Error Propagation-Aware Scale Design for Efficient Homomorphic Encryption-Based LLM Inference¶
Authors: Seo, Min
Year: 2026
Venue: IACR ePrint 2026/2242
type: system collection: catalogue status: auto
Links: ePrint 2026/2242
Summary¶
Auto-extracted summary (unreviewed)
Characterizes the numerical error of each CKKS operation and how it propagates through Transformer computations, then uses the quantified contribution of each local error to the final inference error to assign per-operation scales and modulus levels, avoiding over-conservative precision. For a standard 12-layer Transformer on MRPC this cuts modulus consumption by 30.0% and bootstraps per encoder from 61 to 10 versus THOR (inference 602 s to 135 s at 85.0% accuracy); for an HE-friendly Transformer on RTE it cuts modulus by 33.6% and bootstraps by 80% versus PowerFormer (344 s to 86.7 s, 69.6% to 69.3% accuracy).
Abstract¶
Homomorphic encryption (HE) enables privacy-preserving inference by allowing neural networks to operate directly on encrypted data, but its computational cost remains a major obstacle to deploying large language models in practice. In particular, CKKS-based inference consumes ciphertext modulus through homomorphic multiplications and requires costly bootstrapping when the available modulus is exhausted. In this work, we propose an error-propagation-aware scale design for efficient CKKS-based privacy-preserving LLM inference. We first characterize the numerical errors introduced by individual homomorphic operations and analyze how they propagate through subsequent Transformer computations. Based on this analysis, we quantify the contribution of each local error to the final inference error and determine the precision required for individual operations. We then allocate operation-wise scales and modulus levels accordingly, avoiding unnecessarily conservative precision while maintaining the target inference accuracy. As a result, more computation can be performed within a given modulus chain, reducing the frequency of costly bootstrapping operations and improving overall inference efficiency. Compared with THOR, our method reduces modulus consumption by 30.0% and the number of bootstrapping operations by 83.6% for the standard Transformer. For the HE-friendly Transformer, our method reduces modulus consumption by 33.6% and the number of bootstrapping operations by 80% compared with PowerFormer.
Source: pdf
Extracted details¶
| Field | Value | Source | Confidence |
|---|---|---|---|
| contribution | Error characterization of individual CKKS homomorphic operations | llm | 0.80 |
| contribution | Error-propagation analysis through Transformer layers quantifying each operation's contribution to final error | llm | 0.80 |
| contribution | Operation-wise scale and modulus-level allocation meeting a target accuracy | llm | 0.80 |
| contribution | 30-34% lower modulus consumption and 80-84% fewer bootstrappings vs THOR and PowerFormer | llm | 0.80 |
| schemes | CKKS | llm | 0.80 |
| models | 12-layer Transformer encoder (standard) | llm | 0.80 |
| models | HE-friendly Transformer (PowerFormer-style) | llm | 0.80 |
| datasets | MRPC | llm | 0.80 |
| datasets | RTE | llm | 0.80 |
| tasks | private transformer inference | llm | 0.80 |
| interaction | non-interactive FHE | llm | 0.80 |
| limitations | Baseline modulus consumption for THOR and PowerFormer is estimated from their papers and public code; THOR's bootstrap count is taken from ATLAS | llm | 0.80 |
| llm relevance | high | llm | 0.80 |
| key result | modulus consumption reduction: 30.0% (3400 -> 2379 bits per encoder) (standard Transformer on MRPC vs THOR) | llm | 0.80 |
| key result | bootstrapping count reduction: 83.6% (61 -> 10 per encoder) (standard Transformer on MRPC vs THOR) | llm | 0.80 |
| key result | inference time: 135 s vs 602 s (THOR) (12-layer Transformer on MRPC, both 85.0% accuracy) | llm | 0.80 |
| key result | modulus consumption reduction: 33.6% (1040 -> 691 bits per encoder) (HE-friendly Transformer on RTE vs PowerFormer) | llm | 0.80 |
| key result | inference time: 86.7 s vs 344 s (PowerFormer) (HE-friendly Transformer on RTE, 69.3% vs 69.6% accuracy, bootstraps 2 vs 10) | llm | 0.80 |
System card: Error-propagation-aware scale design¶
| Group | 1. Newer CKKS / FHE LLM frameworks (2026, not in the SoK) |
| Family | CKKS parameter design |
| Arithmetic | Approximate real (CKKS) |
| Interaction | Non-interactive FHE |
| Model scale | LLM (see paper) |
| Status | Active SOTA |
| Relation to the SoK | Chooses per-operation CKKS scales from measured error propagation. |
| Note | Saves modulus and bootstrapping by using smaller scales where errors do not propagate. |
| Systems page | Error-propagation-aware scale design |
Keyword mentions¶
Counts of vocabulary terms in the full text (first page of occurrence in brackets).
- scheme: CKKS 16 [p.1], MPC 2 [p.1], TEE / confidential computing 2 [p.1]
- library: Cheddar 2 [p.2]
- model: Transformer 53 [p.1], BERT 9 [p.7], LLaMA 1 [p.2]
- dataset: GLUE tasks (QNLI/RTE/MRPC/CoLA/QQP/MNLI/STS-B) 21 [p.2], SST-2 1 [p.12]
- hardware: GPU 7 [p.2], NVIDIA A100 3 [p.8]
- technique: Bootstrapping 80 [p.1], GELU 25 [p.4], Softmax 21 [p.2], LayerNorm 13 [p.8], Polynomial approximation 12 [p.1], Minimax approximation 6 [p.1], RNS 6 [p.1], Fine-tuning 6 [p.7], Rescaling 4 [p.1], Attention 4 [p.8]
- topic: Homomorphic encryption 79 [p.1], Large language models 9 [p.1], Private / secure inference 4 [p.1]
Files¶
| Path (under papers/) | Collection | Pages | Size | SHA-256 | Note |
|---|---|---|---|---|---|
references/catalogue/1-125_error-propagation-aware-scale-design-for-efficient-homomorph.pdf |
catalogue | 19 | 445.5 KB | cd2e72cc4dcf |
Database key: error-propagation-aware-scale-design-for-efficient-homomorphic · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set error-propagation-aware-scale-design-for-efficient-homomorphic COLUMN VALUE