Resource Efficient Federated Learning for Deep Anomaly Detection in Industrial IoT applications

Anomaly data constitute a thorny problem in numerous industrial applications. In recent years, deep learning enabled anomaly detection has emerged as a critical direction, however the improved detection accuracy and reconstruction is achieved with the utilization of large neural networks with more layers and nodes, increasing their storage and computational cost. Moreover, the data collected in edge devices contain user privacy, introducing challenges that can be successfully addressed by the very recent on-device privacy-preserving distributed machine learning paradigm, known as federated learning (FL). This paradigm allows edge devices to locally train and exchange models increasing also the communication cost. Thus, in order to deal with the increased communication, processing and storage challenges introduced by FL based deep anomaly detection NN pruning is expected to have significant benefits towards both reducing the processing, storage and communication complexity but also towards avoiding the over fitting problem. With this focus, a novel compression-based optimization problem is proposed at the server-side of a FL paradigm that simultaneously fusses the local models broadcast by the edge devices and performs pruning generating a much more compressed model to be deployed at the edge devices with no accuracy loss even if the training is performed using a small subset of data instances, achieving compression rates greater than 99%.