几何不等于鲁棒性：PGD评估的轨迹级研究

AI 摘要

本研究在Fashion-MNIST数据集上对卷积神经网络进行20步PGD攻击的轨迹级分析，比较了干净训练和对抗训练模型的鲁棒性。研究发现，损失轨迹和梯度对齐模式在鲁棒性不同的对抗训练模型间相似，而失败步数分布能更清晰地区分鲁棒性水平。作者认为轨迹级指标描述优化几何特性，但不能独立衡量对抗鲁棒性，应作为补充诊断工具。

正文节选

Geometry Is Not Robustness: A Trajectory-Level Study of PGD Evaluation Abstract Projected Gradient Descent (PGD) is widely used as a standard adversarial attack for evaluating the robustness of deep learning models. Robustness is typically assessed through final adversarial accuracy, which does not capture the dynamic behaviour of models throughout the attack process. Recent work has proposed trajectory-level diagnostics – such as loss evolution, gradient alignment, and steps-to-failure – to pro

发布时间：2026-08-19 12:00

抓取时间：2026-08-18 12:02

来源机构：arXiv