Model Stealing Attack against Multi-Exit Networks

Compared to traditional neural networks with a single exit, a multi-exit network has multiple exits that allow for early output from intermediate layers of the model, thus bringing significant improvement in computational efficiency while maintaining similar recognition accuracy. When attempting to steal such valuable models using traditional model stealing attacks, we found that conventional methods can only steal the model's classification function while failing to capture its output strategy. This results in a significant decrease in computational efficiency for the stolen substitute model, thereby losing the advantages of multi-exit networks.In this paper, we propose the first model stealing attack to extract both the model function and output strategy. We employ bayesian changepoint detection to analyze the target model's output strategy and use performance loss and strategy loss to guide the training of the substitute model. Furthermore, we designed a novel output strategy search algorithm that can find the optimal output strategy to maximize the consistency between the victim model and the substitute model's outputs. Through experiments on multiple mainstream multi-exit networks and benchmark datasets, we thoroughly demonstrates the effectiveness of our method.