Research map / Label-efficient and robust learning / Semi-supervised learning
The Perils of Learning From Unlabeled Data: Backdoor Attacks on Semi-supervised Learning
Virat Shejwalkar, Lingjuan Lyu, Amir Houmansadr
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 clean-label static full-image repetitive trigger instead of DeHiB/Narcissus/CL-Badnets | SSL backdoor poisoning (FixMatch) · CIFAR10 (0.2% poison) | ASR (%): 1.1 (DeHiB) -> 88.1 | Improved |
| use dirty-label adversarial-perturbation poisoning (DeHiB) | SSL backdoor attack · CIFAR10 | ASR (%): 0.9 - 1.2 (DeHiB without D^l) | Hurt |
| use small trigger patch (CL-Badnets) instead of full-image trigger | SSL backdoor attack · CIFAR100 / CIFAR10 | ASR (%): ASR stays below 35% | Hurt |
| apply strong augmentation to supervised training (ablation) | CL-Badnets · CIFAR10 supervised | ASR (%): 83.4 -> 0.0 | Hurt |
| apply strong augmentation to supervised training (ablation) | Our attack · CIFAR10 supervised | ASR (%): 99.8 -> 88.9 | Hurt |
| apply strong augmentation to supervised training (ablation) | Narcissus · CIFAR10 supervised | ASR (%): 100.0 -> 99.5 | No clear effect |
| add the backdoor trigger to target-class unlabeled data | FixMatch · CIFAR10 | clean accuracy (%): 91.1 -> 93.5 | Improved |
| vary poison size outside 0.2-0.4% of training data (ablation) | Our attack · CIFAR10/SVHN ablation | ASR (%): ASR > 60% only within 0.2-0.4% | Hurt |
| increase labeled data size (ablation) | Our attack · CIFAR10 / SVHN | ASR (%): ASRs remain above 70% | Mixed |
| combine Narcissus trigger with static trigger (ablation) | Our attack · SSL backdoor | ASR | Hurt |
| duplicate poisoning data K times (ablation) | Our attack · CIFAR10 UDA / FlexMatch | ASR (%): 81.5 -> 84.3 (UDA), 87.9 -> 89.1 (FlexMatch) | Mixed |
| use interpolation-based poisoning (ablation) | Our attack · CIFAR10 | ASR | Hurt |
Closest papers by meaning
- Backdoor Secrets Unveiled: Identifying Backdoor Data with Optimized Scaled Prediction Consistency (ICLR 2024 · similarity 0.90)
- Backdoor Attacks via Machine Unlearning (AAAI 2024 · similarity 0.88)
- Invisible Backdoor Attack against Self-supervised Learning (CVPR 2025 · similarity 0.88)
- Universal Backdoor Attacks (ICLR 2024 · similarity 0.87)
- Demystifying Poisoning Backdoor Attacks from a Statistical Perspective (ICLR 2024 · similarity 0.87)
- From Toxic to Trustworthy: Using Self-Distillation and Semi-supervised Methods to Refine Neural Networks (AAAI 2024 · similarity 0.87)
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.
