Deep Neural Networks (DNNs) have attracted significant attention, and their internal models are now considered valuable intellectual assets. Extracting such a model via oracle access to a DNN is conceptually similar to extracting a secret key from a block cipher. Consequently, cryptanalytic techniques, particularly differential-like attacks, have been actively explored. ReLU-based DNNs are the most common and widely deployed architectures. While early works (e.g., Crypto 2020, Eurocrypt 2024) assume access to exact output logits, which are typically not exposed, more recent works (e.g., Asiacrypt 2024, Eurocrypt 2025) focus on the hard-label setting, where only the final classification result (e.g., "dog" or "car") is available. Notably, Carlini et al. (Eurocrypt 2025) showed that model extraction is feasible in polynomial time even under this restricted setting. In this paper, we show that a key assumption underlying their attack becomes increasingly unrealistic as the target depth grows. While prior works noted neurons whose activation states rarely change, we analyze their concrete impact on hard-label extraction: even a single neuron that is (almost) always active can prevent the attack from proceeding unless its parameters are recovered, and ignoring it incurs a non-negligible error. A straightforward solution is to extract these parameters by observing a state switch of such a neuron, but observing such a switch becomes exponentially harder as depth increases, implying that hard-label extraction is not always polynomial time. To address this limitation, we propose a novel attack called cross-layer extraction. Rather than extracting secret parameters (e.g., weights and biases) directly, we exploit cross-layer interactions to recover them from deeper layers, reducing query complexity and addressing limitations of existing approaches.

翻译：暂无翻译