Table 7.

Results of the adversarial attack mitigation technique (before and after) on multiple datasets.

Attack type	Dataset	Movie reviews		Hate speech		Clickbait
	Attack	Before	After	Before	After	Before	After
Word-level	TextFloor	0.27	0.79	0.32	0.68	0.65	0.85
	PWWS	0.21	0.58	0.35	0.60	0.66	0.80
	GENETIC	0.15	0.63	0.23	0.58	0.58	0.76
	SememePSO	0.40	0.76	0.67	0.79	0.79	0.90
	BAE	0.85	0.99	0.94	0.96	0.98	0.97
	BERT-ATTACK	0.8	0.71	0.24	0.75	0.52	0.78
	HotFlip	0.47	0.84	0.71	0.88	0.81	0.85
Sentence-level	SEA	1	1	1	1	1	1
	GAN	0.63	0.89	0.60	0.66	0.54	0.56
	SCPN	1	1	1	1	1	1
Char-level	DeepWordBug	0.58	0.92	0.74	0.83	0.62	0.85
	VIPER	0.94	0.93	0.99	0.97	0.98	0.97
	UAT	0.11	0.74	0.18	0.35	0.39	0.36
	TextBugger	0.5	0.64	0.25	0.65	0.61	0.74
Average		0.565	0.815	0.587	0.764	0.723	0.813
Relative $\downarrow$↑%		↑25%		↑17%		↑9%

Significant values are in bold.