Table 8.

Performance comparison of various defense techniques against four adversarial attacks. The bold text indicates improvements in the results, the underlined text indicates the second-best results, and the italics text signifies the worst value.

Attack	Dataset	DT	ADP	Mixup	AdvT	ScRNN	SHIELD	Our method
TextFloor	MR	0.09	0.11	0.13	0.10	0.02	0.51	0.79
	HS	0.25	0.33	0.46	0.40	0.15	0.53	0.68
	CB	0.35	0.50	0.38	0.39	0.40	0.78	0.85
TextBugger	MR	0.15	0.22	0.11	0.20	0.11	0.43	0.64
	HS	0.33	0.36	0.33	0.54	0.30	0.46	0.65
	CB	0.45	0.52	0.63	0.55	0.62	0.77	0.74
DeepWordBug	MR	0.18	0.11	0.05	0.13	0.09	0.61	0.92
	HS	0.58	0.31	0.39	0.45	0.25	0.73	0.82
	CB	0.54	0.49	0.55	0.47	0.36	0.74	0.85
PWWS	MR	0.17	0.14	0.12	0.3	0.13	0.45	0.58
	HS	0.27	0.35	0.55	0.40	0.19	0.58	0.60
	CB	0.45	0.49	0.47	0.60	0.43	0.73	0.80
AVG		0.385	0.327	0.347	0.344	0.254	0.61	0.74