Special Session 183: Mathematics in Cryptography and Codes

Cryptanalytic Parameter Recovery of Neural Networks: Some Recent Advances
Yi Chen
Institute of Advanced Study, Tsinghua University
Peoples Rep of China
Co-Author(s):    
Abstract:
The problem of model extraction in machine learning has been studied for over thirty years. Its most challenging goal --- functionally equivalent extraction in the black-box setting --- is achieved via parameter recovery. Since Crypto 2020, researchers have made significant progress by approaching it through the lens of cryptanalysis. In this talk, we will briefly review this problem and introduce some recent results in this emerging direction, with a special focus on attacks in the hard-label setting.