Robust Secure Aggregation For Co-located IoT Devices With Corruption Localization

Many Internet-of-Things (IoT) applications benefit from secure Federated (Machine) Learning (FL) techniques, e.g., industrial automation and smart cities. Since IoT devices have limited resources, such techniques need to be very efficient. In this paper, we propose a formal model as well as a new means for provably secure and efficient (i.e., IoT-friendly) aggregation for FL in IoT settings. Besides input privacy, it offers result correctness and malicious input localization, both robust against up to a threshold of malicious devices. The proposed techniques involve no interaction among devices, which only send a short message, and perform lightweight encryption and integrity detection.