TADW: Traceable and Anti-detection Dynamic Watermarking of Deep Neural Networks

Table 3

Testing accuracy on clean models and watermarked models.

Model	SST-2			AG-News
	TextCNN (%)	TextRNN (%)	BERT (%)	TextCNN (%)	TextRNN (%)	BERT (%)
Clean	89.07	88.19	91.49	93.88	93.30	94.25
SN-10-1	88.58	87.75	91.05	93.66	93.87	93.87
SN-10-2	88.58	87.75	91.10	93.55	92.99	93.87
SN-10-3	88.91	87.75	91.05	93.43	92.97	93.79
SN-10-4	88.58	87.70	91.05	93.39	92.80	94.01
SN-10-5	88.63	87.70	91.27	93.63	92.84	93.78
SN-10-6	88.85	87.70	91.05	93.45	92.83	93.84
SN-10-7	88.58	88.25	91.21	93.39	92.86	94.11
SN-10-8	88.58	87.70	91.65	93.38	92.84	94.01
SN-10-9	88.58	87.70	91.32	93.39	92.80	94.17
SN-10-10	88.63	87.81	91.27	93.47	92.87	93.92