537d9b6c927223c796cac288cced29df-AuthorFeedback.pdf

CATrequires almostthesamenumber ofepochsto15 converge compared with baseline PGD-AT,and thus almost identical training time. One application of CAT is to sketch (approximately estimate) the empirical achievable trade-off between accuracy27 and robustness of the same model trained under differentλ values, in a cheap and efficient way (avoiding training28 manytimes).