Stability-Aware Aggregation for Binary Federated Learning in the Presence of Noisy Client Data
4th Cognitive Models and Artificial Intelligence Conference, AICCONF 2026, Prague, Czech Republic, 24 - 25 April 2026, (Full Text)
- Publication Type: Conference Paper / Full Text
- Doi Number: 10.1109/aicconf69182.2026.11600726
- City: Prague
- Country: Czech Republic
- Keywords: Binary Federated Learning, Edge AI Security, Loss Stability, Noisy Client Data, Robust Aggregation
- İstanbul Ticaret University Affiliated: Yes
Abstract
Binary Neural Networks (BNNs) enable efficient edge deep learning, yet their sensitivity makes them vulnerable to noisy client data in Federated Learning (FL). In scenarios where clients have corrupted labels or noisy features, standard aggregation methods such as FedAvg struggle to mitigate the negative effects of erroneous updates, leading to severe global model instability. This paper proposes a robust aggregation method specifically designed to handle noisy client data by leveraging the local loss stability of BNNs. In our approach, the global model monitors the stability of each client's training process to assess their reliability and adjust their participation accordingly. Experimental results demonstrate that BNNs are highly sensitive to label noise, and the proposed model mitigates the negative effects of noisy clients while maintaining reasonable accuracy under aggressive noise conditions.