Federated deep reinforcement learning for privacy-preserving offloading in vehicular edge computing

Area of Science:

• Vehicular communication networks
• Edge computing
• Machine learning for intelligent transportation systems

Background:

• Internet of Vehicles (IoV) applications demand high computation and low latency.
• Limited vehicular computing power necessitates edge offloading.
• Centralized learning for offloading increases communication overhead and data exposure.

Purpose of the Study:

• To propose a privacy-aware federated deep reinforcement learning (FDRL) framework for vehicular edge computing task offloading.
• To address challenges of limited vehicle computing power, latency requirements, and data privacy in dynamic mobility environments.
• To improve system cost, delay, energy efficiency, and deadline reliability through intelligent offloading decisions.

Main Methods:

• Developed a novel FDRL framework integrating federated learning and deep reinforcement learning (DRL).
• Incorporated four key mechanisms: hybrid action representation, personalized federated aggregation, task-criticality-aware deadline reliability, and a handover-aware multi-RSU model.
• Trained local Soft Actor-Critic (SAC) policies using model parameters shared with the federated coordinator, not raw data.

Main Results:

• The proposed FDRL framework demonstrated competitive or superior performance in system cost, delay, energy consumption, and deadline violation compared to baseline methods.
• Achieved reduced communication overhead through federated learning, sharing model parameters instead of raw vehicular data.
• The framework effectively handles non-IID data and dynamic handover scenarios in multi-RSU environments.

Conclusions:

• The FDRL framework offers an effective solution for privacy-aware task offloading in vehicular edge computing.
• It balances computational demands, latency constraints, and data privacy concerns in IoV environments.
• Future work could explore formal privacy guarantees against inference attacks on model updates.