Research map / Federated learning and privacy / Differential privacy
Private and Stable Test-time Adaptation with Differential Privacy
Zefeng Li, Qiaoyue Tang, Mathias Lécuyer, Evan Shelhamer
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 |
|---|---|---|---|
| add per-sample gradient clipping (no noise) to Tent | Tent · ImageNet-C sev 5, continual, ViT-B/16 | avg top-1 accuracy (%): 60.8 -> 64.9 | Improved |
| add per-sample gradient clipping to EATA | EATA · ImageNet-C sev 5, continual, ViT-B/16 | avg top-1 accuracy (%): 65.9 -> 66.0 | No clear effect |
| add per-sample gradient clipping to SAR | SAR · ImageNet-C sev 5, continual, ViT-B/16 | avg top-1 accuracy (%): 60.8 -> 61.7 | Improved |
| add per-sample gradient clipping to DeYO | DeYO · ImageNet-C sev 5, continual, ViT-B/16 | avg top-1 accuracy (%): 61.0 -> 62.7 | Improved |
| add per-sample gradient clipping to DeYO-COME | DeYO-COME · ImageNet-C sev 5, continual, ViT-B/16 | avg top-1 accuracy (%): 66.4 -> 67.5 | Improved |
| make Tent differentially private at epsilon=10 to 20 | Tent · ImageNet-C sev 5, continual, ViT-B/16 | avg top-1 accuracy (%): 60.8 -> 62.9 (epsilon=20) | Improved |
| make EATA, SAR, DeYO, DeYO-COME private at epsilon=20 | EATA, SAR, DeYO, DeYO-COME · ImageNet-C sev 5, continual, ViT-B/16 | avg top-1 accuracy (%): -2.9%, -1.2%, -2.4%, -1.7% | Hurt |
| strengthen privacy to epsilon=1 for DP-Tent (ablation) | DP-Tent · ImageNet-C sev 5, continual, ViT-B/16 | avg top-1 accuracy (%): accuracy degrades by 2.3% | Hurt |
| use batch-level clipping instead of per-sample clipping (ablation) | Tent · ImageNet-C sev 5, continual, ViT-B/16 | avg top-1 accuracy (%): 60.8 -> 59.8 (C=1; Tent in Table 2 59.7) | No clear effect |
| replace hard filtering by grad norm with rescaling by per-sample clipping (ablation) | Tent · ImageNet-C sev 5, continual, ViT-B/16 | avg top-1 accuracy (%): 55.0 -> 64.9 at C=1 | Improved |
| filter by entropy loss instead of clipping (ablation) | Tent · ImageNet-C sev 5, continual, ViT-B/16 | avg top-1 accuracy (%): 64.9 -> 62.9 | Hurt |
| add EATA filters (Filter-1 and Filter-2) to Tent (ablation) | Tent · ImageNet-C sev 5, episodic and continual | avg top-1 accuracy (%): episodic 63.0 -> 65.5, continual 60.8 -> 63.7 | Improved |
Closest papers by meaning
- Evaluation of Test-Time Adaptation Under Computational Time Constraints (ICML 2024 · similarity 0.87)
- PTTA: Purifying Malicious Samples for Test-Time Model Adaptation (ICML 2025 · similarity 0.86)
- DiSK: Differentially Private Optimizer with Simplified Kalman Filter for Noise Reduction (ICLR 2025 · similarity 0.85)
- On Inherent Privacy of Posterior Sampling: A Unified R\'enyi-Divergence Framework (NeurIPS 2026 · similarity 0.85)
- Backpropagation-Free Test-Time Adaptation via Probabilistic Gaussian Alignment (NeurIPS 2025 · similarity 0.85)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.85)
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.
