Catastrophic Overfitting, Entropy Gap and Participation Ratio: A Noiseless $l^p$ Norm Solution for Fast Adversarial Training

In three linesarXiv paper addressing Catastrophic Overfitting (CO) in fast adversarial training. Authors propose controlling the lp training norm instead of noise injection or regularization. They quantify gradient concentration via Participation Ratio and entropy measures, developing an adaptive lp-FGSM that automatically tunes the training norm based on gradient information.

Summary generated by Claude — human-verified