Privacypreserving llm inference in practice: A comparative survey of techniques, trade-offs, and deployability¶
Authors: D. Andreoletti; A. Rudi; E. Carpanzano; T. Leidi
Year: 2026
type: survey collection: sok-references status: auto
Links: ePrint 2026/105
Summary¶
Auto-extracted summary (unreviewed)
Surveys privacy-preserving inference for transformer-based LLMs from a deployment perspective, adopting end-to-end confidentiality of prompt and completion as the operational privacy goal and organising the review by PET family: MPC, FHE, hybrid HE/MPC designs, TEEs and client-side information reduction. It compares trust assumptions, scalability and deployment maturity, concluding that TEEs (SGX, TDX, H100 confidential computing) are today the only practical route for large LLMs at near-production latency, hybrid crypto designs reduce hardware trust at higher cost, and FHE is the principled long-term endpoint for non-interactive confidentiality.
Abstract¶
Large Language Models (LLMs) are increasingly deployed as cloud services, raising practical concerns about the confidentiality of user prompts and generated completions. In this paper, we survey privacy-preserving inference solutions for Transformer-based LLMs with the explicit goal of supporting operational choices in real-world deployments. We adopt a strong operational notion of privacy: only the client can read the prompt and the corresponding completion, end to end. The review is organised around the main families of Privacy- Enhancing Technologies (PETs). For each family, we examine representative systems and how they address key bottlenecks in confidential LLM inference, such as non-linear layers and autoregressive decoding. We then compare these approaches in terms of trust assumptions, scalability, and deployment maturity. This comparison characterises the current practical landscape of privacy-preserving LLM inference and motivates a trustminimising deployment trajectory: from TEE-based solutions that enable large-scale confidential inference today; through crypto-augmented designs that reduce reliance on hardware trust at higher computational cost; toward Fully Homomorphic Encryption as a principled long-term endpoint for non-interactive confidentiality.
Source: pdf
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | Deployment-oriented survey of PETs for confidential LLM inference | llm | 0.85 |
| contribution | Analysis of how each family handles non-linear layers and autoregressive decoding | llm | 0.85 |
| contribution | Comparison by trust assumptions, scalability and deployment maturity | llm | 0.85 |
| contribution | Trust-minimising deployment trajectory from TEEs through hybrid crypto to FHE | llm | 0.85 |
| schemes | FHE | llm | 0.85 |
| schemes | MPC | llm | 0.85 |
| schemes | FSS | llm | 0.85 |
| schemes | TEE | llm | 0.85 |
| models | BERT-base | llm | 0.85 |
| models | BERT-large | llm | 0.85 |
| models | GPT-2 | llm | 0.85 |
| models | LLaMA-2 | llm | 0.85 |
| models | OPT-30B | llm | 0.85 |
| models | OPT-66B | llm | 0.85 |
| tasks | private LLM inference | llm | 0.85 |
| tasks | autoregressive decoding | llm | 0.85 |
| tasks | retrieval-augmented generation in TEEs | llm | 0.85 |
| hardware | Intel SGX | llm | 0.85 |
| hardware | Intel TDX | llm | 0.85 |
| hardware | NVIDIA H100 Confidential Computing | llm | 0.85 |
| threat model | end-to-end confidentiality of prompt and completion against the cloud operator; content privacy prioritised over model privacy | llm | 0.85 |
| interaction | other | llm | 0.85 |
| limitations | Cryptographic approaches remain constrained by interaction (MPC) and computational cost (FHE) | llm | 0.85 |
| limitations | Hybrid HE/MPC designs add protocol-engineering complexity and HE-MPC conversion cost | llm | 0.85 |
| limitations | TEE GPU confidential computing suffers encrypted memory-movement bottlenecks for models exceeding device memory | llm | 0.85 |
| llm relevance | high | llm | 0.85 |
SoK reference [26]¶
D. Andreoletti, A. Rudi, E. Carpanzano, and T. Leidi, “Privacypreserving llm inference in practice: A comparative survey of techniques, trade-offs, and deployability,” Cryptology ePrint Archive, 2026.
Listed on the reference page of the site.
Keyword mentions¶
Counts of vocabulary terms in the full text (first page of occurrence in brackets).
- scheme: MPC 67 [p.2], TEE / confidential computing 48 [p.1], Secret sharing 7 [p.2], Function secret sharing 5 [p.4], CKKS 3 [p.6], Differential privacy 3 [p.8]
- model: Transformer 23 [p.1], GPT 11 [p.2], LLaMA 7 [p.1], BERT 4 [p.2]
- hardware: GPU 18 [p.2], NVIDIA H100 4 [p.7]
- technique: GELU 12 [p.2], Softmax 11 [p.2], Non-interactive 10 [p.1], LayerNorm 8 [p.2], KV cache 6 [p.2], Malicious model 6 [p.5], Bootstrapping 5 [p.3], Polynomial approximation 5 [p.4], SIMD packing 4 [p.3], Matrix multiplication 4 [p.3], Two-party computation 4 [p.5], Semi-honest model 4 [p.3], Three-party computation 3 [p.5], Attention 2 [p.2]
- task: Language modeling / generation 18 [p.1], Medical / healthcare 2 [p.1]
- topic: Large language models 68 [p.1], Homomorphic encryption 61 [p.2], Private / secure inference 11 [p.2], Benchmark 2 [p.6]
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/sok-references/026_privacy-preserving-llm-inference-in-practice-a-comparative-s.pdf |
sok-references | 13 | 431.9 KB | 8aabff33c1f1 |
Database key: privacypreserving-llm-inference-in-practice-a-comparative-survey-of · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set privacypreserving-llm-inference-in-practice-a-comparative-survey-of COLUMN VALUE