Purpose: This paper focuses on examining the robustness of a cooperative Multi-Agent Reinforcement Learning (MARL)-based Intrusion Detection System (IDS) for intrusion detection in decentralised 5G security settings. Even though MARL techniques have proven effective against dynamic threats in decentralized 5G networks, current research has not considered any adversarial scenarios at all. Methods: A cooperative MARL-based Intrusion Detection System was developed through the CRISP-DM approach. Radio Access Network (RAN), MEC, and Core agents were trained using Centralised Training with Decentralised Execution (CTDE) and Deep Q-Network (DQN) methods. The algorithm was tested on the NSL-KDD and UNSW-NB15 datasets against Fast Gradient Sign Method (FGSM) evasion attacks (ε = 0.05-0.30) and Byzantine poisoning attacks with 5%, 10%, and 20% compromised agents. Result: The model achieved 96.94% accuracy on NSL-KDD and 85.15% on UNSW-NB15 in clean scenarios. The FGSM attack at ε = 0.20 resulted in substantial performance deterioration, leading to accuracy drops of 50.14 and 45.26 percentage points, respectively, and a simultaneous increase in false positives. Byzantine poisoning produced smaller but persistent decreases in accuracy of 12.03 and 2.62 percentage points, respectively. Novelty: This study provides among the first empirical evaluations of adversarial fragility in cooperative MARL-based intrusion detection within distributed 5G-oriented security abstractions, demonstrating that cooperative intelligence alone does not guarantee adversarial robustness.
Empirical evaluations reveal a fundamental brittleness in existing defenses: with a single trainable 7B planner, Trident reduces blue agent defensive performance by an average of 522% compared to static red agent baselines while autonomously discovering emergent behaviors such as decoy avoidance and adaptive state prioritization that static heuristics entirely fail to uncover.
The DT environment allows for real-time synchronization between the physical and virtual world, which makes it extremely vulnerable to advanced cyber-attacks. Most of the security techniques used so far emphasize the ability to detect attacks but are weak when it comes to responding to them in an automated and dynamic manner. To tackle this challenge, in this research, we propose an MARL-based automated response solution for Digital Twin security applications. The MARL model uses a decentralized agent architecture where the agents learn how to respond optimally under different circumstances. We formulate the problem as a multi-agent markov decision process, and use the Q-learning approach combined with the idea of experience replay. Performance of the developed solution will be measured with various criteria, such as accuracy, threat mitigation rate, attack success rate, response time, system downtime, cumulative rewards, and system resilience. The findings from the experiments clearly show that the suggested model is able to obtain an accuracy of 94.4%, threat mitigation capability of 93.1%, and lower response times than previous models. Moreover, the learning curve clearly illustrates that stable convergence and better optimization of policies are obtained over episodes. The suggested MARL-based framework is efficient in providing automated incident response for Digital Twin ecosystems.
To address the performance degradation in vehicle formation control caused by communication disruptions under denial-of-service (DoS) attacks, this article proposes a secure control method that integrates a nonzero-sum game and inverse reinforcement learning (IRL). At the game-theoretic layer, a dynamic game model accounting for the attacker’s energy cost is constructed to capture the adversarial interaction and strategy spaces between DoS attacks and the formation controller, with an approximate Nash equilibrium solution derived for both attack and defense strategies. At the reward learning layer, an IRL approach is introduced to adaptively learn the weight parameters of multiobjective performance indices from offline demonstration data, thereby mitigating the sensitivity of system performance to manual weight tuning. At the control implementation layer, based on the learned reward weights, a neural network (NN) is employed to approximate the solution of the Hamilton–Jacobi (HJ) equation, thereby constructing a computable feedback control law. Simulation results demonstrate that the proposed method can effectively suppress formation tracking errors in DoS attack scenarios and exhibits superior robustness, adaptability, and energy efficiency compared to traditional methods.
Securing the decision-making process of reinforcement learning (RL) agents during deployment is crucial in privacy-sensitive and security-critical domains. However, as deployed policies interact with the environment, they generate observable trajectories that can inadvertently leak sensitive decision patterns. External observers can readily exploit these trajectories via trajectory-based attacks such as imitation learning (IL), inverse reinforcement learning (IRL) to extract the policy or infer the underlying reward structure. Current deployment-time defenses typically tackle policy extraction or reward inference in isolation, and often fail to provide strict guarantees on the agent’s task utility. To bridge this gap, we propose the Dual-Critic Constrained Deceptive Q-Learning (DCD-Q) method, a deployment-time trajectory protection framework that aims to reduce the information leaked by released trajectories while preserving acceptable task performance. DCD-Q employs a utility critic to constrain decisions to a near-optimal candidate action set, aiming to preserve task utility while enabling controlled deceptive behavior. Simultaneously, it constructs a dynamic anti-reward from sliding-window visitation statistics to bias execution toward less-visited feasible actions. This induces a controlled, time-varying deceptive behavior in the released trajectories. We theoretically analyze the utility-preserving component by deriving a lower bound on the expected return under constrained action selection. Experiments on benchmark environments evaluate DCD-Q against representative trajectory-based attackers, which show that DCD-Q reduces the effectiveness of policy and reward recovery while maintaining a task-dependent utility–protection trade-off.