Robust federated learning for detecting false data injection attacks in advanced metering infrastructure under non-IID data

Advanced metering infrastructure exposes fine-grained residential consumption to false data injection attacks that tamper with meter readings while evading conventional bad-data screening. Federated learning keeps raw measurements on the device, yet real smart-meter populations are naturally non-independent and identically distributed and the distributed training loop is open to malicious clients that poison model updates. This study proposes a dual-robust federated detection framework that couples a lightweight two-branch gated recurrent detector with a server-side aggregation pipeline built from exponential moving average smoothing, client-level trimmed filtering, and deviation-aware cosine weighting. Real non-IID scenarios are constructed from the public Low Carbon London dataset by ACORN group, and generalization is checked on the Spanish GoiEner dataset across a different region and period. Under a strict fixed false-positive operating point the framework holds balanced accuracy near 61% from zero to thirty% malicious clients, whereas standard federated baselines collapse toward chance, confirming the value of client-level robust aggregation for edge-deployable attack detection.