notesum.ai

Published at October 21

Conflict-Aware Adversarial Training

cs.CV

cs.AI

cs.LG

Released Date: October 21, 2024

Authors: Zhiyu Xue¹, Haohan Wang², Yao Qin¹, Ramtin Pedarsani¹

Aff.: ¹UC Santa Barbara; ²University of Illinois Urbana-Champaign

Arxiv: https://arxiv.org/abs/2410.16579v1

Refer to caption

		Standard Accuracy		PGD		AutoPGD		MIFGSM		FAB		T-FAB		FGSM
	$p=\infty$	CA-AT	Vanilla AT	CA-AT	Vanilla AT	CA-AT	Vanilla AT	CA-AT	Vanilla AT	CA-AT	Vanilla AT	CA-AT	Vanilla AT	CA-AT	Vanilla AT
ResNet18	8/255	0.8659	0.8239	0.7442	0.4703	0.6301	0.3996	0.7419	0.4745	0.8177	0.809	0.6861	0.6538	0.7649	0.519
	16/255			0.7311	0.4248	0.5555	0.2486	0.7233	0.4225	0.7475	0.78	0.5445	0.5104	0.7435	0.4387
	24/255			0.7189	0.405	0.4886	0.1963	0.7182	0.413	0.6858	0.7333	0.4599	0.4783	0.7235	0.403
	32/255			0.7033	0.3877	0.4455	0.1589	0.7182	0.413	0.6402	0.6836	0.4044	0.4507	0.7066	0.379
		Standard Accuracy		PGD		AutoPGD		MIFGSM		FAB		T-FAB		FGSM
	$p=\infty$	CA-AT	Vanilla AT	CA-AT	Vanilla AT	CA-AT	Vanilla AT	CA-AT	Vanilla AT	CA-AT	Vanilla AT	CA-AT	Vanilla AT	CA-AT	Vanilla AT
ResNet34	8/255	0.8753	0.8305	0.8098	0.5973	0.7285	0.4417	0.8111	0.5983	0.8247	0.8068	0.7274	0.6951	0.8149	0.5327
	16/255			0.8034	0.5756	0.6793	0.3395	0.8077	0.5791	0.7738	0.7613	0.6013	0.5762	0.7916	0.2762
	24/255			0.7957	0.5602	0.6445	0.2859	0.8067	0.5743	0.7307	0.6937	0.5142	0.5174	0.7743	0.1428
	32/255			0.785	0.5443	0.6165	0.2498	0.8067	0.5743	0.6918	0.6221	0.4424	0.4822	0.7616	0.088