GanFinger: GAN-Based Fingerprint Generation for Deep Neural Network Ownership Verification

As deep neural networks become fundamental in various fields, protecting these models as valuable  assets has become increasingly crucial To achieve this, various neural network fingerprint methods have been  proposed. However, the existing approaches often have deficiencies in terms of efficiency, stealthiness, and  discriminability. To address these issues, we present GanFinger, which constructs network fingerprints based  on network behavior, characterized by the outputs of pairs of original examples and conferrable adversarial  examples. Specifically, GanFinger leverages generative adversarial networks (GANs) to effectively generate  conferrable adversarial examples with imperceptible perturbations. These examples produce identical outputs  on copyrighted networks while yielding different results on irrelevant networks. Moreover, to enhance the  accuracy of verification, the network similarity is computed based on the accuracy-robustness distance of  fingerprint outputs. To evaluate the performance of GanFinger, we construct a comprehensive benchmark  consisting of 186 networks with five network structures and four popular network post-processing techniques.  Experiments show that GanFinger significantly outperforms the state-of-the-art in efficiency, stealthiness, and  discriminability. It achieves 6.57 times faster fingerprint generation and improves the area under the  uniqueness curve by 0.175, resulting in a relative improvement of approximately 26%.