Research map / Label-efficient and robust learning / Self-supervised and contrastive
SSL-Cleanse: Trojan Detection and Mitigation in Self-Supervised Learning
Mengxin Zheng, Jiaqi Xue, Zihao Wang, Xun Chen, Qian Lou, Lei Jiang, Xiaofeng Wang
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 |
|---|---|---|---|
| replace Kneedle with Sliding Window Kneedle for cluster number estimation (ablation) | SSL-Cleanse detector · ImageNet-100 | DACC: predicted K 20 -> 100; 30% DACC improvement (80%) | Improved |
| increase fraction of unlabeled data used for detection from 5% to 10% (ablation) | SSL-Cleanse detector · CIFAR-10 (CTRL, BYOL) | DACC: 71 -> 80 | Improved |
| increase fraction of unlabeled data used for detection from 5% to 10% (ablation) | SSL-Cleanse detector · ImageNet-100 (CTRL, BYOL) | DACC: 82 -> 85 | Improved |
| apply SCU mitigation to backdoored encoders | Trojaned SSL encoders · CIFAR-10 SSL-Backdoor BYOL | ASR (%): 48.32 -> 1.14 | Improved |
| apply SCU mitigation to backdoored encoders | Trojaned SSL encoders · CIFAR-10 CTRL SimCLR | ASR (%): 81.84 -> 1.15 | Improved |
| apply SCU mitigation to backdoored encoders | Trojaned SSL encoders · ImageNet-100 SSL-Backdoor SimCLR | clean ACC (%): 60.18 -> 58.58 | Hurt |
| apply SCU mitigation to encoders for other downstream tasks | Trojaned SSL encoders · Downstream STL-10 | ASR (%): 36.55 -> 1.84 | Improved |
| apply SCU mitigation to encoders for other downstream tasks | Trojaned SSL encoders · Downstream GTSRB | ASR (%): 38.28 -> 1.03 | Improved |
| use NC or ABS linear-probe detection when downstream data differs from encoder data | NC and ABS detectors · SSL-Backdoor, STL-10 downstream | NC Anomaly Index / ABS REASR: 2.05 -> 1.42 (NC); 0.89 -> 0.34 (ABS) | Hurt |
Closest papers by meaning
- DeDe: Detecting Backdoor Samples for SSL Encoders via Decoders (CVPR 2025 · similarity 0.90)
- Invisible Backdoor Attack against Self-supervised Learning (CVPR 2025 · similarity 0.87)
- TrojanDec: Data-free Detection of Trojan Inputs in Self-supervised Learning (AAAI 2025 · similarity 0.87)
- Stealthy Backdoor Attack in Self-Supervised Learning Vision Encoders for Large Vision Language Models (CVPR 2025 · similarity 0.83)
- HPE: Hallucinated Positive Entanglement for Backdoor Attacks in Federated Self-Supervised Learning (NeurIPS 2026 · similarity 0.82)
- Views Can Be Deceiving: Improved SSL Through Feature Space Augmentation (ICLR 2024 · similarity 0.81)
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.
