Research Article

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

Table 3

Testing accuracy on clean models and watermarked models.

ModelSST-2AG-News
TextCNN (%)TextRNN (%)BERT (%)TextCNN (%)TextRNN (%)BERT (%)

Clean89.0788.1991.4993.8893.3094.25
SN-10-188.5887.7591.0593.6693.8793.87
SN-10-288.5887.7591.1093.5592.9993.87
SN-10-388.9187.7591.0593.4392.9793.79
SN-10-488.5887.7091.0593.3992.8094.01
SN-10-588.6387.7091.2793.6392.8493.78
SN-10-688.8587.7091.0593.4592.8393.84
SN-10-788.5888.2591.2193.3992.8694.11
SN-10-888.5887.7091.6593.3892.8494.01
SN-10-988.5887.7091.3293.3992.8094.17
SN-10-1088.6387.8191.2793.4792.8793.92