AI in Chaos: Adaptive and Secure Communication Via Deep Reinforcement Learning and Moving Target Defense
Zhangying He, Hossein Sayadi · IEEE Access · 2025
This study investigates how Machine Learning (ML) techniques can enhance the security and adaptability of encrypted communication systems, characterized by their suitability for resource-limited environments such as IoT and edge computing platforms. Leveraging chaotic encryption, we present an intelligent and secure communication framework in which messages are encrypted during transmission and decrypted upon receipt, preserving both confidentiality and resilience against attacks. Central to this work is the introduction ofREACT(Randomized Encryption with AI-Controlled Targeting), a novel, adaptive encryption-decryption framework leveraging Deep Reinforcement Learning (DRL) and Moving Target Defense (MTD) to dynamically randomize encryption patterns and strengthen security.REACTemploys a random controller to assign varying chaotic encryption modes, effectively thwarting attackers by making the encryption pattern unpredictable and difficult to intercept. At the receiver’s end, a parallel bank of DRL agents collaborates to accurately identify the active encryption mode and route the incoming signal to the corresponding decryption model, enabling accurate and efficient message recovery. We comprehensively evaluate the decryption performance of various ML models, analyzing their ability to reconstruct encrypted signals and maintain a strong correlation with the original data. Experimental results demonstrate high decryption accuracy, with F1-scores reaching 100% for the easiest mode, 95% for normal and hard modes, and 90% for the most challenging encryption mode. Moreover,REACTreduces the probability of successful attacks by up to 51%, effectively limiting adversaries to guess-level accuracy. These findings underscoreREACT’s potential as a robust and adaptive defense framework for secure and efficient communication across diverse application domains.