Security-Aware Adaptive Computation Offloading in Mobile Edge Computing Using Reinforcement Learning and Deep Q-Networks

Mobile Edge Computing (MEC) enables resource-constrained mobile devices to offload computation-intensive tasks to nearby edge servers. Existing computation offloading approaches primarily optimise latency, energy consumption, or resource allocation, but often do not consider security constraints and multi-user queue stability within a unified decision framework. This paper studies, in simulation, a security-aware adaptive offloading framework that combines an analytical bandwidth threshold, workload classification, reinforcement learning, Lyapunov drift-plus-penalty queue control, Object Dependency Graph (ODG)-based vulnerability scoring, and a Deep Q-Network (DQN) over a continuous state. The framework derives a break-even bandwidth of 13.71 Mbps for time-beneficial offloading. A drift-plus-penalty admission controller bounds the edge queue; in a deterministic illustration it holds the queue at 21 tasks against 297 for greedy admission. ODG gating keeps labelled sensitive objects on the device, although a ratio-based score is shown to let transitively dependent objects leak. Over 10 random seeds, a DQN behind hard gates reaches a reward of -1.91 per step, against -2.07 for local execution and -1.97 for a rule-based pipeline, while a Q-table at the same discount factor falls to -2.42. The study is simulation-only, uses N-Queens as a workload proxy, assumes a single edge server and a reward derived from its own cost model, and these limitations are discussed explicitly. Contents of this record: the preprint draft (PDF, not peer reviewed) and an archive of the simulation code, results, figures and cited open-access papers (MIT-licensed code). Source repository: https://github.com/brindeshwar/security-aware-mec-offloading