Research map / Adversarial robustness and security / Backdoor and poisoning
Backdooring Self-Supervised Contrastive Learning by Noisy Alignment
Tuo Chen, Jie Gui, Minjing Dong, Ju Jia, Lanting Fang, Jian Liu
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 Noisy Alignment poisons instead of SSLBKD | MoCo v2 · ImageNet-100 | ASR (%): 50.9 -> 84.8 | Improved |
| use Noisy Alignment poisons instead of CorruptEncoder | SimSiam · ImageNet-100 | ASR (%): 26.1 -> 97.1 | Improved |
| use Noisy Alignment poisons instead of SSLBKD | SimCLR · CIFAR-10 | ASR (%): 69.2 -> 85.6 | Improved |
| use NA instead of BLTO | MoCo v2 · CIFAR-10 | ASR (%): 99.1 -> 75.9 | Hurt |
| add explicit noise compression via oracle-poisoning pairs (training-controlled) | NA · ImageNet-100 | ASR (%): 84.8 -> 97.3 | Improved |
| add blended trigger in place of patch trigger (ablation) | NA · ImageNet-100 | ASR (%): 82.3 -> 88.2 | Improved |
| defender raises minimal crop ratio to 0.8 (ablation) | NA · ImageNet-100 | ASR (%): 82.3 -> 0.5 | Hurt |
| defender removes random cropping (ablation) | NA · ImageNet-100 | ASR (%): 82.3 -> 0.9 | Hurt |
| increase number of attacked classes from two to four (ablation) | NA · ImageNet-100 (SimSiam) | ASR (%): 97.4 -> 92.7 | Hurt |
| use reference images from STL-10 instead of ImageNet-100-O (ablation) | NA · ImageNet-100 | ASR (%): 77.1 -> 59.0 | Hurt |
| train CLIP from scratch instead of fine-tuning (ablation) | NA · image-text CL | ASR (%): 100.0 -> 93.1 | Hurt |
| apply unsupervised distillation defense (ComPress) (ablation) | NA · ImageNet-100 | ASR (%): ASR below 1% | Hurt |
Closest papers by meaning
- Data Poisoning based Backdoor Attacks to Contrastive Learning (CVPR 2024 · similarity 0.89)
- Backdoor Contrastive Learning via Bi-level Trigger Optimization (ICLR 2024 · similarity 0.86)
- Invisible Backdoor Attack against Self-supervised Learning (CVPR 2025 · similarity 0.86)
- Semantic Shield: Defending Vision-Language Models Against Backdooring and Poisoning via Fine-grained Knowledge Alignment (CVPR 2024 · similarity 0.85)
- BadCLIP: Dual-Embedding Guided Backdoor Attack on Multimodal Contrastive Learning (CVPR 2024 · similarity 0.85)
- Clean Data Can Still Carry Backdoors: Support-Persistent Backdoors for Model Reuse (NeurIPS 2026 · similarity 0.85)
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.
