MsaaDI: A Heterogeneity-Resilient Federated Learning Framework for IoT Device Identification With Multi-Scale Adaptive Aggregation

This research paper presents MsaaDI, a federated learning (FL, a technique where AI models are trained across many devices without sending raw data to a central server) framework designed to identify IoT devices (internet-connected hardware like cameras and sensors) more accurately. The framework addresses two main problems that make federated learning difficult: Non-IID distributions (when different devices have data in different formats or proportions) and class imbalance (when some types of devices are underrepresented in training data), using a Multi-Scale Adaptive Aggregation mechanism on the server side and improved local training strategies on client devices to achieve up to 94% accuracy in testing.