Research map / Federated learning and privacy / Privacy attacks and secure learning
DataMix: Efficient Privacy-Preserving Edge-Cloud Inference
Zhijian Liu, Zhanghao Wu, Chuang Gan, Ligeng Zhu, Song Han
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 |
|---|---|---|---|
| use DataMix instead of Face Anonymizer | edge-cloud inference · CelebA | person-ID attack success rate (%): 62.6% -> 0.6% | Improved |
| use DataMix instead of Face Anonymizer | edge-cloud inference · CelebA | edge FLOPs: 47.13G -> 0.09G | Improved |
| use DataMix instead of Face Anonymizer | edge-cloud inference · CelebA | test accuracy (%): 89.8% -> 90.7% | Improved |
| add more preprocess conv blocks on edge (Npre 1 -> 3) (ablation) | DataMix · CelebA | test accuracy (%): 90.7% -> 91.0% | Improved |
| add more preprocess conv blocks on edge (Npre 1 -> 3) (ablation) | DataMix · Speech Commands | test accuracy (%): 92.2% -> 94.6% | Improved |
| add more preprocess conv blocks on edge (Npre 1 -> 3) (ablation) | DataMix · Speech Commands | speaker-ID attack success rate (%): 19.4% -> 12.7% | Improved |
| increase Gaussian noise from N(0,4) to N(0,8) (ablation) | adding noise · CelebA | test accuracy (%): 88.6% -> 87.9% | Hurt |
| increase Gaussian noise from N(0,4) to N(0,8) (ablation) | adding noise · CelebA | person-ID attack success rate (%): 46.5% -> 35.3% | Improved |
| mix two images instead of raw images (ablation) | DataMix · CelebA | person-ID attack success rate: reduced by 43.7% | Improved |
| increase mixing group size SG (ablation) | DataMix · CelebA | attribute accuracy vs attack success rate | Mixed |
| use DataMix instead of adding noise N(0,8) | edge-cloud inference · Speech Commands | speaker-ID attack success rate (%): 89.9% -> 12.7% | Improved |
| use DataMix instead of hand-crafted noise or blurring | edge-cloud inference · CelebA | person-ID attack success rate: more than 20x lower | Improved |
Closest papers by meaning
- PRISM: Privacy-Aware Routing for Adaptive Cloud–Edge LLM Inference via Semantic Sketch Collaboration (AAAI 2026 · similarity 0.88)
- United We Stand: Accelerating Privacy-Preserving Neural Inference by Conjunctive Optimization with Interleaved Nexus (AAAI 2024 · similarity 0.87)
- Posthoc privacy guarantees for collaborative inference with modified Propose-Test-Release (NeurIPS 2023 · similarity 0.86)
- ObCLIP: Oblivious CLoud-Device Hybrid Image Generation with Privacy Preservation (NeurIPS 2025 · similarity 0.86)
- All Rivers Run to the Sea: Private Learning with Asymmetric Flows (CVPR 2024 · similarity 0.86)
- Towards Robust and Efficient Cloud-Edge Elastic Model Adaptation via Selective Entropy Distillation (ICLR 2024 · similarity 0.83)
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.
