Quantum-Resistant Machine Learning-Based Cryptographic Key Generation System

Abstract With the advent of large-scale quantum computing, there is the existential threat to the broadly used public-key cryptography, such as RSA and Elliptic-Curve Cryptography (ECC). They are based on computational problems integer factorization and the discrete logarithm which Shor algorithm [5] can solve in a polynomials time on a quantum processor powerful enough. As a reaction, the National Institute of Standards and Technology (NIST) formalized the Module Lattice-Based Key Encapsulation Mechanism (ML-KEM/Kyber) as FIPS 203 in 2024 [10]. Despite the mathematical nature of the lattice-based schemes, a key weakness of real-world implementation is the quality of entropy provided when generating keys. The entropy is so weak that it can be used to defeat even the most robust lattice construction, allowing key recovery attacks that do not rely on mathematical hardness conjectures. This paper describes a Quantum-Resistant Machine Learning-based Cryptographic Key Generation System (QR-ML-KGS) that is an addition of an intelligent entropy quality measurement layer on the PQC key generation pipeline. The system uses five independent hardware sources to obtain entropy and extract a 27 dimensional feature vector. A final, calibrated Voting Ensemble classifier that used Random Forest, Gradient Boosting Machine, and Multilayer Perceptron gave an ROC-AUC of 0.9981 and F1-score of 0.9975. Entering the ML quality gate, entropy samples seed a Ring-LWE key generation algorithm, which implements Kyber-512, Kyber-768, and Kyber-1024, congruent with FIPS 203. The full pipeline had a latency average of 473.9 ms per key pair on commodity machines, with about 2.1 keys/sec. Keywords — Post-Quantum Cryptography, Kyber (ML-KEM), FIPS 203, Lattice-Based Cryptography, Ring-LWE, Machine Learning, Entropy Assessment, Random Forest, Gradient Boosting, Hybrid Key Derivation, NIST PQC.