SARG-MORL: Security-Aware Risk-Guided Multi-Objective Reinforcement Learning for Adaptive Routing in Software-Defined Networks

Abstract

Software-Defined Networking (SDN) offers a programmable control plane and a network-wide view that can be exploited for adaptive traffic engineering. However, routing decisions based only on latency, throughput, congestion, or link utilization can select paths that are operationally efficient but exposed to anomalous traffic or attack conditions. Conversely, security mechanisms that operate independently of routing often detect a threat without allowing the routing policy to react to its spatial and temporal consequences. This paper proposes SARG-MORL, a security-aware risk-guided multi-objective reinforcement learning framework for adaptive SDN routing. The central idea is to transform security observations into a continuously updated path-risk field and expose that field to the routing agent as both a state feature and a safety constraint. A topology-aware graph encoder represents nodes, links, flows, and security indicators; a constrained multi-objective actor-critic agent then selects routes by jointly considering delay, packet loss, residual bandwidth, load variance, recovery cost, and security exposure. Unlike a conventional weighted reward formulation, the proposed design separates performance preferences from non-negotiable security constraints, preventing the agent from compensating for an unsafe path merely through gains in latency or throughput. The paper defines the architecture, mathematical formulation, learning procedure, threat model, Mininet/Ryu experimental design, baselines, ablation strategy, and reproducibility requirements. The framework is intended for evaluation under normal, bursty, congestion, link-failure, and attack-injected conditions.