Research map / Architectures and efficiency / RNNs and normalization
FuseFSS: Efficient Secure LLM Inference with Function Secret Sharing
Yuhan Ma, Yong Li, Stefan Schmid
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Sigma's hand-written nonlinear protocols with FuseFSS compiled two-call gates | FuseFSS · BERT/GPT models, seq 128 | online latency: BERT-base 1613.80 -> 1149.50 ms; 19-33% reduction | Improved |
| use FuseFSS compiled gates | FuseFSS · BERT/GPT models, seq 128 | key size and key generation time: BERT-base key size 18.076 -> 13.678 GB; keygen 1.32 -> 1.08 s | Improved |
| use FuseFSS instead of Sigma on fixed-point inference | FuseFSS · GLUE and LAMBADA | accuracy: BERT-base SST-2 89.33 -> 89.33; GPT-2 LAMBADA 60.59 -> 60.90 | No clear effect |
| use FuseFSS compiled path for nExp and reciprocal in softmax | FuseFSS · BERT-base softmax, seq 128 | time (ms): 223 -> 49 | Improved |
| FuseFSS does not speed up max-reduction in softmax much | FuseFSS · BERT-base softmax, seq 128 | time (ms): 93 -> 77 | No clear effect |
| increase sequence length for BERT-base (ablation) | FuseFSS · BERT-base, seq 32 to 512 | online time reduction: 15.1% at 32 -> 30.8% at 512 | Improved |
| pad compiled instances to mask-independent public shape (required for security) (ablation) | FuseFSS GELU gate · BERT-base GELU gate | per-gate time (ms): 0.264 -> 1.014 | Hurt |
| cache compiled gate program across layers (ablation) | FuseFSS · BERT-base, 12 GELU calls | total init time (ms): 10.389 -> 0.903 | Improved |
| use FuseFSS on LLaMA-family models | FuseFSS · LLaMA-7B, seq 16 to 64 | end-to-end time speedup: 1.09x at seq 16 -> 1.22x at seq 64 | Improved |
| use FuseFSS compiled activation gates | FuseFSS · GELU/SiLU gate microbenchmark L=128 | per-gate speedup and key size: 1.64-2.46x speedup; 4.96-6.25x smaller keys | Improved |
Closest papers by meaning
- TensorCommitments: A Lightweight Verifiable Inference for Large Language Models (NeurIPS 2026 · similarity 0.89)
- EncryptedLLM: Privacy-Preserving Large Language Model Inference via GPU-Accelerated Fully Homomorphic Encryption (ICML 2025 · similarity 0.88)
- Hidden No More: Attacking and Defending Private Third-Party LLM Inference (ICML 2025 · similarity 0.87)
- Split-and-Denoise: Protect large language model inference with local differential privacy (ICML 2024 · similarity 0.86)
- Fingerprinting Inference Systems of Large Language Models (NeurIPS 2026 · similarity 0.86)
- Ditto: Quantization-aware Secure Inference of Transformers upon MPC (ICML 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
