Risk-Aware Scheduling in Distributed Systems Using Reinforcement Learning and Mined Traffic Patterns

Large-scale networked systems now face security threats that shift faster than fixed schedulers can follow. Conventional scheduling optimizes throughput and latency, and treats security as an afterthought. This paper takes a different approach: a reinforcement learning framework for risk-aware scheduling, with its risk signal drawn from multi-modal analysis of UNSW-NB15 traffic – statistical profiling, DBSCAN and Gaussian mixture clustering, and sequential pattern mining. The pipeline runs in two phases. First, preprocessing and feature engineering turn raw network data into usable scheduling features. Second, DQN and PPO agents learn scheduling policies under a dynamic, risk-weighted reward. The balance between security and performance shifts with the current mined risk level, not a fixed rule. The risk signal compares live traffic against mined high- risk patterns, restricted to the feature dimensions those patterns actually weighted by each correlation dimension with the real attack label. Across five seeds on real UNSW-NB15 data, both schedulers significantly cut security violations against Random, Round Robin, and Greedy baselines (p < 0.001) and earn significantly higher reward. PPO reaches the lowest violation, cutting violations by roughly three-quarters relative to the baselines. DQN keeps utilization closest to the baselines while still improving reward. The tradeoff is response time weighing risk before every placement costs speed.