Research map / Adversarial robustness and security / Attacks on text and graphs
Adversarial Examples Make Strong Poisons
Liam H Fowl, Micah Goldblum, Ping-yeh Chiang, Jonas Geiping, Wojciech Czaja, Tom Goldstein
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 |
|---|---|---|---|
| Train on PGD adversarial poisons rather than clean data | ResNet-18 · CIFAR-10 validation | accuracy (%): 94.56 -> 6.25 | Hurt |
| Use adversarial poisoning instead of Unlearnable Examples | ResNet-18 · CIFAR-10 validation | accuracy (%): 19.85 -> 6.25 | Hurt |
| Use class-targeted rather than untargeted poisons (ablation) | Adversarial Poisoning · ImageNet validation | accuracy (%): 45.07 -> 3.57 | Hurt |
| Increase the perturbation bound from ε=4/255 to ε=8/255 (ablation) | Adversarial Poisoning · CIFAR-10 validation | accuracy (%): 52.97 ± 1.23 -> 10.98 ± 0.27 | Improved |
| Use class-targeted instead of untargeted poisons at ε=4/255 (ablation) | Adversarial Poisoning · CIFAR-10 validation | accuracy (%): 52.97 ± 1.23 -> 28.49 ± 0.65 | Improved |
| Increase PGD crafting steps from 50 to 250 (ablation) | Adversarial Poisoning · CIFAR-10 validation | accuracy (%): 16.53 ± 0.26 -> 6.25 ± 0.17 | Improved |
| Replace PGD with Carlini-Wagner crafting loss (ablation) | Adversarial Poisoning · CIFAR-10 validation | accuracy (%): 6.25 ± 0.17 -> 48.40 ± 3.24 | Hurt |
| Craft with an adversarially trained model (ablation) | Adversarial Poisoning · CIFAR-10 validation | accuracy (%): 6.25 ± 0.17 -> 81.79 ± 0.57 | Hurt |
| Relabel poisons using the crafting network's adversarial class (ablation) | Adversarially poisoned CIFAR-10 · CIFAR-10 validation | accuracy (%): 6.25 ± 0.17 -> 78.98 ± 0.25 | Improved |
| Train with adversarially perturbed SVHN images rather than CIFAR-10 poisons (ablation) | Corrected-label training · CIFAR-10 validation | accuracy (%): 78.98 ± 0.25 -> 30.19 ± 0.16 | Hurt |
| Apply adversarial training to the poisoned-data victim | Class-targeted adversarial poisoning · CIFAR-10 validation | accuracy (%): 6.25 -> 83.01 | Improved |
| Use Gaussian smoothing as a defense | Class-targeted adversarial poisoning · CIFAR-10 validation | accuracy (%): 11.94 | No clear effect |
Closest papers by meaning
- Witches' Brew: Industrial Scale Data Poisoning via Gradient Matching (ICLR 2021 · similarity 0.90)
- Attack To Defend: Exploiting Adversarial Attacks for Detecting Poisoned Models (CVPR 2024 · similarity 0.89)
- From Trojan Horses to Castle Walls: Unveiling Bilateral Data Poisoning Effects in Diffusion Models (NeurIPS 2024 · similarity 0.88)
- Machine Unlearning Fails to Remove Data Poisoning Attacks (ICLR 2025 · similarity 0.87)
- Adversarial-Inspired Backdoor Defense via Bridging Backdoor and Adversarial Attacks (AAAI 2025 · similarity 0.87)
- Adversarial Training for Defense Against Label Poisoning 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.
