Paper 2026/1678

Cryptanalytic Extraction of Multi-Head Softmax Attention Models

Abstract

Since the seminal work of Carlini et al. at CRYPTO 2020, cryptanalytic model extraction has shown neural-networks parameters can be recovered from black-box queries. Existing attacks are largely built around piecewise-linear phenomena. Softmax attention, as the key component of the transformer architecture, presents a different extraction landscape: its nonlinearity is smooth and sequence-dependent, which renders the existing piecewise-linear-based method inapplicable. Recent work has investigated the learnability of a single-head attention model, while in the multi-head case, the parameters of the multi-head attention layer cannot be uniquely identified from value queries alone. In this paper, we propose the first attack against multi-head attention models. We formalize the extractable representative of multi-head attention and give a polynomial-time algorithm for extracting the parameters of the canonical representative model. We also test our algorithm end to end under finite precision, and successfully extract the parameters of a softmax attention model with token dimension 8 and 6 heads to accuracy $2^{-51}$. Moreover, we overcome the limitation that existing parameter extraction algorithms for one-layer single-head Transformers fail when the ReLU feedforward networks (FFNs) include bias terms. The effectiveness of our approach is demonstrated through model extraction attacks in finite-precision experiments. These results show that softmax normalization itself exposes exploitable algebraic structure, extending cryptanalytic extraction beyond ReLU-centric techniques.