Privacy-Preserving Federated Learning for Iot Devices

Abstract

Home appliance manufacturers seek customer input to enhance their goods and services for the development of a smart home system. We create a federated learning (FL) system including a reputation mechanism to aid home appliance makers in training a machine learning model using consumer data.     
Subsequently, producers may anticipate consumers' needs and consumption patterns in the future. The operational process of the system has two phases: in the first phase, customers train the preliminary model supplied by the manufacturer using both the mobile device and the mobile edge computing (MEC) server.
Customers gather data from several household appliances via smart phones, then downloading and training the basic model with their local data. Upon generating local models, clients endorse their models and transmit them to the blockchain. To mitigate any malice from consumers or manufacturers, we use blockchain technology to substitute the centralized aggregator in the conventional federated learning system. Due to the immutable nature of blockchain data, the behaviors of hostile consumers or producers are traceable. In the second step, manufacturers designate consumers or organizations as miners to compute the averaged model using the received models from customers. Upon completion of the crowdsourcing work, one miner, designated as the interim leader, uploads the model to the blockchain. To safeguard client privacy and enhance test accuracy, we implement differential privacy on the collected features and provide a novel normalizing approach.
We experimentally show that our normalizing approach surpasses batch normalization when features are subjected to differential privacy protection. Furthermore, to entice additional consumers to engage in the crowdsourced FL work, we have devised an incentive system to reward participants.