Reproduce “SoK: Certified Robustness for Deep Neural Networks” on OpenResearch

About this paper

Linyi Li, Tao Xie, Bo Li

Researchers from the University of Illinois Urbana-Champaign and Peking University systematically categorized and benchmarked certifiably robust approaches for Deep Neural Networks against adversarial evasion attacks. The work provides the first comprehensive taxonomy and an open-source platform, demonstrating the trade-off between scalability and tightness in verification methods and identifying the current state-of-the-art certified accuracies across various datasets.

Key insights

• A fundamental trade-off exists between the scalability and tightness of certified robustness methods; complete verification offers exact guarantees but has exponential complexity, while incomplete and probabilistic methods offer better scalability with looser bounds.
• Certified robust training approaches significantly boost certifiable accuracy, but their effectiveness often depends on being verified by a compatible verification method.
• Probabilistic verification methods based on smoothed DNNs are currently the only approaches scalable enough for large-scale datasets like ImageNet, though they struggle with tight $\mathcal{L}_\infty$ robustness in high dimensions.

Summary by alphaXiv.