A Deep Neural Network Fingerprinting Detection Method Based on Active Learning of Generative Adversarial Networks

The issue of copyright protection for deep learning models is critical. We propose a novel deep neural network fingerprint detection method based on active learning of generative adversarial network (AL-GAN), which uses classifier to help generate potential fingerprint samples near the low-density boundary of normal samples to assist generative adversarial network training. The experiment results show that AL-GAN algorithm can generate informative latent fingerprint samples through GAN with active learning. It can clearly separate the fingerprint samples from training samples basing on the distance to conditional normal distribution. Meanwhile, it reduces the matching rate of model fingerprint and improves the success rate of the evasive model fingerprint inspections.