Target-Group-Aware Robust Federated Reinforcement Learning for Heterogeneous EV Charging Under Targeted Poisoning

July 31, 2026

The growing complexity of EV charging and V2G operations challenges distribution networks. Controllers must meet user and station needs while avoiding overload and voltage violations. In federated learning, poisoning attacks via manipulated updates threaten the global model. This thesis studies target‑group‑aware robust federated RL for heterogeneous EV charging. We build an EV2Gym framework with private, workplace, and public clients. Each client trains a local PI‑TD3 controller; the server aggregates updates without raw data. We consider targeted poisoning, where adversaries degrade a specific group. We propose TA‑GMA, which extends geometric‑median aggregation with target‑group inference, soft gating of suspects, and preservation of honest clients. Simulations under public‑target poisoning show TA‑GMA improves target group reward and reduces overload vs. the best geometric‑median baseline. Diagnostics confirm that our coefficient control lowers attacked clients' weight, while retaining useful contributions from non‑suspects. Thus, target‑group‑aware aggregation is promising for robustness and interpretability in federated RL for heterogeneous charging.

Authors in the community:

• 

  B

  Binglong Yang

  ist1115701

Supervisors of this institution:

• 

  A

  António José Castelo Branco Rodrigues

  ist12288

electrical-engineering-electronic-engineering-information-engineering - Electrical engineering, electronic engineering, information engineering

por - Portuguese

Instituto Superior Técnico