Research map / Adversarial robustness and security / Backdoor and poisoning
The Victim and The Beneficiary: Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data
Zixuan Zhu, Rui Wang, Cong Zou, Lihua Jing
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 |
|---|---|---|---|
| add AttentionMix augmentation to filtered training (ablation) | V&B · CIFAR-10 Blend | ASR: 76.70% -> 1.42% | Improved |
| add AttentionMix augmentation to filtered training (ablation) | V&B · CIFAR-10 WaNet | ASR: 90.97% -> 25.16% | Improved |
| add AttentionMix augmentation (ablation) | V&B · CIFAR-10 | BA: 93.50% -> 92.27% (BadNets) | Hurt |
| add standard semi-supervised training on top of filter and AttentionMix (ablation) | V&B · CIFAR-10 WaNet | ASR: 25.16% -> 1.23% | Improved |
| add semi-supervised training without suppression (ablation) | V&B · CIFAR-10 BadNets | ASR: 0.42% -> 1.21% | Hurt |
| add suppression of suspicious samples with Victim predictions (ablation) | V&B · CIFAR-10 Dynamic | ASR: 28.79% -> 1.43% | Improved |
| use AttentionMix instead of Mixup, Cutmix or Attentive Cutmix (ablation) | V&B · CIFAR-10 WaNet/Dynamic | ASR: 13.38% (Attentive Cutmix) -> 1.13% on Dynamic | Improved |
| AttentionMix vs other augmentations on simple triggers (BadNets, Blend) (ablation) | V&B · CIFAR-10 BadNets | ASR: 0.27% (Mixup) vs 0.62% (AttentionMix) | No clear effect |
| increase mixing threshold t_m from 0.1 toward 0.5 | V&B · CIFAR-10 | BA | Hurt |
| increase mixing threshold t_m from 0.1 toward 0.5 | V&B · ImageNet subset | BA | No clear effect |
| apply V&B instead of no defense on ImageNet subset | ResNet-18 · ImageNet subset | BA: 94.94% -> 95.42% (BadNets) | Improved |
Closest papers by meaning
- A Theoretical Analysis of Backdoor Poisoning Attacks in Convolutional Neural Networks (ICML 2024 · similarity 0.89)
- Wicked Oddities: Selectively Poisoning for Effective Clean-Label Backdoor Attacks (ICLR 2025 · similarity 0.89)
- From Toxic to Trustworthy: Using Self-Distillation and Semi-supervised Methods to Refine Neural Networks (AAAI 2024 · similarity 0.88)
- BaDExpert: Extracting Backdoor Functionality for Accurate Backdoor Input Detection (ICLR 2024 · similarity 0.88)
- Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness (NeurIPS 2024 · similarity 0.88)
- Efficient Backdoor Attacks for Deep Neural Networks in Real-world Scenarios (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.
