DisGUIDE: Disagreement-Guided Data-Free Model Extraction

Recent model-extraction attacks on Machine Learning as a Service (MLaaS) systems have moved towards data-free approaches, showing the feasibility of stealing models trained with difficult-to-access data. However, these are ineffective or limited due to low accuracy extracted and high number queries under attack. The query cost makes such techniques infeasible for online MLaaS that charge per query. We create novel approach get higher efficiency than prior model extraction techniques. Specifically, we introduce generator training scheme maximizes disagreement loss between two clone attempt copy This loss, combined diversity experience replay, enables produce better instances train models. Our evaluation popular datasets CIFAR-10 CIFAR-100 shows our improves final by up 3.42% 18.48% respectively. average required achieve state art is reduced 64.95%. hope this will promote future work feasible defenses against attacks.