Research map / Adversarial robustness and security / Attacks on text and graphs
Autoregressive Perturbations for Data Poisoning
Pedro Sandoval-Segura, Vasu Singla, Jonas Geiping, Micah Goldblum, Tom Goldstein, David W. Jacobs
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 sample-wise autoregressive perturbations instead of the strongest listed baseline poison | ResNet-18 trained on CIFAR-10 · CIFAR-10 | test accuracy: 16.84 -> 11.75 | Improved |
| Use the same AR processes across datasets instead of the strongest listed baseline poison | ResNet-18 poison training · STL-10 | test accuracy: 17.26 -> 11.65 | Improved |
| Use sample-wise AR poisoning instead of the strongest listed baseline poison | ResNet-18 poison training · SVHN | test accuracy: 6.39 -> 6.77 | Hurt |
| Use AR poisoning instead of the strongest listed baseline poison | ResNet-18 poison training · CIFAR-100 | test accuracy: 2.99 -> 4.24 | Hurt |
| Use AR poisoning instead of the strongest listed baseline poison | CIFAR-10 models across architectures · CIFAR-10 | test accuracy: 16.84 -> 11.75; 9.96 -> 14.90 | Mixed |
| Add Cutout, CutMix, or Mixup during training | ResNet-18 trained on AR-poisoned CIFAR-10, epsilon=1 · CIFAR-10 | test accuracy: 11.75 -> 11.90; 11.75 -> 11.23; 11.75 -> 11.40 | No clear effect |
| Reduce AR perturbation size from epsilon=1 to epsilon=0.5 | ResNet-18 poison training with standard augmentation · CIFAR-10 | test accuracy: 11.75 -> 14.28 | Hurt |
| Increase adversarial-training radius from 0.125 to 0.25 | ResNet-18 trained on AR-poisoned CIFAR-10 · CIFAR-10 | test accuracy: $33.22_{\pm0.77}$ -> $57.08_{\pm0.75}$ | Hurt |
| Increase adversarial-training radius from 0.25 to 0.50 | ResNet-18 trained on AR-poisoned CIFAR-10 · CIFAR-10 | test accuracy: $57.08_{\pm0.75}$ -> $81.27_{\pm2.61}$ | Hurt |
Closest papers by meaning
- From Trojan Horses to Castle Walls: Unveiling Bilateral Data Poisoning Effects in Diffusion Models (NeurIPS 2024 · similarity 0.86)
- DataElixir: Purifying Poisoned Dataset to Mitigate Backdoor Attacks via Diffusion Models (AAAI 2024 · similarity 0.85)
- Attack To Defend: Exploiting Adversarial Attacks for Detecting Poisoned Models (CVPR 2024 · similarity 0.85)
- Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples (NeurIPS 2026 · similarity 0.85)
- Universal Backdoor Attacks (ICLR 2024 · similarity 0.85)
- Phantom Transfer: Data Poisoning can Survive Data-Level Defences (NeurIPS 2026 · similarity 0.84)
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.
