Research map / Adversarial robustness and security / Attacks on text and graphs
Exploiting Supervised Poison Vulnerability to Strengthen Self-Supervised Defense
Jeremy Styborski, Mingzhi Lyu, Yi Huang, Adams Kong
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 SL-guided adversarial training to SSL+SL (VESPR) (ablation) | SSL+SL · ImageNet-100 poisoned training | Psn Min accuracy (%): 21.82 -> 45.20 | Improved |
| add SL-guided adversarial training to SSL+SL (VESPR) (ablation) | SSL+SL · ImageNet-100 poisoned training | Psn Avg accuracy (%): 51.17 -> 55.37 | Improved |
| use gradient-based adversarial perturbation instead of Gaussian noise augmentation (ablation) | SSL+SL · ImageNet-100 poisoned training | Psn Avg accuracy (%): 52.27 -> 55.37 | Improved |
| generate adversarial examples with SL cross-entropy loss instead of SSL loss (ablation) | VESPR · ImageNet-100 poisoned training | Psn Min accuracy (%): 36.64 -> 45.20 | Improved |
| generate adversarial examples with CE loss instead of combined loss (ablation) | VESPR · ImageNet-100 poisoned training | Psn Avg accuracy (%): 54.10 -> 55.37 | Improved |
| apply adversarial training to SSL (SSL+AT) instead of plain SSL (ablation) | SSL · ImageNet-100 poisoned training | Psn Avg accuracy (%): 46.05 -> 44.81 | Hurt |
| add adversarial training to supervised learning (SL+AT) (ablation) | SL · ImageNet-100 poisoned training | Psn Avg accuracy (%): 21.41 -> 44.29 | Improved |
| use SSL-style augmentations with supervised training (ablation) | SL · ImageNet-100 poisoned training | Psn Avg accuracy (%): 21.41 -> 41.54 | Improved |
| add supervised CE loss to SSL without adversarial training (SSL+SL) (ablation) | SSL · ImageNet-100 poisoned training | Psn Min accuracy (%): Psn Avg 46.05 -> 51.17, Psn Min 14.80 -> 21.82 | Mixed |
| use cutout, mixup or CutMix augmentation defenses | SL · ImageNet-100 poisoned training | Psn Min accuracy (%): Psn Min 6.10 -> 5.06 / 5.16 / 5.70 | No clear effect |
| use SSL (SimCLR) against the SSL-targeted CP8 poison | SSL · ImageNet-100 CP8 | clean test accuracy (%): 14.80 vs SL 57.46 | Hurt |
| add adversarial training in VESPR instead of SSL+SL, trained on clean data (ablation) | SSL+SL · ImageNet-100 clean data | clean test accuracy (%): 79.66 -> 74.84 | Hurt |
Closest papers by meaning
- Not All Poisons are Created Equal: Robust Training against Data Poisoning (ICML 2022 · similarity 0.89)
- Attack To Defend: Exploiting Adversarial Attacks for Detecting Poisoned Models (CVPR 2024 · similarity 0.89)
- Adversarial Examples Make Strong Poisons (NeurIPS 2021 · similarity 0.89)
- Witches' Brew: Industrial Scale Data Poisoning via Gradient Matching (ICLR 2021 · similarity 0.88)
- Adversarial Training for Defense Against Label Poisoning Attacks (ICLR 2025 · similarity 0.87)
- Wicked Oddities: Selectively Poisoning for Effective Clean-Label Backdoor Attacks (ICLR 2025 · 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.
