Research map / Adversarial robustness and security / Attacks on text and graphs
Uncovering Adversarial Risks of Test-Time Adaptation
Tong Wu, Feiran Jia, Xiangyu Qi, Jiachen T. Wang, Vikash Sehwag, Saeed Mahloujifar, Prateek Mittal
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 |
|---|---|---|---|
| Inject malicious test-batch samples using DIA | TeBN, TENT, and pseudo-label TTA · CIFAR-10-C, CIFAR-100-C, and ImageNet-C | targeted attack success rate: ImageNet-C: 10 malicious samples (5%) yields 99.47 (TeBN), 98.67 (TENT), and 96.53 (Hard PL) | Improved |
| Increase the number of malicious samples in the batch | DIA on TTA · CIFAR-10-C, CIFAR-100-C, and ImageNet-C | targeted attack success rate: CIFAR-10-C TeBN: 25.87 -> 55.47 -> 92.80; CIFAR-100-C TeBN: 46.80 -> 93.73 -> 100.00 | Improved |
| Inject malicious samples for an indiscriminate attack | TeBN on benign corrupted samples · CIFAR-10-C | corruption error rate: 10.73 -> 28.02 | Hurt |
| Inject malicious samples for an indiscriminate attack | TeBN on benign corrupted samples · CIFAR-100-C | corruption error rate: 34.52 -> 58.41 | Hurt |
| Inject 20 malicious samples for an indiscriminate attack | TeBN on benign corrupted samples · ImageNet-C | corruption error rate: 47.79 -> 79.03 | Hurt |
| Use adversarially optimized snow or fog corruptions as injected samples | TeBN · ImageNet-C | targeted attack success rate: Snow: 64.00; Fog: 84.00 | Improved |
| Use adversarially trained source models | TeBN on ImageNet-C · ImageNet-C | targeted attack success rate: With 8 malicious samples, ASR degrades by ~80%; with 32 malicious samples, DIA still exceeds 70% success | Hurt |
| Blend training-time BN statistics with test-time BN statistics (τ = 0.6) | TTA BN estimation on ImageNet-C · ImageNet-C | targeted attack success rate: ASR to ~20% for 20 malicious samples with robust models | Hurt |
| Use training-time BN for the last 8 layers, with τ = 0.6 elsewhere | TTA BN estimation on ImageNet-C · ImageNet-C | targeted attack success rate: ASR drops ~40% for the standard model with N_tr = 8 | Hurt |
| Combine the mitigation approaches: adversarially trained model and robust BN estimation | TTA on ImageNet-C · ImageNet-C | targeted attack success rate: mitigate ASR by ~50% for the standard model and ~40% for the robust model | Hurt |
| Increase the weight ω to 0.1 in the stealthy targeted objective (ablation) | DIA on TeBN · CIFAR-10-C | targeted attack success rate: ASR remains more than 75%; corruption accuracy degradation drops to ~2% | Mixed |
| Use stronger data augmentation (AugMix or DeepAugment) (ablation) | DIA on ImageNet-C · ImageNet-C | targeted attack success rate: TeBN: 99.47 (Standard), 98.40 (AugMix), 96.00 (DeepAugment) | Hurt |
Closest papers by meaning
- On the Adversarial Risk of Test Time Adaptation: An Investigation into Realistic Test-Time Data Poisoning (ICLR 2025 · similarity 0.93)
- MedBN: Robust Test-Time Adaptation against Malicious Test Samples (CVPR 2024 · similarity 0.89)
- PTTA: Purifying Malicious Samples for Test-Time Model Adaptation (ICML 2025 · similarity 0.89)
- Monitoring Risks in Test-Time Adaptation (NeurIPS 2025 · similarity 0.86)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.85)
- Less Diverse, Less Safe: The Indirect But Pervasive Risk of Test-Time Scaling in Large Language Models (ICML 2026 · 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.
