This paper introduces a unified, data-driven framework for the end-to-end synthesis and security verification of lattice-based cryptosystems. Beginning with a curated benchmark of real-world implementations (e.g., Kyber, NewHope, FrodoKEM), we extract normalized performance and security features to train a machine-learning model that predicts key-generation and encapsulation latencies with high fidelity (MAE < 5 μs, R² ≈ 0.91). We then enumerate parameter candidates within NIST’s 128-bit security envelope, rank them by a composite score of latency, key size, and failure rate, and select the top proposals. Each candidate undergoes rigorous statistical testing Kolmogorov Smirnov distribution checks and Test Vector Leakage Assessment and interactive-theorem-prover proofs for correctness and IND-CPA security, completing in under two minutes per template. Finally, the framework auto-generates PQClean-compliant C stubs with embedded provenance and CI scripts. Experimental results demonstrate that our pipeline yields deployment-ready schemes matching or exceeding manual baselines in performance while carrying machine-verified security guarantees.
Authors: Shalini Pathak, Basu kalyanwat, Ashish Kumar, Pankaj Dadheech
DOI: https://doi.org/10.47974/jdmsc-2426
Publish Year: 2025
