Deep Reinforcement Learning for Security-Aware Task Offloading in IoT Edge Computing: A Systematic Survey of Frameworks, Challenges, and Future Directions

Abstract

Background: IoT devices deployed in hospitals, factories, vehicles, and homes are largely unable to defend themselves. Limited compute and power budgets make running a local intrusion detector impractical, even as the traffic these devices generate grows increasingly attractive to attackers. DRL offers a practical alternative: rather than depending on a fixed environment model, a DRL agent learns by interacting with its surroundings and progressively improves its decisions. Methods: This survey reports a PRISMA based systematic review of 35 papers (2020–2026) drawn from IEEE Xplore, ACM Digital Library, and Google Scholar. We constructed a two-dimensional taxonomy one axis classifying papers by DRL algorithm family, the other reflecting how tightly each work integrates security into its reward design. Two authors screened all candidates independently; inter-rater agreement was measured using Cohen's kappa (κ = 0.93 for eligibility screening; κ = 0.91 for taxonomy assignment). Results: Across all 35 reviewed papers (including 7 non-DRL contextual comparators), 40% use performance-only reward design (Type A), 29% security-constrained (Type B), 31% security-weighted (Type C), and 0% implement asymmetric false-negative-penalized design (Type D). The corpus also shows a directional pattern toward lower packet-loss rates for continuous-action DRL agents (DDPG, TD3, SAC) under bandwidth-constrained conditions. Three things don't appear anywhere in the corpus: using on-device classifier confidence as a DRL state variable; testing on physical hardware; and building a reward that treats a missed attack as more costly than a false alarm. Conclusion: Eight concrete research directions follow, from testbed validation and federated DRL to 5G channel modeling and grounding reward weights in real security-cost data.