Research map / Federated learning and privacy / Privacy attacks and secure learning
HETAL: Efficient Privacy-preserving Transfer Learning with Homomorphic Encryption
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| encrypt the client data and train the classifier with HETAL instead of nonencrypted training | transfer learning with ViT-Base · MNIST | accuracy: 97.24% -> 96.73% | Hurt |
| encrypt the client data and train the classifier with HETAL instead of nonencrypted training | transfer learning with ViT-Base · CIFAR-10 | accuracy: 96.62% -> 96.57% | Hurt |
| encrypt the client data and train the classifier with HETAL instead of nonencrypted training | transfer learning with ViT-Base · Face Mask Detection | accuracy: 95.46% -> 95.46% | No clear effect |
| encrypt the client data and train the classifier with HETAL instead of nonencrypted training | transfer learning with ViT-Base · DermaMNIST | accuracy: 76.01% -> 76.06% | No clear effect |
| encrypt the client data and train the classifier with HETAL instead of nonencrypted training | transfer learning with MPNet-Base · SNIPS | accuracy: 94.43% -> 95.00% | Helped |
| use DiagABT instead of ColMajor | encrypted ABᵀ multiplication · ABᵀ (128, 128, 4) | running time: 0.1104 -> 0.0601 | Helped |
| use DiagATB instead of RowMajor | encrypted AᵀB multiplication · AᵀB (128, 128, 4) | running time: 0.1171 -> 0.0415 | Helped |
| add domain extension to the normalized softmax approximation (ablation) | HETAL softmax approximation · Softmax approximation, c=3, R=128 | maximum error: The norm-only variant has no reported value at R=128; norm+extn: 0.0037 | Helped |
| add precise correction to normalization and domain extension (ablation) | HETAL softmax approximation · Softmax approximation, c=3, R=128 | maximum error: 0.0037 -> 0.0022 | Helped |
Closest papers by meaning
- Encryption-Friendly LLM Architecture (ICLR 2025 · similarity 0.86)
- Privacy-Preserving Embedding via Look-up Table Evaluation with Fully Homomorphic Encryption (ICML 2024 · similarity 0.85)
- United We Stand: Accelerating Privacy-Preserving Neural Inference by Conjunctive Optimization with Interleaved Nexus (AAAI 2024 · similarity 0.85)
- EncryptedLLM: Privacy-Preserving Large Language Model Inference via GPU-Accelerated Fully Homomorphic Encryption (ICML 2025 · similarity 0.84)
- Blind-Touch: Homomorphic Encryption-Based Distributed Neural Network Inference for Privacy-Preserving Fingerprint Authentication (AAAI 2024 · similarity 0.84)
- Converting Transformers to Polynomial Form for Secure Inference Over Homomorphic Encryption (ICML 2024 · similarity 0.84)
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.