Paper ID: 2409.14122

Efficient and Effective Model Extraction

Hongyu Zhu, Wentao Hu, Sichu Liang, Fangqi Li, Wenwen Wang, Shilin Wang

Model extraction aims to create a functionally similar copy from a machine learning as a service (MLaaS) API with minimal overhead, often for illicit purposes or as a precursor to further attacks, posing a significant threat to the MLaaS ecosystem. However, recent studies show that model extraction is inefficient, especially when the target task distribution is unavailable. In such cases, even significantly increasing the attack budget fails to yield a sufficiently similar model, reducing the adversary's incentive. In this paper, we revisit the basic design choices throughout the extraction process and propose an efficient and effective algorithm, Efficient and Effective Model Extraction (E3), which optimizes both query preparation and the training routine. E3 achieves superior generalization over state-of-the-art methods while minimizing computational costs. For example, with only 0.005 times the query budget and less than 0.2 times the runtime, E3 outperforms classical generative model-based data-free model extraction with over 50% absolute accuracy improvement on CIFAR-10. Our findings highlight the ongoing risk of model extraction and propose E3 as a useful benchmark for future security evaluations.

Submitted: Sep 21, 2024