Federated Learning and Custom Aggregation Schemes

Federated Learning (FL) enables model training on decentralized data, where custom aggregation schemes are crucial for accuracy, robustness, and security. The Scaleout Edge AI platform facilitates this through a feature called "Server Functions," which allows practitioners to design and test their own aggregation methods. A practical example demonstrates how to build a custom aggregator to defend against a label-flipping attack, a scenario where a malicious client corrupts its local training data. The proposed mitigation strategy uses cosine similarity to compare client model updates, identify outliers, and exclude them from the global model aggregation process.