Intelligent Security Scheme for Backdoor Attacks in High Speed Heterogeneous Communication Network

In the realm of cyber security within multi-user intelligent dynamic spectrum access (DSA) environments, Deep Reinforcement Learning (DRL) has emerged as a potent decision-making tool. Despite its capabilities, the inherent lack of interpretability in DRL's deep neural networks creates vulnerabilities, notably to backdoor attacks. This research delves into DSA in heterogeneous networks and proposes a subtle, yet effective method for executing backdoor attacks. The approach involves an attacker discreetly monitoring channel usage to select unobtrusive backdoor triggers, which are then slyly embedded into the training data of secondary users during the DRL model's training phase. In the model's operational or inference phase, the attacker can then activate these embedded backdoors. This activation manipulates the secondary users' behaviors, compelling them to execute specific actions that hinder their ability to successfully access communication channels. Simulations conducted as part of this research demonstrate the high efficacy of this backdoor attack method, achieving over a 90% success rate in varying DSA scenarios. This method is not only effective but also efficient, reducing the overhead associated with continuous attacks by 20% to 30%. Remarkably, the approach proves versatile, being applicable across three different DRL model types. These findings underscore a crucial aspect of DRL application in cyber-secure environments, especially in the context of dynamic spectrum access, where the balance of advanced functionality and cyber cyber security is essential.