Research map / Federated learning and privacy / Privacy attacks and secure learning
Iron: Private Inference on Transformers
Meng Hao, Hongwei Li, Hanxiao Chen, Pengzhi Xing, Guowen Xu, Tianwei Zhang
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 |
|---|---|---|---|
| Compact packing of whole matrix into one ciphertext for HE matmul vs Cheetah row-wise packing | Iron matmul protocol · Matmul (32,8,16) | communication (MB): 2.79 -> 0.11 | Improved |
| Customized Softmax/GELU/LayerNorm protocols (merged LayerNorm, square trick) vs SIRNN | Iron non-linear protocols · Softmax | communication (MB): 347.71 -> 206.265 | Improved |
| Iron end-to-end vs SIRNN on four BERT models | Iron · BERT end-to-end private inference | runtime: 3.30x to 11.83x faster; 3.47x to 14.11x less communication | Improved |
| Iron fixed-point vs plaintext float on BERT-Tiny | Iron · GLUE (MNLI, MRPC, QNLI, SST-2) | accuracy: loss not exceeding 0.3%; +0.85% on MNLI | No clear effect |
| Lower fractional scale in fixed-point encoding (below 12) (ablation) | Iron · MRPC BERT-Tiny | accuracy: scale 12 matches plaintext; loss under 1% for scale >= 6 | Hurt |
Closest papers by meaning
- Converting Transformers to Polynomial Form for Secure Inference Over Homomorphic Encryption (ICML 2024 · similarity 0.91)
- Nimbus: Secure and Efficient Two-Party Inference for Transformers (NeurIPS 2024 · similarity 0.91)
- CipherPrune: Efficient and Scalable Private Transformer Inference (ICLR 2025 · similarity 0.91)
- Encryption-Friendly LLM Architecture (ICLR 2025 · similarity 0.89)
- Ditto: Quantization-aware Secure Inference of Transformers upon MPC (ICML 2024 · similarity 0.87)
- Jaguar: Fast Private CNN Inference with Power-of-Two Homomorphic Arithmetic (NeurIPS 2026 · similarity 0.87)
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.
