Perturbing BEAMs: EEG adversarial attack to deep learning models for epilepsy diagnosing

Table 9 Performance of the modified methods in the case of maintaining the same \({DL}_{E}\) as Table 5. The parameter N for the modified GPBEAM-DE is set to be 5

Architecture				Maxpool		Temporal convolution		LSTM		Mixed LSTM
\({\mathrm{DL}}_{E}\)	\({\epsilon }_{B}\)	\({\epsilon }_{E}\)	Evaluation Criteria	Acc	SR	Acc	SR	Acc	SR	Acc	SR
EEG-related models
0	-	-	-	0.92	-	0.84	-	0.90	-	0.88	-
0.024	0.1	0.023	modified GPBEAM	0.42	0.49	0.28	0.55	0.71	0.18	0.75	0.14
0.021	0.1	0.020	modified GPBEAM-DE	0.47	0.41	0.34	0.50	0.77	0.15	0.77	0.11
0.030	0.3	0.029	modified GPBEAM	0.36	0.55	0.23	0.61	0.68	0.21	0.73	0.16
0.026	0.3	0.025	modified GPBEAM-DE	0.40	0.51	0.26	0.58	0.70	0.20	0.75	0.14
0.036	0.5	0.035	modified GPBEAM	0.34	0.58	0.19	0.64	0.67	0.23	0.69	0.19
0.030	0.5	0.029	modified GPBEAM-DE	0.36	0.55	0.23	0.61	0.68	0.21	0.71	0.17
BEAM-related models
0	-	-	-	0.92	-	0.92	-	0.93	-	0.94	-
0.024	0.1	0.023	modified GPBEAM	0.43	0.50	0.37	0.55	0.76	0.18	0.73	0.20
0.021	0.1	0.020	modified GPBEAM-DE	0.61	0.29	0.51	0.41	0.58	0.37	0.64	0.31
0.030	0.3	0.029	modified GPBEAM	0.18	0.75	0.17	0.75	0.61	0.33	0.61	0.32
0.026	0.3	0.025	modified GPBEAM-DE	0.28	0.67	0.25	0.71	0.26	0.70	0.25	0.70
0.036	0.5	0.035	modified GPBEAM	0.12	0.75	0.12	0.80	0.58	0.36	0.56	0.37
0.030	0.5	0.029	modified GPBEAM-DE	0.14	0.78	0.13	0.80	0.17	0.79	0.15	0.80

ISSN: 1472-6947