Research map / Adversarial robustness and security / Backdoor and poisoning
Progressive Backdoor Erasing via Connecting Backdoor and Adversarial Attacks
Bingxu Mu, Zhenxing Niu, Le Wang, Xue Wang, Qiguang Miao, Rong Jin, Gang Hua
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 progressive adversarial fine-tuning and clean-image filtering to erase the backdoor | PBE · CIFAR-10, Badnet, all-to-one | ASR: 100.00 -> 1.09 | Improved |
| Use PBE with a clean extra dataset rather than PBE without one | PBE · CIFAR-10, Badnet, all-to-one | ASR: 11.30 -> 1.09 | Improved |
| Use PBE without a clean extra dataset | PBE · CIFAR-10, Badnet, all-to-one | ASR: 100.00 -> 11.30 | Improved |
| Use PBE without a clean extra dataset | PBE · CIFAR-10, Blend, all-to-one | ASR: 100.00 -> 1.16 | Improved |
| Use PBE without a clean extra dataset | PBE · CIFAR-10, Badnet, all-to-one | ASR: 100.00 -> 11.30 | Mixed |
| Use PBE with a clean extra dataset | PBE · CIFAR-10, WaNet, all-to-one | ASR: 99.50 -> 0.46 | Improved |
| Use PBE with a clean extra dataset | PBE · GTSRB, Badnet, all-to-one | ASR: 100.00 -> 0.47 | Improved |
| Use PBE with a clean extra dataset | PBE · Tiny-ImageNet, WaNet, all-to-one | ASR: 99.92 -> 5.32 | Improved |
| Use PBE with a clean extra dataset | PBE · CIFAR-10, Badnet, all-to-all | ASR: 94.41 -> 0.62 | Improved |
| Fine-tune on adversarial examples generated by untargeted attacks | Infected model · Backdoor attacks across tested datasets | ASR | Improved |
| Progressively update the extra dataset by ranking training images against the infected and purified models | PBE · CIFAR-10 Blend attack | Average precision: 0.25 -> 0.93 | Improved |
| Plant a backdoor in the model | Classifier adversarial examples · CIFAR-10, five backdoor attacks | Feature L2 distance to triggered-image features: 102.58 -> 85.11; 135.91 -> 78.18; 124.22 -> 75.09; 40.42 -> 28.66; 48.13 -> 15.85 | Improved |
Closest papers by meaning
- Towards Unified Robustness Against Both Backdoor and Adversarial Attacks (TPAMI 2024 · similarity 0.93)
- Breaking the False Sense of Security in Backdoor Defense through Re-Activation Attack (NeurIPS 2024 · similarity 0.92)
- Adversarial-Inspired Backdoor Defense via Bridging Backdoor and Adversarial Attacks (AAAI 2025 · similarity 0.90)
- 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 0.89)
- Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness (NeurIPS 2024 · similarity 0.89)
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.
