Neural-Guided Enumerative SAT Framework for Cryptographic Key Recovery

Xinhao Zheng, Xinhao Song, Jingzhi Wang, Gongshen Liu, Junchi Yan

Keywords: Enumerative, SAT, Cryptanalysis

Abstract: Boolean satisfiability (SAT) provides a principled framework for cryptanalysis by encoding cryptographic problems into conjunctive or algebraic normal forms. While the well-known (conflict-driven clause learning) CDCL-based solvers excel at handling SAT instances, their performance on cryptographic instances deteriorates quickly with increasing cipher complexity due to exponential search spaces. Recent neural approaches, ranging from end-to-end prediction to solver-integrated heuristics, show some promise yet suffer from scalability issues, limited recovery accuracy, or high computational overhead. In this work, we propose two complementary neural-guided enumerative SAT frameworks tailored for cryptographic key recovery. The first integrates lightweight neural predictions with traditional solvers: it predicts a subset of $k$ key variables as critical variables and enumerates the candidate assignments of these predicted variables, yielding up to a 5× speedup on the public benchmark SAT4CryptoBench. The second employs a discriminator trained to distinguish correct versus incorrect assignments in ANF encodings, enabling accurate pruning and propagation. On Simon datasets, this method achieves 79.5\% full-key recovery accuracy, significantly surpassing prior neural approaches. Together, these frameworks bridge the gap between machine learning and classical SAT solving, offering scalable and efficient cryptanalysis.

Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)

Submission Number: 7067