Adaptive Firewall Policy Optimization Based on Multi-Objective Reinforcement Learning

Traditional static firewalls struggle to manage complex, dynamically evolving network traffic, often leading to rule redundancy, conflicts, and latency degradation. To address these challenges, this study proposes an intelligent firewall rule optimization framework based on Proximal Policy Optimization (PPO). The framework models the rule scheduling task as a multi-objective reinforcement learning problem, integrating system latency, throughput, and false positive rate into a composite reward function. A custom simulation environment compatible with OpenAI Gym is developed to represent firewall states and actions, while visualization tools such as reward curves, clip fraction, and action heatmaps enhance interpretability. Experimental evaluations demonstrate that the proposed PPO-based method outperforms baseline algorithms (Q-learning, DQN, A2C) in convergence speed, false positive reduction, and throughput improvement, achieving up to 60% faster stabilization and a 15% lower error rate. The approach offers a scalable and adaptive framework for real-time firewall policy management, contributing to the development of intelligent, self-optimizing network defense systems.