Research map / Adversarial robustness and security / Backdoor and poisoning
Anti-Backdoor Coreset Selection via Cumulative Entropy
Qi Zhao, Christian Wressnegger
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 |
|---|---|---|---|
| accumulate entropy over epochs (Cumulative Entropy) instead of epoch-wise entropy sampling (ablation) | uncertainty-based coreset selection · CIFAR10 Blend, ResNet18 | poisoning rate of coreset | Improved |
| use Cumulative Entropy rather than cumulative Margin or LeastConf (ablation) | anti-backdoor coreset selection · CIFAR10 Blend, ResNet18 | ACC, ASR, poisoning rate | Improved |
| add min-max normalization of per-epoch entropy in CENT (ablation) | ABCS · CIFAR10 WaNet, selection ratio 0.4 | poisoning rate of coreset (%): 0.62 -> 0.50 | Improved |
| add warm-up phase before selection (ablation) | CENT · CIFAR10 WaNet | poisoning rate of coreset | Improved |
| add unlearning step with label smoothing (ablation) | CENT + warm-up · CIFAR10 WaNet | poisoning rate of coreset | Improved |
| use a very small label smoothing factor epsilon (ablation) | ABCS · CIFAR10 WaNet | ASR | Hurt |
| vary label smoothing factor epsilon (ablation) | ABCS · CIFAR10 | natural accuracy | No clear effect |
| exclude mis-predicted samples when computing the threshold (ablation) | ABCS · Tiny-ImageNet Blend | natural accuracy (%): 38.53 -> 61.11 | Improved |
| use a short warm-up (Twa) (ablation) | ABCS · CIFAR10 | ASR | Hurt |
| use a larger regularization gamma (>=0.1) (ablation) | ABCS · CIFAR10 | ASR | Improved |
| use epoch-wise EL2N instead of cumulative EL2N (ablation) | EL2N coreset selection · CIFAR10 Blend | poisoning rate of coreset | Hurt |
| use ABCS on clean datasets | ABCS vs naive training · CIFAR10 clean | natural accuracy (%): 94.64 -> 94.77 | No clear effect |
Closest papers by meaning
- Anti-Backdoor Learning: Training Clean Models on Poisoned Data (NeurIPS 2021 · similarity 0.90)
- Two Sides of the Same Coin: Learning the Backdoor to Remove the Backdoor (AAAI 2025 · similarity 0.90)
- Training with More Confidence: Mitigating Injected and Natural Backdoors During Training (NeurIPS 2022 · similarity 0.90)
- Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness (NeurIPS 2024 · similarity 0.89)
- Demystifying Poisoning Backdoor Attacks from a Statistical Perspective (ICLR 2024 · similarity 0.88)
- Inspecting Prediction Confidence for Detecting Black-Box Backdoor Attacks (AAAI 2024 · 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.
