Latency-Aware Data Allocation Optimization for LEO Satellite IoT Networks with Federated Learning

Federated learning (FL) has been deployed on low earth orbit (LEO) satellites Internet of Things (IoT), where learning models can be trained collaboratively, thus preserving IoT data privacy without centralizing. However, the efficiency of FL is significantly hindered by the straggler that cause maximum latency. The data allocation strategy that parallelizes learning could potentially increase efficiency of FL for LEO satellite IoT networks since multiple LEO satellites can access a terrestrial IoT gateway concurrently. However, modeling and optimizing data allocation poses a significant challenge. To address this challenge, this paper proposes a collaborative learning method with latency-aware data allocation for LEO satellite IoT networks. Particularly, we formulate the data allocation strategy as an optimization problem of minimizing the maximum latency which is the sum of training time of the learning model and signal propagation delay, while considering the constraint of limited energy at each satellite. Next, we use a line search sequential quadratic programming (SQP) method to decompose the problem into a sequence of quadratic programming (QP) subproblems, which are further solved by the active-set algorithm. Simulation results show that nearly the half of the maximum latency per round can be decreased and the procedure of convergence is accelerated about 25 % in a large LEO satellite constellation with 1000 satellites and 10 IoT gateways.