Research map / Adversarial robustness and security / Attacks on text and graphs
Exploring Model Dynamics for Accumulative Poisoning Discovery
Jianing Zhu, Xiawei Guo, Jiangchao Yao, Chao Du, LI He, Shuo Yuan, Tongliang Liu, Liang Wang, Bo Han
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 discrepancy-aware selective sample correction | ResNet-18 · CIFAR-10, CIFAR-100, and SVHN Accu. Poison | test accuracy + Trigger (%): CIFAR-10: 75.3±0.26 -> 77.3±0.58; SVHN: 88.7±0.20 -> 88.8±0.26; CIFAR-100: 44.4±0.21 -> 45.4±1.39 | Improved |
| Apply DSC rather than adversarial training to clean samples | ResNet-18 · CIFAR-10 Clean Oracle | test accuracy (%): 77.2 -> 84.7 | Improved |
| Use a later Epoch 30 auxiliary model without adjusting the discrepancy threshold (ablation) | DSC on ResNet-18 · CIFAR-10 Accu. Poison | test accuracy + Trigger (%): 77.3±0.58% -> 63.6±3.39% | Hurt |
| Adjust P when using the Epoch 30 auxiliary model (ablation) | DSC on ResNet-18 · CIFAR-10 Accu. Poison | test accuracy + Trigger (%): 63.6±3.39% -> 76.8±0.27% | Improved |
| Replace DSC correction with TRADES, MART, MMD, or FAT correction (ablation) | ResNet-18 · CIFAR-10 Accu. Poison | test accuracy + Trigger (%): DSC: 77.3±0.58%; TRADES: 72.5±0.45%; MART: 68.4±0.66%; MMD: 71.4±0.77%; FAT: 76.2±0.23% | Hurt |
| Use KL (TRADES) instead of PGD for the adversarial optimization loss (ablation) | DSC on ResNet-18 · CIFAR-10 Accu. Poison | test accuracy + Trigger (%): 78.7 -> 77.3 | Hurt |
| Use CW∞ instead of PGD for the adversarial optimization loss (ablation) | DSC on ResNet-18 · CIFAR-10 Accu. Poison | test accuracy + Trigger (%): 78.7 -> 79.3 | Improved |
| Increase the adaptive attack constraint β (ablation) | Accumulative poisoning attacker against ST · CIFAR-10 adaptive attack | test accuracy + Trigger (%): 51.3 -> 74.9 | Mixed |
| Use a black-box Epoch 10 surrogate rather than the white-box setting (ablation) | DSC on ResNet-18 · CIFAR-10 Accu. Poison | test accuracy + Trigger (%): 77.3±0.58% -> 79.7±0.03% | Improved |
Closest papers by meaning
- Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples (NeurIPS 2026 · similarity 0.85)
- Attack To Defend: Exploiting Adversarial Attacks for Detecting Poisoned Models (CVPR 2024 · similarity 0.85)
- From Trojan Horses to Castle Walls: Unveiling Bilateral Data Poisoning Effects in Diffusion Models (NeurIPS 2024 · similarity 0.85)
- Machine Unlearning Fails to Remove Data Poisoning Attacks (ICLR 2025 · similarity 0.84)
- Sharpness-Aware Data Poisoning Attack (ICLR 2024 · similarity 0.84)
- Demystifying Poisoning Backdoor Attacks from a Statistical Perspective (ICLR 2024 · 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.
