Research map / Adversarial robustness and security / Backdoor and poisoning
Reconstructive Neuron Pruning for Backdoor Defense
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 RNP instead of the strongest baseline NAD | backdoored ResNet-18 · CIFAR-10, 12 attacks (average) | ASR (%): 13.65 -> 5.03 | Helped |
| Apply RNP instead of ANP | backdoored ResNet-34 · ImageNet-12, 5 attacks (average) | ASR (%): 18.43 -> 8.87 | Helped |
| Add filter recovering after neuron unlearning (ablation) | RNP · CIFAR-10, BadNets | CA (%): 40.5 -> 92.22 | Helped |
| Add neuron unlearning before filter recovering (ablation) | filter recovering · CIFAR-10, BadNets | ASR (%): 100 -> 0.20 | Helped |
| Replace loss maximization with minimizing loss on incorrectly labeled clean samples (ablation) | RNP unlearning · CIFAR-10, BadNets | ASR (%): 0.20 -> 3.64 | Hurt |
| Use threshold-based pruning instead of fraction-based pruning (ablation) | RNP · CIFAR-10, BadNets | CA (%): 91.82 -> 92.22 | Helped |
| Restrict filter recovering to Block 4 rather than all blocks (ablation) | RNP · CIFAR-10, BadNets | ASR (%): 0.20 -> 6.50 | Hurt |
| Add fine-tuning after RNP pruning (ablation) | RNP · CIFAR-10, BadNets, poisoning rate 0.05% | CA (%): 56.32 -> 77.29 | Helped |
| Use more clean defense data for unlearning (ablation) | RNP · CIFAR-10, Dynamic and CL attacks | CA (%) | Hurt |
| Continue unlearning from epoch 10 to epoch 15 (ablation) | RNP · CIFAR-10, five tested attacks | CA (%) | Hurt |
Closest papers by meaning
- Fusing Pruned and Backdoored Models: Optimal Transport-based Data-free Backdoor Mitigation (AAAI 2025 · similarity 0.92)
- REFINE: Inversion-Free Backdoor Defense via Model Reprogramming (ICLR 2025 · similarity 0.90)
- Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness (NeurIPS 2024 · similarity 0.89)
- Backdoor Mitigation via Invertible Pruning Masks (NeurIPS 2025 · similarity 0.89)
- BAN: Detecting Backdoors Activated by Adversarial Neuron Noise (NeurIPS 2024 · similarity 0.89)
- Uncovering, Explaining, and Mitigating the Superficial Safety of Backdoor Defense (NeurIPS 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.