New research highlights loss function misalignment in AI robustness evaluations

Researchers have identified a misalignment between surrogate loss functions and the zero-one loss in Wasserstein distributionally robust optimization (WDRO). This misalignment can lead to hypotheses with perfect robust accuracy having higher WDRO risk than those with near-zero robust accuracy, particularly when using common loss functions like cross-entropy. To address this, the paper introduces Wasserstein distributional attacks (WDA), a novel method that optimizes the allocation of empirical mass under Wasserstein constraints to directly evaluate robust accuracy. Experiments on CIFAR-10, CIFAR-100, and ImageNet demonstrate that WDA achieves superior adversarial accuracy compared to existing methods. AI

IMPACT Introduces a new method for evaluating and improving model robustness against adversarial attacks.

RANK_REASON Academic paper detailing new theoretical findings and experimental methods in AI robustness. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Bach Le Tran
• CIFAR-10
• CIFAR-100
• cross entropy
• ImageNet
• Wasserstein distributional attacks
• Wasserstein distributionally robust optimization
• Wasserstein Robust Classification Error
• WDA
• WDRO
• zero-one loss

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →