Research map / Label-efficient and robust learning / Semi-supervised learning
Backdoor Defense via Decoupling the Training Process
Kunzhe Huang, Yiming Li, Baoyuan Wu, Zhan Qin, Kui Ren
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 |
|---|---|---|---|
| Decouple training with self-supervised backbone learning, followed by classifier training and semi-supervised fine-tuning | ResNet-18 · CIFAR-10, BadNets | ASR: 100 -> 0.96 | Improved |
| Decouple training with self-supervised backbone learning, followed by classifier training and semi-supervised fine-tuning | ResNet-18 · ImageNet, BadNets | ASR: 90.49 -> 0.26 | Improved |
| Replace the self-supervised backbone with a supervised-trained backbone (ablation) | DBD · CIFAR-10, BadNets | ASR: 0.96 -> 100 | Hurt |
| Use SCE rather than CE for second-stage classifier training (ablation) | Self-supervised backbone pipeline · CIFAR-10, Blended | ASR: 12.19 -> 6.24 | Improved |
| Use SCE rather than CE for second-stage classifier training (ablation) | Self-supervised backbone pipeline · CIFAR-10, BadNets | ASR: 5.20 -> 5.12 | No clear effect |
| Filter out low-credible samples before tuning FC layers (ablation) | SS with SCE · CIFAR-10, BadNets | ASR: 5.12 -> 4.02 | Improved |
| Filter out low-credible samples before tuning FC layers (ablation) | SS with SCE · CIFAR-10, BadNets | BA: 82.34 -> 78.94 | Hurt |
| Fine-tune the whole model semi-supervised rather than only tuning FC layers on high-credible samples (ablation) | SS with SCE + Tuning · CIFAR-10, BadNets | ASR: 4.02 -> 0.96 | Improved |
| Fine-tune the whole model semi-supervised rather than only tuning FC layers on high-credible samples (ablation) | SS with SCE + Tuning · CIFAR-10, BadNets | BA: 78.94 -> 92.41 | Improved |
| Increase WaNet kernel size from 4 to 224 (ablation) | WaNet on ImageNet · ImageNet, WaNet | ASR: 6.87 -> 96.22 | Improved |
| Increase WaNet strength from 0.4 to 1 (ablation) | WaNet on ImageNet · ImageNet, WaNet | ASR: 80.06 -> 96.22 | Improved |
| Remove WaNet noise mode (ablation) | WaNet on ImageNet · ImageNet, WaNet | ASR: 31.72 -> 96.22 | Improved |
Closest papers by meaning
- Wicked Oddities: Selectively Poisoning for Effective Clean-Label Backdoor Attacks (ICLR 2025 · similarity 0.91)
- Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness (NeurIPS 2024 · similarity 0.91)
- A Theoretical Analysis of Backdoor Poisoning Attacks in Convolutional Neural Networks (ICML 2024 · similarity 0.90)
- Backdoor Defense via Test-Time Detecting and Repairing (CVPR 2024 · similarity 0.90)
- BaDExpert: Extracting Backdoor Functionality for Accurate Backdoor Input Detection (ICLR 2024 · similarity 0.90)
- Adversarial-Inspired Backdoor Defense via Bridging Backdoor and Adversarial Attacks (AAAI 2025 · similarity 0.90)
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.
