costs some precision.",
  "id": 26
 },
 {
  "group": "2. CKKS core and variants",
  "name": "Multi-key CKKS",
  "year": 2019,
  "venue": "CCS 2019",
  "authors": "Chen, Dai, Kim, Song",
  "title": "Efficient Multi-Key Homomorphic Encryption with Packed Ciphertexts with Application to Oblivious Neural Network Inference",
  "paper": "https://eprint.iacr.org/2019/524",
  "code": "https://github.com/SNUCP/MKHE-KKLSS",
  "family": "Multi-key RLWE",
  "arith": "Approximate (CKKS) and exact (BFV)",
  "interaction": "Non-interactive, multiple key owners",
  "scale": "Small CNN",
  "status": "Niche",
  "rel": "Lets client and model owner encrypt under different keys.",
  "note": "Enables protecting both inputs and weights. Ciphertext size grows with the number of parties.",
  "id": 27
 },
 {
  "group": "2. CKKS core and variants",
  "name": "Threshold / multiparty CKKS (MHE)",
  "year": 2021,
  "venue": "PETS 2021",
  "authors": "Mouchet, Troncoso-Pastoriza, Bossuat, Hubaux",
  "title": "Multiparty Homomorphic Encryption from Ring-Learning-with-Errors",
  "paper": "https://eprint.iacr.org/2020/304",
  "code": "https://github.com/tuneinsight/lattigo",
  "family": "Threshold RLWE (CKKS/BFV)",
  "arith": "Approximate or exact",
  "interaction": "Interactive key generation and decryption",
  "scale": "-",
  "status": "Active",
  "rel": "Distributes the CKKS secret key across parties.",
  "note": "Used in Lattigo and OpenFHE for federated and collaborative analytics. Fits double-blind model and data privacy.",
  "id": 28
 },
 {
  "group": "2. CKKS core and variants",
  "name": "Bit-CKKS / bootstrapping bits with CKKS",
  "year": 2024,
  "venue": "EUROCRYPT 2024",
  "authors": "Bae, Cheon, Kim, Stehlé",
  "title": "Bootstrapping Bits with CKKS",
  "paper": "https://eprint.iacr.org/2024/767",
  "code": null,
  "family": "CKKS used for binary/small-integer data",
  "arith": "Bits and small integers via CKKS",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Active SOTA",
  "rel": "Turns CKKS into a fast boolean or integer FHE engine.",
  "note": "Faster amortized bit bootstrapping than TFHE/FHEW for large batches. Bridges approximate and exact worlds.",
  "id": 29
 },
 {
  "group": "2. CKKS core and variants",
  "name": "BLEACH (discrete CKKS)",
  "year": 2024,
  "venue": "J. Cryptology 2024",
  "authors": "Drucker, Moshkowich, Pelleg, Shaul",
  "title": "BLEACH: Cleaning Errors in Discrete Computations over CKKS",
  "paper": "https://eprint.iacr.org/2022/1298",
  "code": null,
  "family": "CKKS with error cleaning",
  "arith": "Discrete values via CKKS",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Niche",
  "rel": "Runs exact integer/boolean logic inside CKKS.",
  "note": "Periodically cleans CKKS noise with polynomial maps. Avoids scheme switching.",
  "id": 30
 },
 {
  "group": "2. CKKS core and variants",
  "name": "General functional bootstrapping in CKKS",
  "year": 2025,
  "venue": "CRYPTO 2025",
  "authors": "Alexandru, Kim, Polyakov",
  "title": "General Functional Bootstrapping using CKKS",
  "paper": "https://eprint.iacr.org/2024/1623",
  "code": "https://github.com/openfheorg/openfhe-development",
  "family": "CKKS functional bootstrapping",
  "arith": "Lookup tables on small integers",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Active SOTA",
  "rel": "Evaluates LUTs during CKKS bootstrapping, like TFHE PBS.",
  "note": "The SoK lists it as a robustness direction for piecewise non-linear approximation in LLMs.",
  "id": 31
 },
 {
  "group": "4. Alternative FHE schemes",
  "name": "Gentry-Lee matrix FHE (GL scheme)",
  "year": 2025,
  "venue": "IACR ePrint 2025/1935",
  "authors": "Gentry, Lee",
  "title": "Fully Homomorphic Encryption for Matrix Arithmetic",
  "paper": "https://eprint.iacr.org/2025/1935",
  "code": null,
  "family": "New lattice FHE, native matrices",
  "arith": "Approximate real matrices",
  "interaction": "Non-interactive",
  "scale": "None yet",
  "status": "Active SOTA (emerging)",
  "rel": "Same arithmetic as CKKS but natively on matrices.",
  "note": "Removes most rotations and rotation keys in linear algebra. The SoK expects its CKKS recommendations to transfer directly. 2026 follow-ups now cover GL bootstrapping, key size and a first GPU Transformer framework (Terrazzo).",
  "id": 32
 },
 {
  "group": "3. CKKS bootstrapping and operation advances",
  "name": "First CKKS bootstrapping",
  "year": 2018,
  "venue": "EUROCRYPT 2018",
  "authors": "Cheon, Han, Kim, Kim, Song",
  "title": "Bootstrapping for Approximate Homomorphic Encryption",
  "paper": "https://eprint.iacr.org/2018/153",
  "code": null,
  "family": "CKKS bootstrapping",
  "arith": "-",
  "interaction": "-",
  "scale": "-",
  "status": "Foundational",
  "rel": "Makes CKKS fully homomorphic.",
  "note": "Introduced CoeffToSlot, EvalMod and SlotToCoeff. Every LLM framework relies on its descendants.",
  "id": 33
 },
 {
  "group": "3. CKKS bootstrapping and operation advances",
  "name": "Improved CKKS bootstrapping (Chen-Chillotti-Song)",
  "year": 2019,
  "venue": "EUROCRYPT 2019",
  "authors": "Chen, Chillotti, Song",
  "title": "Improved Bootstrapping for Approximate Homomorphic Encryption",
  "paper": "https://eprint.iacr.org/2018/1043",
  "code": null,
  "family": "CKKS bootstrapping",
  "arith": "-",
  "interaction": "-",
  "scale": "-",
  "status": "Foundational",
  "rel": "Faster linear transforms and modular reduction.",
  "note": "Basis of Microsoft SEAL-era bootstrapping implementations.",
  "id": 34
 },
 {
  "group": "3. CKKS bootstrapping and operation advances",
  "name": "Non-sparse-key bootstrapping (Bossuat et al.)",
  "year": 2021,
  "venue": "EUROCRYPT 2021",
  "authors": "Bossuat, Mouchet, Troncoso-Pastoriza, Hubaux",
  "title": "Efficient Bootstrapping for Approximate Homomorphic Encryption with Non-Sparse Keys",
  "paper": "https://eprint.iacr.org/2020/1203",
  "code": "https://github.com/tuneinsight/lattigo",
  "family": "CKKS bootstrapping",
  "arith": "-",
  "interaction": "-",
  "scale": "-",
  "status": "Widely used",
  "rel": "Secure, precise bootstrapping implemented in Lattigo.",
  "note": "Removes the sparse-key assumption that weakened security.",
  "id": 35
 },
 {
  "group": "3. CKKS bootstrapping and operation advances",
  "name": "High-precision bootstrapping (minimax + inverse sine)",
  "year": 2021,
  "venue": "EUROCRYPT 2021",
  "authors": "Lee, Lee, Lee, Kim, No",
  "title": "High-Precision Bootstrapping of RNS-CKKS Homomorphic Encryption Using Optimal Minimax Polynomial Approximation and Inverse Sine Function",
  "paper": "https://eprint.iacr.org/2020/552",
  "code": null,
  "family": "CKKS bootstrapping",
  "arith": "-",
  "interaction": "-",
  "scale": "-",
  "status": "Foundational",
  "rel": "Higher bootstrapping precision.",
  "note": "Optimal minimax EvalMod approximations, important for deep LLM circuits.",
  "id": 36
 },
 {
  "group": "3. CKKS bootstrapping and operation advances",
  "name": "META-BTS",
  "year": 2022,
  "venue": "CCS 2022",
  "authors": "Bae, Cheon, Cho, Kim, Kim",
  "title": "META-BTS: Bootstrapping Precision Beyond the Limit",
  "paper": "https://eprint.iacr.org/2022/1167",
  "code": null,
  "family": "CKKS bootstrapping",
  "arith": "-",
  "interaction": "-",
  "scale": "-",
  "status": "Active",
  "rel": "Iterative bootstrapping that doubles precision.",
  "note": "Trades extra bootstrapping for precision. Useful when deep models accumulate error.",
  "id": 37
 },
 {
  "group": "3. CKKS bootstrapping and operation advances",
  "name": "Hybrid key switching",
  "year": 2020,
  "venue": "CT-RSA 2020",
  "authors": "Han, Ki",
  "title": "Better Bootstrapping for Approximate Homomorphic Encryption",
  "paper": "https://eprint.iacr.org/2019/688",
  "code": null,
  "family": "Key switching / bootstrapping",
  "arith": "-",
  "interaction": "-",
  "scale": "-",
  "status": "Widely used",
  "rel": "The key-switching method used in most libraries.",
  "note": "Balances key size and noise. Key switching dominates rotation and bootstrapping cost.",
  "id": 38
 },
 {
  "group": "3. CKKS bootstrapping and operation advances",
  "name": "KLSS key switching",
  "year": 2023,
  "venue": "CRYPTO 2023",
  "authors": "Kim, Lee, Seo, Song",
  "title": "Accelerating HE Operations from Key Decomposition Technique",
  "paper": "https://eprint.iacr.org/2023/413",
  "code": null,
  "family": "Key switching",
  "arith": "-",
  "interaction": "-",
  "scale": "-",
  "status": "Active SOTA",
  "rel": "Asymptotically faster key switching for large parameters.",
  "note": "Reduces NTT count in key switching. Relevant to rotation-heavy linear layers.",
  "id": 39
 },
 {
  "group": "3. CKKS bootstrapping and operation advances",
  "name": "Fused plaintext-ciphertext matmul and bootstrapping",
  "year": 2024,
  "venue": "CRYPTO 2024",
  "authors": "Bae, Cheon, Hanrot, Park, Stehlé",
  "title": "Plaintext-Ciphertext Matrix Multiplication and FHE Bootstrapping: Fast and Fused",
  "paper": "https://eprint.iacr.org/2024/1284",
  "code": null,
  "family": "CKKS bootstrapping + matmul",
  "arith": "-",
  "interaction": "-",
  "scale": "-",
  "status": "Active SOTA",
  "rel": "Folds matrix multiplication into bootstrapping transforms.",
  "note": "Related to the operation fusion used by SecretVoice in the SoK.",
  "id": 40
 },
 {
  "group": "3. CKKS bootstrapping and operation advances",
  "name": "Grafting (CKKS modulus engineering)",
  "year": 2024,
  "venue": "IACR ePrint 2024/1014",
  "authors": "Cheon, Hanrot, Kim, Stehlé",
  "title": "Grafting: Decoupled Scale Factors and Modulus in RNS-CKKS",
  "paper": "https://eprint.iacr.org/2024/1014",
  "code": null,
  "family": "CKKS parameterization",
  "arith": "-",
  "interaction": "-",
  "scale": "-",
  "status": "Active SOTA",
  "rel": "Decouples scale factors from RNS primes.",
  "note": "Uses word-sized primes with arbitrary scale. Reduces wasted modulus and speeds GPU implementations.",
  "id": 41
 },
 {
  "group": "3. CKKS bootstrapping and operation advances",
  "name": "SHIP bootstrapping",
  "year": 2025,
  "venue": "EUROCRYPT 2025",
  "authors": "Cheon, Hanrot, Kim, Stehlé",
  "title": "SHIP: A Shallow and Highly Parallelizable CKKS Bootstrapping Algorithm",
  "paper": "https://eprint.iacr.org/2025/784",
  "code": null,
  "family": "CKKS bootstrapping (blind-rotation inspired)",
  "arith": "-",
  "interaction": "-",
  "scale": "-",
  "status": "Active SOTA",
  "rel": "Very shallow bootstrapping consuming few levels.",
  "note": "SoK notes its evaluation keys are larger than classical bootstrapping, so it needs refinement for LLMs.",
  "id": 42
 },
 {
  "group": "3. CKKS bootstrapping and operation advances",
  "name": "PaCo bootstrapping",
  "year": 2025,
  "venue": "IACR ePrint",
  "authors": "Coron, Seuré",
  "title": "PaCo: Bootstrapping for CKKS via Partial CoeffToSlot",
  "paper": "https://eprint.iacr.org/2025/886",
  "code": null,
  "family": "CKKS bootstrapping",
  "arith": "-",
  "interaction": "-",
  "scale": "-",
  "status": "Active SOTA",
  "rel": "Small-key bootstrapping (about 0.25 GB keys).",
  "note": "SoK highlights it as appealing for LLM inference because key memory is a key bottleneck.",
  "id": 43
 },
 {
  "group": "3. CKKS bootstrapping and operation advances",
  "name": "Large-interval evaluation",
  "year": 2022,
  "venue": "IEEE Access 2022",
  "authors": "Cheon, Kim, Park",
  "title": "Efficient Homomorphic Evaluation on Large Intervals",
  "paper": "https://eprint.iacr.org/2022/280",
  "code": null,
  "family": "CKKS function evaluation",
  "arith": "-",
  "interaction": "-",
  "scale": "-",
  "status": "Active",
  "rel": "Extends approximation domains for non-linear functions.",
  "note": "SoK cites it for robustness when encrypted activations fall outside the fitted interval.",
  "id": 44
 },
 {
  "group": "3. CKKS bootstrapping and operation advances",
  "name": "Fast homomorphic Softmax",
  "year": 2024,
  "venue": "CCS 2024",
  "authors": "Cho, Hanrot, Kim, Park, Stehlé",
  "title": "Fast and Accurate Homomorphic Softmax Evaluation",
  "paper": "https://arxiv.org/abs/2410.11184",
  "code": null,
  "family": "CKKS non-linear evaluation",
  "arith": "-",
  "interaction": "-",
  "scale": "-",
  "status": "Active SOTA",
  "rel": "Normalize-and-square Softmax used by SplitPrompt and THOR.",
  "note": "SoK's recommended model-preserving Softmax algorithm.",
  "id": 45
 },
 {
  "group": "4. Alternative FHE schemes",
  "name": "Gentry's FHE",
  "year": 2009,
  "venue": "STOC 2009",
  "authors": "Gentry",
  "title": "Fully Homomorphic Encryption Using Ideal Lattices",
  "paper": "https://doi.org/10.1145/1536414.1536440",
  "code": "https://crypto.stanford.edu/craig/craig-thesis.pdf",
  "family": "Ideal lattices + bootstrapping",
  "arith": "Bits",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Foundational",
  "rel": "First FHE construction, ancestor of all schemes.",
  "note": "Introduced bootstrapping. Impractical today but defines the paradigm.",
  "id": 46
 },
 {
  "group": "4. Alternative FHE schemes",
  "name": "Paillier (partially HE)",
  "year": 1999,
  "venue": "EUROCRYPT 1999",
  "authors": "Paillier",
  "title": "Public-Key Cryptosystems Based on Composite Degree Residuosity Classes",
  "paper": "https://doi.org/10.1007/3-540-48910-X_16",
  "code": "https://github.com/data61/python-paillier",
  "family": "Additive partially homomorphic",
  "arith": "Exact integers, additions only",
  "interaction": "Non-interactive for linear ops",
  "scale": "Linear models",
  "status": "Foundational",
  "rel": "Supports only addition, no multiplication of ciphertexts.",
  "note": "Used in early PPML and federated aggregation. Cannot evaluate Transformers alone.",
  "id": 47
 },
 {
  "group": "4. Alternative FHE schemes",
  "name": "BGV",
  "year": 2012,
  "venue": "ITCS 2012",
  "authors": "Brakerski, Gentry, Vaikuntanathan",
  "title": "(Leveled) Fully Homomorphic Encryption without Bootstrapping",
  "paper": "https://eprint.iacr.org/2011/277",
  "code": "https://github.com/homenc/HElib",
  "family": "RLWE, leveled, modulus switching",
  "arith": "Exact integers mod p, SIMD",
  "interaction": "Non-interactive",
  "scale": "CNNs, quantized models",
  "status": "Widely used",
  "rel": "Exact counterpart of CKKS with plaintext modulus p.",
  "note": "Exact arithmetic suits quantized integer LLMs. Non-linear functions and bootstrapping are costlier than in CKKS.",
  "id": 48
 },
 {
  "group": "4. Alternative FHE schemes",
  "name": "BFV",
  "year": 2012,
  "venue": "CRYPTO 2012 / ePrint 2012",
  "authors": "Brakerski; Fan, Vercauteren",
  "title": "Somewhat Practical Fully Homomorphic Encryption",
  "paper": "https://eprint.iacr.org/2012/144",
  "code": "https://github.com/microsoft/SEAL",
  "family": "RLWE, scale-invariant",
  "arith": "Exact integers mod p, SIMD",
  "interaction": "Non-interactive",
  "scale": "Linear layers in HE+MPC hybrids",
  "status": "Widely used",
  "rel": "Exact scheme widely used for the linear layers of hybrid protocols.",
  "note": "Basis of Cheetah, BOLT, BumbleBee and Nimbus linear layers. No native approximate arithmetic.",
  "id": 49
 },
 {
  "group": "4. Alternative FHE schemes",
  "name": "Brakerski scale-invariant FHE",
  "year": 2012,
  "venue": "CRYPTO 2012",
  "authors": "Brakerski",
  "title": "Fully Homomorphic Encryption without Modulus Switching from Classical GapSVP",
  "paper": "https://eprint.iacr.org/2012/078",
  "code": null,
  "family": "LWE, scale-invariant",
  "arith": "Exact integers",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Foundational",
  "rel": "Theoretical basis of BFV.",
  "note": "Removes modulus switching through scale invariance.",
  "id": 50
 },
 {
  "group": "4. Alternative FHE schemes",
  "name": "Revisited BGV/BFV for finite fields",
  "year": 2021,
  "venue": "ASIACRYPT 2021",
  "authors": "Kim, Polyakov, Zucca",
  "title": "Revisiting Homomorphic Encryption Schemes for Finite Fields",
  "paper": "https://eprint.iacr.org/2021/204",
  "code": "https://github.com/openfheorg/openfhe-development",
  "family": "RNS BGV/BFV",
  "arith": "Exact integers",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Widely used",
  "rel": "Modern BGV/BFV variants implemented in OpenFHE.",
  "note": "Lower noise and faster RNS BFV/BGV multiplication.",
  "id": 51
 },
 {
  "group": "4. Alternative FHE schemes",
  "name": "GSW",
  "year": 2013,
  "venue": "CRYPTO 2013",
  "authors": "Gentry, Sahai, Waters",
  "title": "Homomorphic Encryption from Learning with Errors: Conceptually-Simpler, Asymptotically-Faster, Attribute-Based",
  "paper": "https://eprint.iacr.org/2013/340",
  "code": null,
  "family": "LWE, matrix ciphertexts",
  "arith": "Bits / small integers",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Foundational",
  "rel": "Basis of FHEW/TFHE blind rotation.",
  "note": "Slow noise growth for multiplication by bits. Used as the accumulator in modern bootstrapping.",
  "id": 52
 },
 {
  "group": "4. Alternative FHE schemes",
  "name": "FHEW",
  "year": 2015,
  "venue": "EUROCRYPT 2015",
  "authors": "Ducas, Micciancio",
  "title": "FHEW: Bootstrapping Homomorphic Encryption in Less Than a Second",
  "paper": "https://eprint.iacr.org/2014/816",
  "code": "https://github.com/openfheorg/openfhe-development",
  "family": "LWE + RGSW blind rotation",
  "arith": "Bits / small integers via LUT",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Foundational",
  "rel": "Fast per-gate bootstrapping with no SIMD.",
  "note": "Exact LUT evaluation of arbitrary functions. Available in OpenFHE, including scheme switching with CKKS.",
  "id": 53
 },
 {
  "group": "4. Alternative FHE schemes",
  "name": "TFHE (CGGI)",
  "year": 2016,
  "venue": "ASIACRYPT 2016 / J. Cryptology 2020",
  "authors": "Chillotti, Gama, Georgieva, Izabachène",
  "title": "TFHE: Fast Fully Homomorphic Encryption over the Torus",
  "paper": "https://eprint.iacr.org/2018/421",
  "code": "https://github.com/zama-ai/tfhe-rs",
  "family": "Torus LWE + programmable bootstrapping",
  "arith": "Bits / small integers via LUT",
  "interaction": "Non-interactive",
  "scale": "Small CNNs, tiny Transformers",
  "status": "Widely used",
  "rel": "Exact LUT evaluation of any non-linearity; no approximation error.",
  "note": "Main CKKS alternative for non-linear functions. Lacks CKKS-style SIMD, so large matrix products are slow.",
  "id": 54
 },
 {
  "group": "4. Alternative FHE schemes",
  "name": "TFHE programmable bootstrapping for DNNs",
  "year": 2021,
  "venue": "CSCML 2021",
  "authors": "Chillotti, Joye, Paillier",
  "title": "Programmable Bootstrapping Enables Efficient Homomorphic Inference of Deep Neural Networks",
  "paper": "https://eprint.iacr.org/2021/091",
  "code": "https://github.com/zama-ai/concrete-ml",
  "family": "TFHE PBS",
  "arith": "Quantized integers",
  "interaction": "Non-interactive",
  "scale": "Deep quantized NNs",
  "status": "Active",
  "rel": "Evaluates activations exactly with lookup tables.",
  "note": "Basis of Zama Concrete ML, which supports quantized LLM fragments such as GPT-2 attention heads.",
  "id": 55
 },
 {
  "group": "4. Alternative FHE schemes",
  "name": "TFHE floating point",
  "year": 2025,
  "venue": "IACR ePrint 2025/257",
  "authors": "(see paper)",
  "title": "TFHE Gets Real: An Efficient and Flexible Homomorphic Floating-Point Arithmetic",
  "paper": "https://eprint.iacr.org/2025/257",
  "code": null,
  "family": "TFHE",
  "arith": "Floating-point emulation",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Active SOTA",
  "rel": "True floating point instead of CKKS fixed point.",
  "note": "SoK cites it as evidence that FHE floating point is now feasible but costly.",
  "id": 56
 },
 {
  "group": "4. Alternative FHE schemes",
  "name": "FINAL (NTRU + LWE)",
  "year": 2022,
  "venue": "ASIACRYPT 2022",
  "authors": "Bonte, Iliashenko, Park, Pereira, Smart",
  "title": "FINAL: Faster FHE Instantiated with NTRU and LWE",
  "paper": "https://eprint.iacr.org/2022/074",
  "code": "https://github.com/KULeuven-COSIC/FINAL",
  "family": "NTRU-based bootstrapping",
  "arith": "Bits",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Niche",
  "rel": "NTRU alternative to TFHE gate bootstrapping.",
  "note": "Smaller keys and faster bootstrapping than TFHE at the time, with NTRU security caveats.",
  "id": 57
 },
 {
  "group": "4. Alternative FHE schemes",
  "name": "High-precision integer HE (CLPX)",
  "year": 2018,
  "venue": "CT-RSA 2018",
  "authors": "Chen, Laine, Player, Xia",
  "title": "High-Precision Arithmetic in Homomorphic Encryption",
  "paper": "https://eprint.iacr.org/2017/809",
  "code": null,
  "family": "BFV variant",
  "arith": "Exact high-precision rationals",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Niche",
  "rel": "Exact alternative to approximate CKKS for real numbers.",
  "note": "Encodes large integers and fixed-point values exactly. Superseded by CKKS for ML.",
  "id": 58
 },
 {
  "group": "4. Alternative FHE schemes",
  "name": "CHIMERA (scheme switching)",
  "year": 2020,
  "venue": "J. Math. Cryptology 2020",
  "authors": "Boura, Gama, Georgieva, Jetchev",
  "title": "CHIMERA: Combining Ring-LWE-based Fully Homomorphic Encryption Schemes",
  "paper": "https://eprint.iacr.org/2018/758",
  "code": null,
  "family": "Switching TFHE/BFV/CKKS",
  "arith": "Mixed",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Foundational",
  "rel": "Uses CKKS for linear parts and TFHE for non-linear parts.",
  "note": "First framework for switching between ring-LWE schemes.",
  "id": 59
 },
 {
  "group": "4. Alternative FHE schemes",
  "name": "PEGASUS (CKKS + FHEW switching)",
  "year": 2021,
  "venue": "IEEE S&P 2021",
  "authors": "Lu, Huang, Hong, Ma, Qu",
  "title": "PEGASUS: Bridging Polynomial and Non-polynomial Evaluations in Homomorphic Encryption",
  "paper": "https://eprint.iacr.org/2020/1606",
  "code": "https://github.com/Alibaba-Gemini-Lab/OpenPEGASUS",
  "family": "CKKS <-> FHEW switching",
  "arith": "Approximate + LUT",
  "interaction": "Non-interactive",
  "scale": "CNN, decision trees",
  "status": "Active",
  "rel": "CKKS for SIMD linear algebra, FHEW LUTs for exact non-linearities.",
  "note": "Scheme switching avoids high-degree polynomials but switching cost is high. OpenFHE implements a variant.",
  "id": 60
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "Gazelle",
  "year": 2018,
  "venue": "USENIX Security 2018",
  "authors": "Juvekar, Vaikuntanathan, Chandrakasan",
  "title": "GAZELLE: A Low Latency Framework for Secure Neural Network Inference",
  "paper": "https://arxiv.org/abs/1801.05507",
  "code": null,
  "family": "HE (BFV) + garbled circuits",
  "arith": "Exact integers",
  "interaction": "Interactive 2PC",
  "scale": "CNNs",
  "status": "Foundational",
  "rel": "Uses HE only for linear layers and garbled circuits for non-linear ones.",
  "note": "Template for most hybrid protocols. Requires many communication rounds.",
  "id": 61
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "Delphi",
  "year": 2020,
  "venue": "USENIX Security 2020",
  "authors": "Mishra, Lehmkuhl, Srinivasan, Zheng, Popa",
  "title": "Delphi: A Cryptographic Inference Service for Neural Networks",
  "paper": "https://eprint.iacr.org/2020/050",
  "code": "https://github.com/mc2-project/delphi",
  "family": "HE preprocessing + GC",
  "arith": "Exact integers",
  "interaction": "Interactive 2PC",
  "scale": "CNNs",
  "status": "Foundational",
  "rel": "Moves HE to an offline preprocessing phase.",
  "note": "Fast online phase. Also co-designs activations with polynomial approximation.",
  "id": 62
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "CrypTFlow2",
  "year": 2020,
  "venue": "CCS 2020",
  "authors": "Rathee, Rathee, Kumar, Chandran, Gupta, Rastogi, Sharma",
  "title": "CrypTFlow2: Practical 2-Party Secure Inference",
  "paper": "https://eprint.iacr.org/2020/1002",
  "code": "https://github.com/mpc-msri/EzPC",
  "family": "OT + HE 2PC",
  "arith": "Exact fixed point",
  "interaction": "Interactive 2PC",
  "scale": "ResNet-50, DenseNet",
  "status": "Foundational",
  "rel": "Exact 2PC comparison and truncation protocols.",
  "note": "Part of Microsoft EzPC. Basis for SIRNN and later LLM protocols.",
  "id": 63
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "Cheetah",
  "year": 2022,
  "venue": "USENIX Security 2022",
  "authors": "Huang, Lu, Hong, Ding",
  "title": "Cheetah: Lean and Fast Secure Two-Party Deep Neural Network Inference",
  "paper": "https://eprint.iacr.org/2022/207",
  "code": "https://github.com/Alibaba-Gemini-Lab/OpenCheetah",
  "family": "RLWE (BFV) + silent OT",
  "arith": "Exact fixed point",
  "interaction": "Interactive 2PC",
  "scale": "ResNet-50",
  "status": "Foundational",
  "rel": "Coefficient-encoded HE matmul without rotations plus OT non-linearities.",
  "note": "Basis of Iron, BOLT and BumbleBee.",
  "id": 64
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "THE-X",
  "year": 2022,
  "venue": "ACL Findings 2022",
  "authors": "Chen, Bao, Huang et al.",
  "title": "THE-X: Privacy-Preserving Transformer Inference with Homomorphic Encryption",
  "paper": "https://arxiv.org/abs/2206.00216",
  "code": null,
  "family": "HE (SEAL CKKS) + client-side non-linear",
  "arith": "Approximate",
  "interaction": "Interactive (client computes non-linear steps)",
  "scale": "BERT-Tiny/Base",
  "status": "Foundational",
  "rel": "Early HE Transformer that offloads non-linear steps to the client.",
  "note": "Replaces Softmax/GELU and returns some intermediate values to the client, which leaks information.",
  "id": 65
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "Iron",
  "year": 2022,
  "venue": "NeurIPS 2022",
  "authors": "Hao, Li, Chen, Xing, Xu, Zhang",
  "title": "Iron: Private Inference on Transformers",
  "paper": "https://proceedings.neurips.cc/paper/2022/hash/64e2449d74f84e5b1a5c96ba7b3d308e-Abstract.html",
  "code": null,
  "family": "HE (BFV) + OT/SS 2PC",
  "arith": "Exact fixed point",
  "interaction": "Interactive 2PC",
  "scale": "BERT-Base",
  "status": "Foundational",
  "rel": "First efficient 2PC Transformer with HE matmul.",
  "note": "Model-preserving but gigabytes of communication per inference.",
  "id": 66
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "MPCFormer",
  "year": 2023,
  "venue": "ICLR 2023",
  "authors": "Li, Shao, Wang et al.",
  "title": "MPCFormer: Fast, Performant and Private Transformer Inference with MPC",
  "paper": "https://arxiv.org/abs/2211.01452",
  "code": "https://github.com/DachengLi1/MPCFormer",
  "family": "MPC (CrypTen) + distillation",
  "arith": "Fixed point secret sharing",
  "interaction": "Interactive 2PC/3PC",
  "scale": "BERT-Base/Large",
  "status": "Active",
  "rel": "Replaces Softmax/GELU with quadratic forms and distills.",
  "note": "Model-modifying, like the SoK's N-MOD class but for MPC.",
  "id": 67
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "PUMA",
  "year": 2023,
  "venue": "arXiv 2023",
  "authors": "Dong, Lu, Zheng, Wu et al.",
  "title": "PUMA: Secure Inference of LLaMA-7B in Five Minutes",
  "paper": "https://arxiv.org/abs/2307.12533",
  "code": "https://github.com/secretflow/spu",
  "family": "Replicated secret sharing 3PC",
  "arith": "Fixed point",
  "interaction": "Interactive 3PC (honest majority)",
  "scale": "LLaMA-7B",
  "status": "Active SOTA",
  "rel": "Model-preserving 3PC; first private LLaMA-7B inference.",
  "note": "About five minutes per token for LLaMA-7B. Needs three non-colluding servers.",
  "id": 68
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "CipherGPT",
  "year": 2023,
  "venue": "IACR ePrint 2023/1147",
  "authors": "Hou, Liu, Li et al.",
  "title": "CipherGPT: Secure Two-Party GPT Inference",
  "paper": "https://eprint.iacr.org/2023/1147",
  "code": null,
  "family": "VOLE + HE 2PC",
  "arith": "Fixed point",
  "interaction": "Interactive 2PC",
  "scale": "GPT-2",
  "status": "Active",
  "rel": "First 2PC GPT generation framework.",
  "note": "Model-preserving but heavy communication, about 25 GB per token.",
  "id": 69
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "SIGMA",
  "year": 2024,
  "venue": "PETS 2024",
  "authors": "Gupta, Kumar, Chandran et al.",
  "title": "SIGMA: Secure GPT Inference with Function Secret Sharing",
  "paper": "https://eprint.iacr.org/2023/1269",
  "code": "https://github.com/mpc-msri/EzPC",
  "family": "Function secret sharing on GPU",
  "arith": "Fixed point",
  "interaction": "Interactive 2PC with dealer",
  "scale": "Llama-2-13B",
  "status": "Active SOTA",
  "rel": "GPU-accelerated FSS with a trusted dealer for preprocessing.",
  "note": "Seconds to tens of seconds per token for 13B models. Requires offline correlated randomness.",
  "id": 70
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "BOLT",
  "year": 2024,
  "venue": "IEEE S&P 2024",
  "authors": "Pang, Zhu, Lou, Wang",
  "title": "BOLT: Privacy-Preserving, Accurate and Efficient Inference for Transformers",
  "paper": "https://eprint.iacr.org/2023/1893",
  "code": "https://github.com/Clive2312/BOLT",
  "family": "HE (BFV) + OT 2PC",
  "arith": "Fixed point",
  "interaction": "Interactive 2PC",
  "scale": "BERT-Base",
  "status": "Active SOTA",
  "rel": "Hybrid with compact HE packing and word elimination.",
  "note": "About 10x less communication than Iron. Still interactive.",
  "id": 71
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "BumbleBee",
  "year": 2025,
  "venue": "NDSS 2025",
  "authors": "Lu, Huang, Gu, Li et al.",
  "title": "BumbleBee: Secure Two-party Inference Framework for Large Transformers",
  "paper": "https://eprint.iacr.org/2023/1678",
  "code": "https://github.com/AntCPLab/OpenBumbleBee",
  "family": "RLWE HE + OT 2PC",
  "arith": "Fixed point",
  "interaction": "Interactive 2PC",
  "scale": "LLaMA-7B",
  "status": "Active SOTA",
  "rel": "Hybrid 2PC with efficient HE matmul and accurate activations.",
  "note": "Runs LLaMA-7B in about 14 minutes per token. Model-preserving.",
  "id": 72
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "Nimbus",
  "year": 2024,
  "venue": "NeurIPS 2024",
  "authors": "Li, Yang, Tan, Lu et al.",
  "title": "Nimbus: Secure and Efficient Two-Party Inference for Transformers",
  "paper": "https://arxiv.org/abs/2411.15707",
  "code": "https://github.com/secretflow/spu",
  "family": "HE + 2PC, outer-product encoding",
  "arith": "Fixed point",
  "interaction": "Interactive 2PC",
  "scale": "BERT-Base",
  "status": "Active SOTA",
  "rel": "Client-side outer-product HE encoding for linear layers.",
  "note": "Uses low-degree piecewise polynomials for non-linear layers.",
  "id": 73
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "SecFormer",
  "year": 2024,
  "venue": "ACL Findings 2024",
  "authors": "Luo, Zhang, Zhang, Zhang et al.",
  "title": "SecFormer: Fast and Accurate Privacy-Preserving Inference for Transformer Models via SMPC",
  "paper": "https://arxiv.org/abs/2401.00793",
  "code": "https://github.com/jinglong696/SecFormer",
  "family": "Secret sharing (CrypTen)",
  "arith": "Fixed point",
  "interaction": "Interactive 2PC/3PC",
  "scale": "BERT-Base/Large",
  "status": "Active",
  "rel": "Fourier-series GELU and Goldschmidt LayerNorm.",
  "note": "Balances distillation and model fidelity.",
  "id": 74
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "Ditto",
  "year": 2024,
  "venue": "ICML 2024",
  "authors": "Wu, Fang, Zheng, Ma et al.",
  "title": "Ditto: Quantization-aware Secure Inference of Transformers upon MPC",
  "paper": "https://arxiv.org/abs/2405.05525",
  "code": "https://github.com/secretflow/spu",
  "family": "3PC + quantization",
  "arith": "Mixed bit-width fixed point",
  "interaction": "Interactive 3PC",
  "scale": "BERT, GPT-2",
  "status": "Active",
  "rel": "Quantization-aware MPC with type conversion.",
  "note": "Speeds MPC by lowering bit-widths, a direction not yet explored in CKKS.",
  "id": 75
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "MPC-minimized secure LLM inference (Marill)",
  "year": 2024,
  "venue": "arXiv 2024",
  "authors": "Rathee, Li, Stoica, Zhang et al.",
  "title": "MPC-Minimized Secure LLM Inference",
  "paper": "https://arxiv.org/abs/2408.03561",
  "code": null,
  "family": "MPC + LLM fine-tuning co-design",
  "arith": "Fixed point",
  "interaction": "Interactive 2PC/3PC",
  "scale": "Llama-2-7B",
  "status": "Active SOTA",
  "rel": "Fine-tunes LLMs so fewer layers need MPC.",
  "note": "Model-modifying. Combines public pre-trained layers with private ones.",
  "id": 76
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "Centaur",
  "year": 2025,
  "venue": "ACL Findings 2025",
  "authors": "Luo, Chen, Zhang, Liu et al.",
  "title": "Centaur: Bridging the Impossible Trinity of Privacy, Efficiency, and Performance in Privacy-Preserving Transformer Inference",
  "paper": "https://arxiv.org/abs/2412.10652",
  "code": null,
  "family": "Random permutation + 2PC",
  "arith": "Fixed point",
  "interaction": "Interactive 2PC",
  "scale": "BERT, GPT-2",
  "status": "Active",
  "rel": "Permutes weights so non-linear steps run in plaintext.",
  "note": "Weaker privacy than full cryptography but much faster.",
  "id": 77
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "CrypTen",
  "year": 2021,
  "venue": "NeurIPS 2021",
  "authors": "Knott, Venkataraman, Hannun et al.",
  "title": "CrypTen: Secure Multi-Party Computation Meets Machine Learning",
  "paper": "https://arxiv.org/abs/2109.00984",
  "code": "https://github.com/facebookresearch/CrypTen",
  "family": "Secret sharing (PyTorch)",
  "arith": "Fixed point",
  "interaction": "Interactive multi-party",
  "scale": "BERT (via MPCFormer)",
  "status": "Widely used",
  "rel": "PyTorch-native MPC framework.",
  "note": "Used as the backend of MPCFormer and SecFormer. Supports GPUs.",
  "id": 78
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "SecretFlow-SPU",
  "year": 2023,
  "venue": "USENIX ATC 2023",
  "authors": "Ma, Zheng, Fang et al.",
  "title": "SecretFlow-SPU: A Performant and User-Friendly Framework for Privacy-Preserving Machine Learning",
  "paper": "https://www.usenix.org/conference/atc23/presentation/ma",
  "code": "https://github.com/secretflow/spu",
  "family": "MPC compiler (ABY3, Cheetah, SEMI2K)",
  "arith": "Fixed point",
  "interaction": "Interactive 2PC/3PC",
  "scale": "LLaMA-7B (PUMA)",
  "status": "Widely used",
  "rel": "JAX-to-MPC compiler used by PUMA, BumbleBee, Ditto, Nimbus.",
  "note": "Ant Group open-source stack. Comparable role to OpenFHE for the MPC world.",
  "id": 79
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "MP-SPDZ",
  "year": 2020,
  "venue": "CCS 2020",
  "authors": "Keller",
  "title": "MP-SPDZ: A Versatile Framework for Multi-Party Computation",
  "paper": "https://eprint.iacr.org/2020/521",
  "code": "https://github.com/data61/MP-SPDZ",
  "family": "Generic MPC (30+ protocols)",
  "arith": "Integers / fixed point",
  "interaction": "Interactive multi-party",
  "scale": "Research models",
  "status": "Widely used",
  "rel": "Benchmark framework for many MPC protocols, including malicious security.",
  "note": "Useful when a stronger threat model than semi-honest is needed.",
  "id": 80
 },
 {
  "group": "6. Trusted hardware, statistical and integrity-only alternatives",
  "name": "NVIDIA Confidential Computing (H100/Blackwell)",
  "year": 2023,
  "venue": "Product",
  "authors": "NVIDIA",
  "title": "NVIDIA Confidential Computing",
  "paper": "https://www.nvidia.com/en-us/data-center/solutions/confidential-computing/",
  "code": null,
  "family": "GPU TEE",
  "arith": "Native FP16/BF16/FP8",
  "interaction": "Trusted hardware",
  "scale": "Llama-3-70B and larger",
  "status": "Commercial",
  "rel": "Near-cleartext speed with trust in hardware vendor.",
  "note": "Low single-digit overhead in many LLM workloads, but vulnerable to side channels and requires attestation trust.",
  "id": 81
 },
 {
  "group": "6. Trusted hardware, statistical and integrity-only alternatives",
  "name": "Confidential computing on H100: benchmark",
  "year": 2024,
  "venue": "arXiv 2024",
  "authors": "Zhu, Yin, Deng, Almeida et al.",
  "title": "Confidential Computing on NVIDIA Hopper GPUs: A Performance Benchmark Study",
  "paper": "https://arxiv.org/abs/2409.03992",
  "code": null,
  "family": "GPU TEE evaluation",
  "arith": "Native floating point",
  "interaction": "Trusted hardware",
  "scale": "Llama-3-70B",
  "status": "Active",
  "rel": "Quantifies TEE overhead for LLM inference.",
  "note": "Reports under 7% overhead for most LLM queries, versus about 10^4x for CKKS.",
  "id": 82
 },
 {
  "group": "6. Trusted hardware, statistical and integrity-only alternatives",
  "name": "Apple Private Cloud Compute",
  "year": 2024,
  "venue": "Apple Security Research",
  "authors": "Apple",
  "title": "Private Cloud Compute: A new frontier for AI privacy in the cloud",
  "paper": "https://security.apple.com/blog/private-cloud-compute/",
  "code": null,
  "family": "Custom silicon TEE + attestation",
  "arith": "Native floating point",
  "interaction": "Trusted hardware",
  "scale": "Production foundation models",
  "status": "Commercial",
  "rel": "Production private LLM serving based on hardware trust and transparency logs.",
  "note": "No cryptographic guarantee against a compromised operator with hardware access.",
  "id": 83
 },
 {
  "group": "6. Trusted hardware, statistical and integrity-only alternatives",
  "name": "Intel TDX / AMD SEV-SNP",
  "year": 2021,
  "venue": "Product",
  "authors": "Intel; AMD",
  "title": "Confidential VMs (Intel TDX, AMD SEV-SNP)",
  "paper": "https://www.intel.com/content/www/us/en/developer/tools/trust-domain-extensions/overview.html",
  "code": "https://www.amd.com/en/developer/sev.html",
  "family": "CPU TEE (confidential VM)",
  "arith": "Native floating point",
  "interaction": "Trusted hardware",
  "scale": "CPU LLM inference",
  "status": "Commercial",
  "rel": "CPU-side TEE often paired with GPU TEEs.",
  "note": "Lower throughput than GPU TEE. Known side-channel history.",
  "id": 84
 },
 {
  "group": "6. Trusted hardware, statistical and integrity-only alternatives",
  "name": "Slalom",
  "year": 2019,
  "venue": "ICLR 2019",
  "authors": "Tramèr, Boneh",
  "title": "Slalom: Fast, Verifiable and Private Execution of Neural Networks in Trusted Hardware",
  "paper": "https://arxiv.org/abs/1806.03287",
  "code": "https://github.com/ftramer/slalom",
  "family": "SGX TEE + outsourced GPU linear layers",
  "arith": "Native / field arithmetic",
  "interaction": "Trusted hardware",
  "scale": "VGG, MobileNet",
  "status": "Foundational",
  "rel": "Hybrid TEE plus untrusted GPU with blinding.",
  "note": "Template for TEE-GPU split designs.",
  "id": 85
 },
 {
  "group": "6. Trusted hardware, statistical and integrity-only alternatives",
  "name": "Split-N-Denoise (SnD)",
  "year": 2024,
  "venue": "ICML 2024",
  "authors": "Mai, Yan, Huang et al.",
  "title": "Split-and-Denoise: Protect Large Language Model Inference with Local Differential Privacy",
  "paper": "https://arxiv.org/abs/2310.09130",
  "code": null,
  "family": "Local DP on embeddings + client denoiser",
  "arith": "Native",
  "interaction": "Non-interactive (single round)",
  "scale": "Llama-2, GPT-2",
  "status": "Active",
  "rel": "Formal DP but only statistical privacy.",
  "note": "Orders of magnitude cheaper than CKKS, but leaks information within the privacy budget.",
  "id": 86
 },
 {
  "group": "6. Trusted hardware, statistical and integrity-only alternatives",
  "name": "InferDPT",
  "year": 2023,
  "venue": "arXiv 2023",
  "authors": "Tong, Chen, Zhang, Qi et al.",
  "title": "InferDPT: Privacy-Preserving Inference for Closed-box Large Language Model",
  "paper": "https://arxiv.org/abs/2310.12214",
  "code": null,
  "family": "Text-level DP perturbation",
  "arith": "Native",
  "interaction": "Non-interactive",
  "scale": "GPT-4 (black box)",
  "status": "Active",
  "rel": "Works with closed APIs without server cooperation.",
  "note": "Utility drops at strong privacy levels. No cryptographic protection.",
  "id": 87
 },
 {
  "group": "6. Trusted hardware, statistical and integrity-only alternatives",
  "name": "PrivateLoRA",
  "year": 2023,
  "venue": "arXiv 2023",
  "authors": "Wang, Lin, Zeng, Zhang et al.",
  "title": "PrivateLoRA for Efficient Privacy Preserving LLM",
  "paper": "https://arxiv.org/abs/2311.14030",
  "code": null,
  "family": "Split inference (edge-cloud)",
  "arith": "Native",
  "interaction": "Interactive split",
  "scale": "Llama-2-13B",
  "status": "Niche",
  "rel": "Keeps private low-rank layers on device.",
  "note": "Activations are exposed to the cloud, so privacy is heuristic.",
  "id": 88
 },
 {
  "group": "6. Trusted hardware, statistical and integrity-only alternatives",
  "name": "zkLLM",
  "year": 2024,
  "venue": "CCS 2024",
  "authors": "Sun, Li, Zhang",
  "title": "zkLLM: Zero Knowledge Proofs for Large Language Models",
  "paper": "https://arxiv.org/abs/2404.16109",
  "code": "https://github.com/jvhs0706/zkllm-ccs2024",
  "family": "Zero-knowledge proofs",
  "arith": "Field arithmetic",
  "interaction": "Non-interactive proof",
  "scale": "OPT-13B",
  "status": "Active SOTA",
  "rel": "Proves correct inference; does not hide the user's prompt from the server.",
  "note": "Complements CKKS: integrity rather than input confidentiality.",
  "id": 89
 },
 {
  "group": "6. Trusted hardware, statistical and integrity-only alternatives",
  "name": "EZKL",
  "year": 2023,
  "venue": "Open-source toolkit",
  "authors": "Zkonduit",
  "title": "EZKL: zkSNARKs for ONNX models",
  "paper": "https://docs.ezkl.xyz/",
  "code": "https://github.com/zkonduit/ezkl",
  "family": "Halo2 zkSNARK",
  "arith": "Field arithmetic",
  "interaction": "Non-interactive proof",
  "scale": "Small models",
  "status": "Active",
  "rel": "Verifiable ML inference toolkit.",
  "note": "Proof generation is slow for large models.",
  "id": 90
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "OpenFHE",
  "year": 2022,
  "venue": "WAHC 2022",
  "authors": "Al Badawi, Bates, Bergamaschi et al.",
  "title": "OpenFHE: Open-Source Fully Homomorphic Encryption Library",
  "paper": "https://eprint.iacr.org/2022/915",
  "code": "https://github.com/openfheorg/openfhe-development",
  "family": "CKKS, BGV, BFV, FHEW, TFHE, scheme switching, threshold",
  "arith": "C++ (Python wrapper)",
  "interaction": "Library",
  "scale": "Used by RL, PRITRAN, TRICYCLE, POLARIS, ELLMO, ENSI",
  "status": "Widely used",
  "rel": "Most complete open CKKS library; successor of PALISADE.",
  "note": "BSD-2 license. CPU-only core, with GPU back-ends from CEP and FIDESlib.",
  "id": 91
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "Microsoft SEAL",
  "year": 2017,
  "venue": "Microsoft Research",
  "authors": "Chen, Laine, Player et al.",
  "title": "Simple Encrypted Arithmetic Library - SEAL",
  "paper": null,
  "code": "https://github.com/microsoft/SEAL",
  "family": "CKKS, BFV, BGV",
  "arith": "C++",
  "interaction": "Library",
  "scale": "Used by NEXUS, MOAI, EUSTON, STIP, SecretVoice",
  "status": "Widely used",
  "rel": "Popular CKKS library without built-in CKKS bootstrapping.",
  "note": "MIT license. Frameworks pair it with Phantom on GPU.",
  "id": 92
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "Lattigo",
  "year": 2020,
  "venue": "Tune Insight / EPFL",
  "authors": "Mouchet, Bossuat et al.",
  "title": "Lattigo: A Multiparty Homomorphic Encryption Library in Go",
  "paper": null,
  "code": "https://github.com/tuneinsight/lattigo",
  "family": "CKKS, BGV, BFV, multiparty",
  "arith": "Go",
  "interaction": "Library",
  "scale": "Used by ARION, CACHEMIR",
  "status": "Widely used",
  "rel": "Go library with strong bootstrapping and multiparty support.",
  "note": "Apache-2.0. CPU only.",
  "id": 93
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "HEaaN / HEaaN2",
  "year": 2018,
  "venue": "CryptoLab",
  "authors": "CryptoLab",
  "title": "HEaaN library",
  "paper": "https://heaan.it/",
  "code": "https://heaan.it/",
  "family": "CKKS",
  "arith": "C++/Python",
  "interaction": "Library (commercial)",
  "scale": "Used by EFLA, SplitPrompt",
  "status": "Commercial",
  "rel": "Reference CKKS implementation by the CKKS authors with GPU support.",
  "note": "Free for non-commercial use. Closed source.",
  "id": 94
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "HElib",
  "year": 2014,
  "venue": "IBM",
  "authors": "Halevi, Shoup",
  "title": "Algorithms in HElib",
  "paper": "https://eprint.iacr.org/2014/106",
  "code": "https://github.com/homenc/HElib",
  "family": "BGV, CKKS",
  "arith": "C++",
  "interaction": "Library",
  "scale": "-",
  "status": "Niche",
  "rel": "First major HE library; BGV focus.",
  "note": "Apache-2.0. Maintenance mode.",
  "id": 95
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "IBM HELayers",
  "year": 2023,
  "venue": "PETS 2023",
  "authors": "Aharoni, Adir, Baruch et al.",
  "title": "HeLayers: A Tile Tensors Framework for Large Neural Networks on Encrypted Data",
  "paper": "https://arxiv.org/abs/2011.01805",
  "code": "https://ibm.github.io/helayers/",
  "family": "CKKS, BGV (via SEAL/HElib/HEaaN)",
  "arith": "C++/Python",
  "interaction": "ML-on-HE toolkit",
  "scale": "Used by Polyformer",
  "status": "Active",
  "rel": "Tile-tensor packing abstraction for neural networks.",
  "note": "Free community edition. Not open source.",
  "id": 96
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "Phantom",
  "year": 2024,
  "venue": "IEEE TDSC 2024",
  "authors": "Yang, Shen, Zhang et al.",
  "title": "Phantom: A CUDA-Accelerated Word-Wise Homomorphic Encryption Library",
  "paper": "https://eprint.iacr.org/2023/049",
  "code": "https://github.com/encryptorion-lab/phantom-fhe",
  "family": "CKKS, BFV, BGV",
  "arith": "CUDA C++",
  "interaction": "GPU library",
  "scale": "Used by NEXUS, MOAI, EUSTON, STIP, ENSI, CACHEMIR",
  "status": "Widely used",
  "rel": "Open GPU CKKS library used by most SoK GPU frameworks.",
  "note": "GPL-3.0. Bootstrapping added by downstream frameworks.",
  "id": 97
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "FIDESlib",
  "year": 2025,
  "venue": "arXiv 2025",
  "authors": "Agulló-Domingo, Vera-López, Guzelhan, Daksha et al.",
  "title": "FIDESlib: A Fully-Fledged Open-Source FHE Library for Efficient CKKS on GPUs",
  "paper": "https://arxiv.org/abs/2507.04775",
  "code": "https://github.com/CAPS-UMU/FIDESlib",
  "family": "CKKS with bootstrapping",
  "arith": "CUDA C++",
  "interaction": "GPU library",
  "scale": "Used by ELLMO, POLARIS",
  "status": "Active SOTA",
  "rel": "OpenFHE-compatible GPU CKKS server library.",
  "note": "Full GPU bootstrapping. Interoperates with OpenFHE client keys.",
  "id": 98
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "Liberate.FHE",
  "year": 2023,
  "venue": "Desilo",
  "authors": "Desilo",
  "title": "Liberate.FHE: A New FHE Library for Bridging the Gap between Theory and Practice",
  "paper": null,
  "code": "https://github.com/Desilo/liberate-fhe",
  "family": "CKKS",
  "arith": "Python/CUDA",
  "interaction": "GPU library",
  "scale": "Used by THOR, Powerformer",
  "status": "Deprecated",
  "rel": "GPU CKKS library in Python.",
  "note": "SoK notes it is deprecated.",
  "id": 99
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "Cheddar",
  "year": 2024,
  "venue": "arXiv 2024",
  "authors": "Choi, Kim, Ahn",
  "title": "Cheddar: A Swift Fully Homomorphic Encryption Library Designed for GPU Architectures",
  "paper": "https://arxiv.org/abs/2407.13055",
  "code": null,
  "family": "CKKS",
  "arith": "CUDA",
  "interaction": "GPU library",
  "scale": "-",
  "status": "Active SOTA",
  "rel": "Among the fastest reported GPU CKKS bootstrapping implementations.",
  "note": "Uses 32-bit word arithmetic. Code availability limited.",
  "id": 100
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "100x faster GPU bootstrapping",
  "year": 2021,
  "venue": "TCHES 2021",
  "authors": "Jung, Kim, Ahn, Cheon, Lee",
  "title": "Over 100x Faster Bootstrapping in Fully Homomorphic Encryption through Memory-centric Optimization with GPUs",
  "paper": "https://eprint.iacr.org/2021/508",
  "code": null,
  "family": "CKKS",
  "arith": "CUDA",
  "interaction": "GPU technique",
  "scale": "-",
  "status": "Foundational",
  "rel": "Showed CKKS bootstrapping is memory-bound on GPUs.",
  "note": "Kernel fusion ideas reused by Phantom, Cheddar and FIDESlib.",
  "id": 101
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "TFHE-rs",
  "year": 2022,
  "venue": "Zama",
  "authors": "Zama",
  "title": "TFHE-rs: Pure Rust implementation of TFHE",
  "paper": null,
  "code": "https://github.com/zama-ai/tfhe-rs",
  "family": "TFHE (booleans, integers)",
  "arith": "Rust, CUDA",
  "interaction": "Library",
  "scale": "-",
  "status": "Widely used",
  "rel": "Leading TFHE implementation with GPU support.",
  "note": "BSD-3-Clause-Clear, free for non-commercial use.",
  "id": 102
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "Concrete ML",
  "year": 2022,
  "venue": "Zama",
  "authors": "Zama",
  "title": "Concrete ML: Privacy-Preserving ML with FHE",
  "paper": null,
  "code": "https://github.com/zama-ai/concrete-ml",
  "family": "TFHE",
  "arith": "Python",
  "interaction": "ML-on-HE toolkit",
  "scale": "-",
  "status": "Active",
  "rel": "Compiles quantized scikit-learn and PyTorch models to TFHE.",
  "note": "Has LLM examples such as GPT-2 attention and LoRA fine-tuning. Practical mainly for small models.",
  "id": 103
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "Orion",
  "year": 2025,
  "venue": "ASPLOS 2025",
  "authors": "Ebel, Garimella, Reagen",
  "title": "Orion: A Fully Homomorphic Encryption Framework for Deep Learning",
  "paper": "https://arxiv.org/abs/2311.03470",
  "code": "https://github.com/baahl-nyu/orion",
  "family": "CKKS (Lattigo back-end)",
  "arith": "Python/Go",
  "interaction": "Compiler/framework",
  "scale": "CACHEMIR extends its bootstrapping placement",
  "status": "Active SOTA",
  "rel": "PyTorch-to-CKKS compiler with automatic bootstrapping placement.",
  "note": "Shortest-path bootstrapping placement reused by CACHEMIR.",
  "id": 104
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "HEIR",
  "year": 2023,
  "venue": "Google",
  "authors": "Google HEIR team",
  "title": "HEIR: Homomorphic Encryption Intermediate Representation",
  "paper": "https://heir.dev/",
  "code": "https://github.com/google/heir",
  "family": "CKKS, BGV, BFV, CGGI (via OpenFHE, Lattigo, TFHE-rs)",
  "arith": "MLIR/C++",
  "interaction": "Compiler/framework",
  "scale": "-",
  "status": "Active SOTA",
  "rel": "MLIR-based FHE compiler targeting multiple libraries and accelerators.",
  "note": "Apache-2.0. Aims to become standard FHE compiler infrastructure.",
  "id": 105
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "DaCapo",
  "year": 2024,
  "venue": "USENIX Security 2024",
  "authors": "Cheon, Seo, Kang et al.",
  "title": "DaCapo: Automatic Bootstrapping Management for Efficient Fully Homomorphic Encryption",
  "paper": "https://www.usenix.org/conference/usenixsecurity24/presentation/cheon",
  "code": null,
  "family": "CKKS compiler",
  "arith": "-",
  "interaction": "Compiler/framework",
  "scale": "-",
  "status": "Active",
  "rel": "Automatic bootstrapping placement for CKKS programs.",
  "note": "SoK notes optimal bootstrapping placement is NP-complete in general.",
  "id": 106
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "Fhelipe",
  "year": 2024,
  "venue": "PLDI 2024",
  "authors": "Krastev, Samardzic, Langowski, Devadas, Sanchez",
  "title": "A Tensor Compiler with Automatic Data Packing for Simple and Efficient Fully Homomorphic Encryption",
  "paper": "https://doi.org/10.1145/3656382",
  "code": "https://github.com/fhelipe-compiler/fhelipe",
  "family": "CKKS",
  "arith": "Python/C++",
  "interaction": "Compiler/framework",
  "scale": "-",
  "status": "Active",
  "rel": "Automatic packing and bootstrapping for tensor programs.",
  "note": "Simpler than hand-tuned packing with competitive performance.",
  "id": 107
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "EVA / CHET",
  "year": 2020,
  "venue": "PLDI 2020",
  "authors": "Dathathri, Kostova, Saarikivi, Dai, Laine, Musuvathi",
  "title": "EVA: An Encrypted Vector Arithmetic Language and Compiler for Efficient Homomorphic Computation",
  "paper": "https://arxiv.org/abs/1912.11951",
  "code": "https://github.com/microsoft/EVA",
  "family": "CKKS (SEAL)",
  "arith": "Python/C++",
  "interaction": "Compiler/framework",
  "scale": "-",
  "status": "Deprecated",
  "rel": "Early CKKS compiler with automatic rescaling.",
  "note": "Archived but influential.",
  "id": 108
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "TenSEAL",
  "year": 2021,
  "venue": "arXiv 2021",
  "authors": "Benaissa, Retiat, Cebere, Belfedhal",
  "title": "TenSEAL: A Library for Encrypted Tensor Operations Using Homomorphic Encryption",
  "paper": "https://arxiv.org/abs/2104.03152",
  "code": "https://github.com/OpenMined/TenSEAL",
  "family": "CKKS, BFV (SEAL)",
  "arith": "Python",
  "interaction": "ML-on-HE toolkit",
  "scale": "-",
  "status": "Niche",
  "rel": "Easy Python tensor API over SEAL.",
  "note": "Good for teaching and prototypes. No bootstrapping.",
  "id": 109
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "HEonGPU",
  "year": 2024,
  "venue": "IACR ePrint",
  "authors": "Özcan, Savaş",
  "title": "HEonGPU: a GPU-based Fully Homomorphic Encryption Library 1.0",
  "paper": "https://eprint.iacr.org/2024/1543",
  "code": "https://github.com/Alisah-Ozcan/HEonGPU",
  "family": "CKKS, BFV, TFHE",
  "arith": "CUDA C++",
  "interaction": "GPU library",
  "scale": "-",
  "status": "Active",
  "rel": "Open GPU library including CKKS bootstrapping.",
  "note": "Apache-2.0. Multi-stream GPU execution.",
  "id": 110
 },
 {
  "group": "7. FHE libraries, GPU back-ends and compilers",
  "name": "TensorFHE",
  "year": 2023,
  "venue": "HPCA 2023",
  "authors": "Fan, Wang, Xu, Hou et al.",
  "title": "TensorFHE: Achieving Practical Computation on Encrypted Data Using GPGPU",
  "paper": "https://arxiv.org/abs/2212.14191",
  "code": null,
  "family": "CKKS",
  "arith": "CUDA (tensor cores)",
  "interaction": "GPU technique",
  "scale": "-",
  "status": "Active",
  "rel": "Uses GPU tensor cores for NTT.",
  "note": "Shows AI accelerators can be repurposed for FHE.",
  "id": 111
 },
 {
  "group": "8. FHE hardware accelerators",
  "name": "F1",
  "year": 2021,
  "venue": "MICRO 2021",
  "authors": "Samardzic, Feldmann, Krastev et al.",
  "title": "F1: A Fast and Programmable Accelerator for Fully Homomorphic Encryption",
  "paper": "https://arxiv.org/abs/2109.05371",
  "code": null,
  "family": "ASIC (BGV/CKKS)",
  "arith": "-",
  "interaction": "Hardware",
  "scale": "-",
  "status": "Foundational",
  "rel": "First programmable FHE ASIC design.",
  "note": "Projected thousands of times faster than CPU. Supports only shallow bootstrapping.",
  "id": 112
 },
 {
  "group": "8. FHE hardware accelerators",
  "name": "CraterLake",
  "year": 2022,
  "venue": "ISCA 2022",
  "authors": "Samardzic, Feldmann, Krastev et al.",
  "title": "CraterLake: A Hardware Accelerator for Efficient Unbounded Computation on Encrypted Data",
  "paper": "https://doi.org/10.1145/3470496.3527393",
  "code": null,
  "family": "ASIC (CKKS)",
  "arith": "-",
  "interaction": "Hardware",
  "scale": "-",
  "status": "Foundational",
  "rel": "Supports deep CKKS with bootstrapping.",
  "note": "Targets unbounded-depth workloads like deep networks.",
  "id": 113
 },
 {
  "group": "8. FHE hardware accelerators",
  "name": "BTS",
  "year": 2022,
  "venue": "ISCA 2022",
  "authors": "Kim, Kim, Kim, Jung et al.",
  "title": "BTS: An Accelerator for Bootstrappable Fully Homomorphic Encryption",
  "paper": "https://arxiv.org/abs/2112.15479",
  "code": null,
  "family": "ASIC (CKKS)",
  "arith": "-",
  "interaction": "Hardware",
  "scale": "-",
  "status": "Foundational",
  "rel": "Accelerator designed around CKKS bootstrapping.",
  "note": "Memory bandwidth centered design.",
  "id": 114
 },
 {
  "group": "8. FHE hardware accelerators",
  "name": "ARK",
  "year": 2022,
  "venue": "MICRO 2022",
  "authors": "Kim, Lee, Kim, Sohn et al.",
  "title": "ARK: Fully Homomorphic Encryption Accelerator with Runtime Data Generation and Inter-Operation Key Reuse",
  "paper": "https://arxiv.org/abs/2205.00922",
  "code": null,
  "family": "ASIC (CKKS)",
  "arith": "-",
  "interaction": "Hardware",
  "scale": "-",
  "status": "Active",
  "rel": "Generates data on the fly to reduce memory traffic.",
  "note": "Addresses the memory wall that the SoK identifies for LLMs.",
  "id": 115
 },
 {
  "group": "8. FHE hardware accelerators",
  "name": "SHARP",
  "year": 2023,
  "venue": "ISCA 2023",
  "authors": "Kim, Park, Kim et al.",
  "title": "SHARP: A Short-Word Hierarchical Accelerator for Robust and Practical Fully Homomorphic Encryption",
  "paper": "https://doi.org/10.1145/3579371.3589053",
  "code": null,
  "family": "ASIC (CKKS)",
  "arith": "-",
  "interaction": "Hardware",
  "scale": "-",
  "status": "Active",
  "rel": "Uses 36-bit words for smaller area and power.",
  "note": "Shows word-size choice matters for CKKS accelerators.",
  "id": 116
 },
 {
  "group": "8. FHE hardware accelerators",
  "name": "Trebuchet",
  "year": 2023,
  "venue": "arXiv 2023",
  "authors": "Cousins, Polyakov, Al Badawi et al.",
  "title": "TREBUCHET: Fully Homomorphic Encryption Accelerator for Deep Computation",
  "paper": "https://arxiv.org/abs/2304.05237",
  "code": null,
  "family": "ASIC (DARPA DPRIVE)",
  "arith": "-",
  "interaction": "Hardware",
  "scale": "-",
  "status": "Active",
  "rel": "DARPA DPRIVE accelerator for deep computation.",
  "note": "Cited by the SoK as evidence for four to five orders of magnitude speedup from custom chips.",
  "id": 117
 },
 {
  "group": "8. FHE hardware accelerators",
  "name": "BASALISC",
  "year": 2023,
  "venue": "USENIX Security 2023",
  "authors": "Geelen, Van Beirendonck, Pereira et al.",
  "title": "BASALISC: Programmable Hardware Accelerator for BGV Fully Homomorphic Encryption",
  "paper": "https://eprint.iacr.org/2022/657",
  "code": null,
  "family": "ASIC (BGV)",
  "arith": "-",
  "interaction": "Hardware",
  "scale": "-",
  "status": "Active",
  "rel": "BGV accelerator (DARPA DPRIVE).",
  "note": "Exact-arithmetic counterpart to CKKS accelerators.",
  "id": 118
 },
 {
  "group": "8. FHE hardware accelerators",
  "name": "Intel HERACLES",
  "year": 2026,
  "venue": "(see SoK ref 103)",
  "authors": "Golder, Kumar, Taneja et al.",
  "title": "HERACLES: 8192-way SIMD Programmable Scalable Fully-Homomorphic Encryption SoC",
  "paper": "https://www.intel.com/content/www/us/en/research/fhe.html",
  "code": null,
  "family": "ASIC (DARPA DPRIVE)",
  "arith": "-",
  "interaction": "Hardware",
  "scale": "-",
  "status": "Active SOTA",
  "rel": "Intel's fabricated FHE SoC.",
  "note": "One of the first silicon FHE accelerators.",
  "id": 119
 },
 {
  "group": "1. Newer CKKS / FHE LLM frameworks (2026, not in the SoK)",
  "name": "Odin",
  "year": 2026,
  "venue": "arXiv 2609.12378",
  "authors": "Fan, Chen, Ye, Ji",
  "title": "An Open-Source End-to-End FHE Implementation for Privacy-Preserving Llama 3 8B Inference",
  "paper": "https://arxiv.org/abs/2609.12378",
  "code": null,
  "family": "CKKS framework",
  "arith": "Approximate real (CKKS)",
  "interaction": "Non-interactive FHE",
  "scale": "Llama-3-8B",
  "status": "Active SOTA",
  "rel": "Open-source end-to-end CKKS Llama-3-8B with co-designed packing.",
  "note": "Feature-major cross-layer layout and transient per-operator layouts avoid repacking between attention and projections.",
  "id": 120
 },
 {
  "group": "1. Newer CKKS / FHE LLM frameworks (2026, not in the SoK)",
  "name": "THEMIS",
  "year": 2026,
  "venue": "IACR ePrint 2026/2264",
  "authors": "Li, Wang, Li, Tang, Liu, Chen",
  "title": "THEMIS: A Co-Designed System for Encrypted Transformer Inference",
  "paper": "https://eprint.iacr.org/2026/2264",
  "code": null,
  "family": "CKKS framework",
  "arith": "Approximate real (CKKS)",
  "interaction": "Non-interactive FHE",
  "scale": "Transformer (see paper)",
  "status": "Active SOTA",
  "rel": "Plans all stages against one shared CKKS level budget.",
  "note": "Bounds depth per stage type and fixes where on the modulus chain each runs. Generalizes coefficient-encoded PCMM.",
  "id": 121
 },
 {
  "group": "1. Newer CKKS / FHE LLM frameworks (2026, not in the SoK)",
  "name": "ROSETTA",
  "year": 2026,
  "venue": "IACR ePrint 2026/2047",
  "authors": "Yu, Zhang, Kong, Ding, Chen, Yu",
  "title": "ROSETTA: Efficient and Accurate Privacy-Preserving LLM Decoding via Hybrid CKKS/TFHE Evaluation",
  "paper": "https://eprint.iacr.org/2026/2047",
  "code": null,
  "family": "Hybrid CKKS + TFHE",
  "arith": "Approximate + LUT",
  "interaction": "Non-interactive FHE",
  "scale": "Generative LLM (see paper)",
  "status": "Active SOTA",
  "rel": "Uses CKKS for linear layers and TFHE for decode-stage non-linearities.",
  "note": "A scheme-switching alternative to the SoK's pure polynomial-approximation approach.",
  "id": 122
 },
 {
  "group": "1. Newer CKKS / FHE LLM frameworks (2026, not in the SoK)",
  "name": "Terrazzo (GL scheme on GPU)",
  "year": 2026,
  "venue": "IACR ePrint 2026/1870",
  "authors": "Shokri, Tsoutsos",
  "title": "Terrazzo: Memory-Aware GPU Framework for Private Transformer Inference",
  "paper": "https://eprint.iacr.org/2026/1870",
  "code": null,
  "family": "Gentry-Lee matrix FHE",
  "arith": "Approximate real matrices",
  "interaction": "Non-interactive FHE",
  "scale": "Small language models",
  "status": "Active SOTA (emerging)",
  "rel": "First private Transformer inference on the GL matrix scheme, the SoK's flagged successor to CKKS.",
  "note": "Co-designs cryptography, algorithms and kernels so GL keys and ciphertexts fit on a commodity GPU.",
  "id": 123
 },
 {
  "group": "1. Newer CKKS / FHE LLM frameworks (2026, not in the SoK)",
  "name": "Zephyr",
  "year": 2026,
  "venue": "IACR ePrint 2026/932",
  "authors": "Seo, Min",
  "title": "Zephyr: GPU-Efficient Homomorphic Encryption for Privacy-Preserving Transformer Inference",
  "paper": "https://eprint.iacr.org/2026/932",
  "code": null,
  "family": "CKKS with 32-bit RNS + grafting",
  "arith": "Approximate real (CKKS)",
  "interaction": "Non-interactive FHE",
  "scale": "Transformer (see paper)",
  "status": "Active SOTA",
  "rel": "Redesigns CKKS for 32-bit GPU integer arithmetic.",
  "note": "Combines 32-bit words with grafting to remove the 64-bit RNS mismatch on GPUs.",
  "id": 124
 },
 {
  "group": "1. Newer CKKS / FHE LLM frameworks (2026, not in the SoK)",
  "name": "Error-propagation-aware scale design",
  "year": 2026,
  "venue": "IACR ePrint 2026/2242",
  "authors": "Seo, Min",
  "title": "Error Propagation-Aware Scale Design for Efficient Homomorphic Encryption-Based LLM Inference",
  "paper": "https://eprint.iacr.org/2026/2242",
  "code": null,
  "family": "CKKS parameter design",
  "arith": "Approximate real (CKKS)",
  "interaction": "Non-interactive FHE",
  "scale": "LLM (see paper)",
  "status": "Active SOTA",
  "rel": "Chooses per-operation CKKS scales from measured error propagation.",
  "note": "Saves modulus and bootstrapping by using smaller scales where errors do not propagate.",
  "id": 125
 },
 {
  "group": "1. Newer CKKS / FHE LLM frameworks (2026, not in the SoK)",
  "name": "ATLAS",
  "year": 2026,
  "venue": "IACR ePrint 2026/1817",
  "authors": "Xie, Tan, Boddeti, Lu",
  "title": "ATLAS: Automated Approximation of Transformers for Efficient Homomorphic Inference in One Hour",
  "paper": "https://eprint.iacr.org/2026/1817",
  "code": null,
  "family": "CKKS approximation search",
  "arith": "Approximate real (CKKS)",
  "interaction": "Non-interactive FHE",
  "scale": "Transformer (see paper)",
  "status": "Active SOTA",
  "rel": "Automatically tunes per-layer polynomial degrees and iteration counts.",
  "note": "Replaces hand-tuned approximation settings with layer-wise search.",
  "id": 126
 },
 {
  "group": "1. Newer CKKS / FHE LLM frameworks (2026, not in the SoK)",
  "name": "HEAT",
  "year": 2026,
  "venue": "IACR ePrint 2026/1862",
  "authors": "Zirilli, Marincione, Kornaropoulos, Ateniese, Rodolà",
  "title": "HEAT: Faster Fully Homomorphic Inference via Approximations-Weights Co-Adaptation",
  "paper": "https://eprint.iacr.org/2026/1862",
  "code": null,
  "family": "CKKS + HE-aware training",
  "arith": "Approximate real (CKKS)",
  "interaction": "Non-interactive FHE",
  "scale": "Language model (see paper)",
  "status": "Active SOTA",
  "rel": "Fine-tunes weights to tolerate cheaper approximations.",
  "note": "Model-modifying (N-MOD style): trades a short training step for fewer bootstraps.",
  "id": 127
 },
 {
  "group": "1. Newer CKKS / FHE LLM frameworks (2026, not in the SoK)",
  "name": "CGF-Softmax",
  "year": 2026,
  "venue": "arXiv 2602.01621",
  "authors": "(see paper)",
  "title": "CGF-Softmax: A Cumulant-Based Softmax Reformulation for Efficient Inference under Homomorphic Encryption",
  "paper": "https://arxiv.org/abs/2602.01621",
  "code": null,
  "family": "CKKS non-linear evaluation",
  "arith": "Approximate real (CKKS)",
  "interaction": "Non-interactive FHE",
  "scale": "ViT and LLMs",
  "status": "Active SOTA",
  "rel": "Softmax without homomorphic division or max subtraction.",
  "note": "Lower depth than normalize-and-square, but it is an approximation of Softmax rather than exact model preservation.",
  "id": 128
 },
 {
  "group": "1. Newer CKKS / FHE LLM frameworks (2026, not in the SoK)",
  "name": "EncFormer",
  "year": 2026,
  "venue": "arXiv 2604.09975",
  "authors": "Zhu, Jin, Aung, Xiao",
  "title": "EncFormer: Secure and Efficient Transformer Inference over Encrypted Data",
  "paper": "https://arxiv.org/abs/2604.09975",
  "code": null,
  "family": "CKKS + MPC hybrid",
  "arith": "Approximate + secret sharing",
  "interaction": "Interactive 2PC",
  "scale": "GPT and BERT style models",
  "status": "Active SOTA",
  "rel": "Hybrid that optimizes FHE-to-MPC boundaries with complex CKKS conversion.",
  "note": "Cost model chooses where to switch between FHE and MPC to minimize conversions.",
  "id": 129
 },
 {
  "group": "2. CKKS core and variants",
  "name": "Discrete CKKS for high-precision bootstrapping",
  "year": 2025,
  "venue": "IACR ePrint 2025/1786",
  "authors": "Choe, Kim, Stehlé, Suvanto",
  "title": "Leveraging Discrete CKKS to Bootstrap in High Precision",
  "paper": "https://eprint.iacr.org/2025/1786",
  "code": null,
  "family": "CKKS bootstrapping",
  "arith": "Approximate real, very high precision",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Active SOTA",
  "rel": "Reaches about 80+ bit precision needed for noise flooding.",
  "note": "Relevant to IND-CPA-D security and threshold CKKS.",
  "id": 130
 },
 {
  "group": "2. CKKS core and variants",
  "name": "Fourier-extension functional bootstrapping",
  "year": 2026,
  "venue": "IACR ePrint 2026/367",
  "authors": "Bian, Fu, Shen, Pan, Wang, Guan",
  "title": "High-Precision Functional Bootstrapping for CKKS from Fourier Extension",
  "paper": "https://eprint.iacr.org/2026/367",
  "code": null,
  "family": "CKKS functional bootstrapping",
  "arith": "LUTs / smooth functions",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Active SOTA",
  "rel": "Evaluates smooth functions during bootstrapping with better error bounds.",
  "note": "Improves on general CKKS functional bootstrapping for piecewise non-linear layers.",
  "id": 131
 },
 {
  "group": "4. Alternative FHE schemes",
  "name": "Efficient bootstrapping for GL matrix FHE",
  "year": 2026,
  "venue": "IACR ePrint 2026/956",
  "authors": "Crockett, Gentry, Kim, Lee, Lee",
  "title": "Efficient Bootstrapping in Fully Homomorphic Encryption for Matrix Arithmetic",
  "paper": "https://eprint.iacr.org/2026/956",
  "code": null,
  "family": "GL scheme bootstrapping",
  "arith": "Approximate real matrices",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Active SOTA (emerging)",
  "rel": "Bootstrapping for the matrix-native GL scheme.",
  "note": "Expresses CoeffToSlot and SlotToCoeff as native matrix products, so key switches per step are a small constant.",
  "id": 132
 },
 {
  "group": "4. Alternative FHE schemes",
  "name": "Low-depth bootstrapping for matrix-native FHE",
  "year": 2026,
  "venue": "IACR ePrint 2026/811",
  "authors": "Shokri, Tsoutsos",
  "title": "Low-Depth Bootstrapping for Matrix-Native FHE",
  "paper": "https://eprint.iacr.org/2026/811",
  "code": null,
  "family": "GL scheme bootstrapping",
  "arith": "Approximate real matrices",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Active SOTA (emerging)",
  "rel": "Reduces bootstrapping depth in GL-style schemes.",
  "note": "Same authors as Terrazzo.",
  "id": 133
 },
 {
  "group": "4. Alternative FHE schemes",
  "name": "Trace-Factored BigSwitch (GL keys)",
  "year": 2026,
  "venue": "IACR ePrint 2026/2058",
  "authors": "Park, Jeong, Jeong, Oh, Lee, Kim",
  "title": "Trace-Factored BigSwitch for Matrix-Friendly FHE",
  "paper": "https://eprint.iacr.org/2026/2058",
  "code": null,
  "family": "GL scheme key switching",
  "arith": "Approximate real matrices",
  "interaction": "Non-interactive",
  "scale": "-",
  "status": "Active SOTA (emerging)",
  "rel": "Removes the largest GL evaluation key.",
  "note": "Targets the evaluation-key memory bottleneck of matrix-native FHE.",
  "id": 134
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "SHAFT",
  "year": 2025,
  "venue": "IACR ePrint 2025/2324",
  "authors": "Kei, Chow",
  "title": "SHAFT: Secure, Handy, Accurate, and Fast Transformer Inference",
  "paper": "https://eprint.iacr.org/2025/2324",
  "code": null,
  "family": "2PC secret sharing",
  "arith": "Fixed point",
  "interaction": "Interactive 2PC",
  "scale": "BERT, GPT-2",
  "status": "Active SOTA",
  "rel": "Constant-round Softmax and cheaper GELU protocols.",
  "note": "Avoids the max extraction that costs logarithmic rounds.",
  "id": 135
 },
 {
  "group": "5. Interactive MPC and HE+MPC hybrids",
  "name": "CRAfT",
  "year": 2026,
  "venue": "IACR ePrint 2026/1292",
  "authors": "Zhao, Yang, Song, Hao, He, Zhang",
  "title": "CRAfT: Constant-Round Non-Polynomial Approximation and P2P Network for Secure Transformer Inference",
  "paper": "https://eprint.iacr.org/2026/1292",
  "code": null,
  "family": "HE + MPC, multi-user",
  "arith": "Mixed",
  "interaction": "Interactive (constant rounds)",
  "scale": "Transformer (see paper)",
  "status": "Active SOTA",
  "rel": "Rotation-free HE packing for many mutually distrusting users.",
  "note": "Moves communication to a peer-to-peer network to scale to many users.",
  "id": 136
 }
]