PPML Is More Vulnerable to Cryptanalytic Extraction Attacks

With the expansion of Machine Learning as a Service (MLaaS), Secure Multi-Party Computation (MPC) is widely used to protect the privacy of both proprietary models and client data during inference. To achieve practical performance, these protocols typically rely on fixed-point arithmetic over finite rings. However, this design choice introduces a unique arithmetic vulnerability: silent modular wraparound. In this paper, we propose a novel model extraction attack that actively exploits this behavi