IEEE Vehicular Technology Conference 2022 Spring Keynote speech

Centralized data collection and training in conventional machine learning (ML) algorithms have raised many concerns including privacy restrictions and communication cost due to massive amount of data transfer. Federated leaning (FL) exploits the rapidly growing computational capacity in small local devices and allows these devices to train ML models locally and only exchange the trained model parameters with the edge server. Through this, FL can greatly alleviate data privacy concern, reduce communication cost, and help build a scalable centralized ML model. FL methods offer a number of prominent advantages, including scalability and data privacy. On the other hand, a large-scale wireless network normally involves many heterogeneous devices with varying constraints and encounters very dynamic channel environments. This raises many challenges such as system heterogeneity, statistical heterogeneity, privacy and security, user scheduling, fairness in FL. This talk will present some of our recent research outcomes on model parameter transmission schemes and user scheduling strategies in FL that tackle these challenges. Techniques such as NOMA and over-the-air computation are introduced to achieve fast ML training. Model parameter compression and sparsification are further introduced to reduce the wireless communication cost. Moreover, model update-based aggregation is applied to defend against Byzantine attacks and individual client model initialization schemes are exploited to enhance privacy protection in FL.