JOINT POWER COST AND LATENCY MINIMIZATION FOR SECURE COLLABORATIVE LEARNING SYSTEM

This work investigates the model update security in a collaborative learning or federated learning network by using the covert communication. The CC uses the jamming signal and multiple friendly jammers (FJs) are deployed that can offer jamming services to the model owner, i.e., a base station (BS). To enable the BS to select the best FJ, i.e., the lowest cost FJ, a truthful auction is adopted. Then, a problem is formulated to optimize the jamming power, transmission power, and local accuracy. The objective is to minimize the training latency, subject to the security performance requirement and budget of the BS. To solve the non-convex problem, we adopt a Successive Convex Approximation algorithm. The simulation results reveals some interesting things. For example, the trustful auction reduces the jamming cost of the BS as the number of FJs increases.