Exploring Potential Vulnerabilities in DRL-Based Congestion Control with Adversarial Policy

While deep reinforcement learning (DRL)-based congestion control (CC) algorithms outperform traditional TCP methods in dynamic networks, their robustness is challenged by the complex influence caused by state perturbations from competing flows. Existing DRL-based CC models, constrained by their black-box nature, often suffer performance degradation under dynamic network conditions, which may indirectly alter their observations. To address this challenge, we propose Mona, an adversarial attack framework that explores vulnerabilities in existing DRL-based CC schemes while helping improve their robustness. Leveraging a dual-Critic architecture and a global state representation, Mona enables centralized training with an explicit estimation mechanism embedded in the reward function, facilitating targeted attacks without direct access to the victim’s state. Based on this design, Mona can be distributedly deployed to indirectly perturb the victim flow’s observations by injecting a stealthy flow into the bottleneck link, ultimately inducing suboptimal decisions. Experiments demonstrate Mona’s effectiveness across diverse CC models, reducing victim flow throughput by 8.69%–26.67% under various network conditions in both simulated environments and real-world network deployments, and we conduct a preliminary exploration of defense strategies through a minimax-based adversarial training framework that improves model robustness by 20%.