Model extraction attacks (MEAs) pose a significant threat to deep learning (DL) models, where adversaries aim to steal the decision behavior of targeted DL models. While several works have shown the ability of a surrogate model to mimic the target DL model, the underlying factors that make a DL model vulnerable to MEAs are unclear. Analyzing these underlying factors is the key to enhancing the security of DL systems. This involves exploring MEAs in diverse scenarios to understand the relationship between their success and the features of DL systems. In this paper, we evaluate the underlying factors influencing MEAs from two crucial perspectives: the model’s intrinsic parameters and the quality of the data used. For the model’s intrinsic parameters, we focus on how the batch size, learning rate, and optimizer influence the effectiveness of MEAs. Regarding data quality, we conduct an in-depth analysis of how data annotation and selection affect MEAs’ success. Our study includes analyzing variations in batch size, five learning rates, eight optimizers, the impact of varying proportions of dirty data, and the effects of subtle changes in data richness. The results of our research reveal a diverse range of susceptibilities to MEAs.

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%.