Paper 2026/1066

Breaking Slope and Structure Restrictions: Broadening Hard-Label Cryptanalytic Extraction of PReLU Neural Networks

Abstract

This paper studies the problem of model parameter extraction of PReLU neural networks in the hard-label setting, the most challenging setting. Existing attacks on PReLU neural networks suffer from two fundamental restrictions: (1) the learnable slopes in PReLU activations are restricted to smaller than 1, not conforming to the standard definitions of PReLU activations; (2) they do not apply to expansive PReLU neural networks. In this paper, for the first time, we break the two restrictions by proposing a new attack in the hard-label setting. Our breakthroughs stem from two new techniques and an important finding. First, we propose a new network isomorphism, called flip-and-scaling, which helps break the slope restriction and build a new extraction framework. Second, we find that there are linear constraints on the internal states of expansive PReLU neural networks, and give the exact number of linear constraints. Third, we propose a new neuron signature recovery method for expansive PReLU neural networks, which overcomes the challenge brought by linear constraints and breaks the structure restriction. The correctness and effectiveness of our work have been fully verified by experiments on several hundred expansive PReLU neural networks. Overall, our work not only overcomes the restrictions of existing attacks but also provides some inspiration for future work.