Machine Learning-Based Evaluation of Binary Sequence Randomness for Secure Communications

Machine learning is an emerging domain with applications across a wide range of fields. This paper investigates a machine learning-based approach for assigning scores to digital sequences based on their level of randomness, using statistical indicators commonly employed in classical randomness evaluation tests. A total of 15,000 binary sequences were generated for training and testing, covering six classes: Advanced Encryption Standard in Counter Mode (AES-CTR), ChaCha20, Mersenne Twister, Linear Congruential Generator (LCG), biased random sequences, and periodic sequences. For each sequence, several statistical features were computed, including Shannon entropy, Exclusive OR (XOR) relation score, and binary matrix rank. These features were used to train a machine learning model based on the XGBoost algorithm, implemented in Python. The proposed model successfully distinguishes statistically degraded sequences (periodic and biased) from statistically uniform ones. However, high-quality pseudorandom generators exhibit overlapping behavior, indicating similar statistical properties, especially for short sequences. The level of randomness is an important factor in determining whether a cryptographic key is suitable for use in secure encryption algorithms.