Explainable machine learning frameworks for cryptography protocol design using discrete structures

This paper introduces Explainable Secure-Net, a new framework that leverages machine learning to aid in the automatic design of cryptographic key-exchange protocols whilst providing transparent human-readable explanations for each decision made. Classical techniques are based on expert-designed algebraic rules and usually lead to long and intricate manual proofs, whereas Explainable Secure-Net encodes some core discrete-mathematical ingredients e.g., finite-field operations and protocol graph structures into a graph neural network. For each step of the protocol that the model proposes, a light-weight explanation module serves to highlight what drove its choice with respect to distinguishing features, keeping the design process transparent and readily auditable. For security, all candidate protocols are run through a hybrid verification process which marries formal symbolic verification to large scale statistical testing. On a 1,500 benchmarks of synthetic protocol examples, Explainable Secure-Net achieves 94 % synthesis accuracy and explores more than 95 % of the potential key-exchange space, while keeping the false-negative vulnerability rate below 1 % and providing 128-bit security guarantees. Our findings show that it is feasible to speed protocol innovation without a loss in terms of rigor and interpretability. We argue that Explainable Secure-Net is an important first step towards machine-augmented cryptographic design tools that can automatically propose, explain, and certify secure protocol definitions for a variety of cryptographic tasks.