Research map / Adversarial robustness and security / Backdoor and poisoning
Data-Free Backdoor Removal Based on Channel Lipschitzness
Runkai Zheng, Rongjun Tang, Jianze Li, Li Liu
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 |
|---|---|---|---|
| apply CLP (data-free channel pruning) to backdoored model | ResNet-18 · CIFAR-10 (average over 7 attacks) | ASR (%): 98.72 -> 2.81 | Improved |
| apply CLP to backdoored model | ResNet-18 · Tiny ImageNet (average over 7 attacks) | ASR (%): 97.47 -> 1.89 | Improved |
| use CLP instead of ANP with 500 benign samples | ANP · CIFAR-10 (average over 7 attacks) | ASR (%): 18.23 -> 2.81 | Improved |
| use CLP instead of ANP with 1,000 benign samples | ANP · Tiny ImageNet (average over 7 attacks) | ASR (%): 17.62 -> 1.89 | Improved |
| use CLP instead of NAD | NAD · CIFAR-10 (average over 7 attacks) | ACC (%): 92.78 -> 93.36 | Improved |
| use CLP instead of NAD | NAD · Tiny ImageNet (average over 7 attacks) | ASR (%): 62.13 -> 1.89 | Improved |
| use CLP instead of fine-tuning on CIFAR-10 for CLA attack | FT · CIFAR-10 (CLA) | ASR (%): 0.39 -> 2.84 | Hurt |
| use CLP instead of NAD on Blended | NAD · CIFAR-10 (Blended) | ASR (%): 1.61 -> 1.90 | Hurt |
| lower poisoning rate from 10% to 1% for CLA, then apply CLP (ablation) | CLP · CIFAR-10 (CLA) | ASR (%): 2.84 -> 10.04 | Hurt |
| vary poisoning rate (1%, 5%, 10%) before applying CLP (ablation) | CLP · CIFAR-10 (BadNets, Trojan, Blended, A2A) | ASR (%): 0.76 | No clear effect |
| lower pruning threshold u toward 0 (ablation) | CLP · CIFAR-10 (ResNet-18) | ASR and ACC | Mixed |
| use u=3 or 4 across ResNet/VGG/SENet architectures instead of per-architecture u (ablation) | CLP · CIFAR-10 (BadNets, various CNNs) | ASR and ACC | Mixed |
Closest papers by meaning
- Fusing Pruned and Backdoored Models: Optimal Transport-based Data-free Backdoor Mitigation (AAAI 2025 · similarity 0.86)
- DP-SGD Without Clipping: The Lipschitz Neural Network Way (ICLR 2024 · similarity 0.86)
- From Toxic to Trustworthy: Using Self-Distillation and Semi-supervised Methods to Refine Neural Networks (AAAI 2024 · similarity 0.85)
- ECLipsE: Efficient Compositional Lipschitz Constant Estimation for Deep Neural Networks (NeurIPS 2024 · similarity 0.85)
- Backdoor Mitigation via Invertible Pruning Masks (NeurIPS 2025 · similarity 0.84)
- Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness (NeurIPS 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.
