Research map / Adversarial robustness and security / Backdoor and poisoning
Effective Backdoor Defense by Exploiting Sensitivity of Poisoned Samples
Weixin Chen, Baoyuan Wu, Haoqian Wang
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 D-ST (FCT-based distinguishment + secure training) instead of DBD | secure training from scratch · CIFAR-10 avg | ASR (%): 96.40 -> 1.21 | Improved |
| use D-BR instead of ANP for backdoor removal | backdoor removal · CIFAR-10 avg | ACC (%): 84.91 -> 92.42 | Improved |
| semi-supervised contrastive learning (SS-CTL) instead of CTL for the feature extractor (ablation) | D-ST · CIFAR-10 BadNets-all2one | ACC (%): 85.63 -> 92.77 | Improved |
| SS-CTL instead of S-CTL for the feature extractor (ablation) | D-ST · CIFAR-10 SIG | ASR (%): 99.77 -> 0.00 | Improved |
| SS-CTL instead of S-CTL for the feature extractor (ablation) | D-ST · CIFAR-10 ACC across attacks | ACC (%): 94.51 -> 90.07 (SIG) | Hurt |
| replace FCT metric with Spectral signatures or DBD metric in SD module (ablation) | D-BR · CIFAR-10 BadNets, Blend, CL | ASR | Hurt |
| replace FCT metric with ABL loss metric (ablation) | D-BR · CIFAR-10 Blend | ASR (%): 24.72 -> <0.5 | Hurt |
| use unlearning+relearning instead of pure unlearning (ablation) | BR module · CIFAR-10 | ACC | Improved |
| use unlearning+relearning instead of pure relearning (ablation) | BR module · CIFAR-10 | ASR | Improved |
| add negative CE term on distinguished poisoned samples (lambda_p > 0) (ablation) | mixed cross-entropy loss · CIFAR-10 | ASR | Improved |
| increase lambda_p from 0.001 to 1 (ablation) | mixed cross-entropy loss · CIFAR-10 | ACC | Hurt |
Closest papers by meaning
- Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness (NeurIPS 2024 · similarity 0.89)
- Backdoor Secrets Unveiled: Identifying Backdoor Data with Optimized Scaled Prediction Consistency (ICLR 2024 · similarity 0.88)
- Activation Gradient based Poisoned Sample Detection Against Backdoor Attacks (ICLR 2025 · similarity 0.88)
- Lie Detector: Unified Backdoor Detection via Cross-Examination Framework (NeurIPS 2025 · similarity 0.88)
- RPP: A Certified Poisoned-Sample Detection Framework for Backdoor Attacks under Dataset Imbalance (NeurIPS 2026 · similarity 0.88)
- Demystifying Poisoning Backdoor Attacks from a Statistical Perspective (ICLR 2024 · similarity 0.88)
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.
