Medical Specialty Classification Based on Semiadversarial Data Augmentation

Table 1

The performance of different attacks.

Attack	Accuracy	Time (minutes)	#Query
None	73.3	—	—
Textfooler [43]	20.1	47 mim	1207.12
Deepwordbug [44]	37.4	33 min	921.78
Textbugger [45]	40.6	40 min	947.50
BERT-attack [37]	17.2	67 min	1089.72
MSAA	40.5	12 min	240.91

The victim model is BioBERT trained on the medical specialty classification dataset. Time is execution times for each method that attacks BioBERT based on 1000 test examples by the plain way and the proposed method. #Query is the number of queries that methods require for the victim model. To make a fair comparison, MSAA is simplified by SemiADA that just generates adversarial examples but does not save the intermediate perturbed examples.