Hardware aware federated learning with sparse models for adaptive anomaly detection in IoT

Securing the enormous and varied network of linked devices has become a major obstacle in the fast changing environment of the Internet of Things (IoT). Traditional anomaly detection methods sometimes find it difficult to handle the range of IoT devices, which vary widely in processing power and energy restrictions. To solve this, we propose HANN-FL (Hardware-Aware Neural Network with Federated Learning), an adaptive platform for detecting abnormalities in IoT devices. HANN-FL provides a scalable and effective approach by combining Federated Learning (FL) with hardware-aware neural network models. One of the most important components of HANN-FL is the application of SPADIS (Sparse Pruned Adaptive Distilled Scaling), a post-training compression method that lowers neural network weights by dynamically scaling activations during inference to keep accuracy intact. With low memory and computational overhead, SPADIS makes it possible to distribute HANN-FL on resource-constrained edge devices. FL helps HANN-FL improve privacy by enabling security models to work together across edge devices without centralizing sensitive data. The proposed framework (HANN-FL) helps in advancing secure and effective AI driven IoT systems by reducing model size, offering strong security and privacy along with adaptability to diverse systems.