Federated Learning Method for Local Differential Privacy in IoT Networks
Pooja Kapila, Baldev Singh · 2022
The Internet of Things is a young technology with great potential. The fact that the creators of crowdsourcing programmes may easily determine users’ location data, traffic data, motor vehicle details, environmental information, etc., raises major sensitivity personal data privacy concerns for users. The cost of connectivity here between vehicles and the cloud server also considerably rises as the number of cars grows. In this research, we propose federated learning with local differential privacy integration to support machine learning model development in crowdsourcing applications while minimising privacy risks and reducing communication costs. We explicitly recommend using four LDP approaches to disrupt slopes caused by moving vehicles. The proposed Three-Outputs method provides three different output options to achieve a high accuracy when the private budget is constrained. The output options of 3 can be coded with just two bits to reduce transmission expenses. Additionally, when the privacy cost is significant, a perfect piecewise approach is recommended to maximise performance. In addition, we provide a suboptimal method with a simple formula and comparable utility to PM-OPT. Then, we develop a completely new hybrid mechanism by fusing Three-Outputs with PM-SUB. In order to synchronise the cars and cloud server for collaborative model training, an LDP-FedSGD technique is suggested. Extensive experiments on real data sets support the possibility that our suggested methods could safeguard privacy while maintaining value.