notesum.ai

Published at November 4

Alignment-Based Adversarial Training (ABAT) for Improving the Robustness and Accuracy of EEG-Based BCIs

cs.HC

cs.AI

cs.LG

Released Date: November 4, 2024

Authors: Xiaoqing Chen, Ziwei Wang, Dongrui Wu

Arxiv: http://arxiv.org/abs/2411.02094v1

Model	EA	Training	No	FGSM	FGSM	FGSM	PGD	PGD	PGD	AutoAttack	AutoAttack	AutoAttack	Avg.
Model	EA	Training	Attack	0.01	0.03	0.05	0.01	0.03	0.05	0.01	0.03	0.05	Avg.
EEGNet	w/o EA	BT	60.78	36.14	8.89	1.66	35.92	7.70	1.05	35.64	7.32	0.94	19.60
		AT-FGSM 0.01	56.71	48.95	34.10	22.61	49.02	33.74	21.63	48.87	33.51	21.17	37.03
		AT-PGD 0.01	55.99	48.33	34.25	22.31	48.47	34.03	21.37	48.28	33.71	20.95	36.77
	with EA	BT	69.79	57.54	35.11	19.98	57.50	34.74	19.01	57.46	34.49	18.45	40.41
		ABAT-FGSM 0.01	71.08	64.11	49.52	35.65	64.12	49.28	35.11	64.08	49.10	34.70	51.67
		ABAT-FGSM 0.03	68.92	65.33	56.71	48.25	65.33	56.61	48.11	65.33	56.58	47.90	57.91
		ABAT-FGSM 0.05	65.46	62.91	57.29	51.34	62.92	57.23	51.35	62.91	57.20	51.16	57.98
		ABAT-PGD 0.01	73.28	66.82	51.53	37.49	66.80	51.29	36.70	66.78	51.09	36.34	53.81
		ABAT-PGD 0.03	70.69	66.85	58.81	49.76	66.85	58.83	49.50	66.85	58.74	49.31	59.62
		ABAT-PGD 0.05	67.48	64.75	59.26	53.16	64.75	59.25	53.15	64.74	59.19	53.00	59.87
DeepCNN	w/o EA	BT	54.06	31.61	6.93	1.02	31.29	6.30	0.82	31.08	5.97	0.68	16.98
		AT-FGSM 0.01	49.19	41.82	27.87	16.23	41.74	27.47	15.65	41.69	27.22	15.30	30.42
		AT-PGD 0.01	49.06	41.53	27.69	16.54	41.51	27.40	15.95	41.49	27.26	15.63	30.41
	with EA	BT	59.70	44.21	20.41	7.70	43.98	19.43	6.42	43.84	18.88	5.93	27.05
		ABAT-FGSM 0.01	62.73	53.20	35.08	21.58	53.15	34.57	20.78	53.15	34.32	20.37	38.89
		ABAT-FGSM 0.03	66.05	58.82	44.71	32.33	58.82	44.55	31.71	58.80	44.37	31.44	47.16
		ABAT-FGSM 0.05	63.88	59.08	48.74	38.71	59.08	48.66	38.49	59.08	48.60	38.28	50.26
		ABAT-PGD 0.01	61.54	52.28	34.23	21.14	52.20	33.77	20.22	52.17	33.56	19.88	38.10
		ABAT-PGD 0.03	64.98	58.64	44.82	32.72	58.60	44.68	32.19	58.59	44.60	31.88	47.17
		ABAT-PGD 0.05	62.83	58.00	48.01	38.25	58.00	47.92	37.96	57.97	47.81	37.80	49.45
ShallowCNN	w/o EA	BT	60.57	31.92	7.27	1.08	31.60	6.11	0.73	31.48	5.80	0.57	17.71
		AT-FGSM 0.01	57.47	47.15	29.57	16.87	47.13	29.37	16.00	47.08	29.05	15.46	33.51
		AT-PGD 0.01	57.47	47.49	29.73	16.80	47.48	29.64	16.06	47.44	29.24	15.44	33.68
	with EA	BT	71.46	57.01	34.40	19.24	56.94	33.90	18.24	56.91	33.74	17.75	39.96
		ABAT-FGSM 0.01	73.82	65.41	48.83	34.45	65.35	48.68	33.90	65.34	48.61	33.71	51.81
		ABAT-FGSM 0.03	74.79	68.58	54.99	42.48	68.57	54.85	42.21	68.56	54.81	42.08	57.19
		ABAT-FGSM 0.05	73.88	68.36	57.42	45.88	68.35	57.39	45.82	68.35	57.38	45.64	58.85
		ABAT-PGD 0.01	73.95	65.61	48.70	34.30	65.59	48.47	33.69	65.59	48.34	33.38	51.76
		ABAT-PGD 0.03	74.99	68.85	55.65	42.26	68.85	55.56	41.96	68.84	55.52	41.67	57.41
		ABAT-PGD 0.05	73.51	68.17	56.97	45.74	68.16	56.94	45.63	68.16	56.93	45.46	58.57