Reproduce “ADBM: Adversarial diffusion bridge model for reliable adversarial purification” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiao Li, Wenxuan Sun, Huanran Chen, Qiongxiu Li, Yining Liu, Yingzhe He, Jie Shi, Xiaolin Hu
The Adversarial Diffusion Bridge Model (ADBM), developed by researchers including those from Tsinghua University and Huawei Technologies, proposes a targeted diffusion model that directly learns to purify adversarial examples by bridging diffused adversarial data to clean data. This approach significantly enhances adversarial robustness against diverse attacks, achieving robustness gains of 4.4% on CIFAR-10 and 5.2% on SVHN over DiffPure against white-box attacks, while outperforming traditional adversarial training against unseen threats under a rigorously established evaluation framework.
Key insights
- Directly learning a diffusion bridge tailored for purification is more effective than repurposing general generative diffusion models, as it overcomes the fundamental trade-off between noise removal and data integrity.
- Reliable evaluation is paramount for adversarial robustness; previous assessments of diffusion-based purification methods were significantly flawed due to weak adaptive attacks.
- ADBM offers superior generalization against unseen norm-bounded attacks (e.g., L1, L2) and various black-box attacks compared to Adversarial Training, while also functioning as a highly efficient and transferable plug-and-play defense.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.