Jamming-Resilient Over-the-Air Computation for Federated Learning in Edge Intelligence

By making full use of the accumulative feature of wireless signals, the over-the-air computation makes it possible for the federated learning to aggregate all the parameters simultaneously from the distributed learning agents, which lessens the delay of updating the parameters caused by the collisions of wireless signals. However, the accumulative feature in the over-the-air computation also opens the door to the jamming attackers. Specifically, an adversary can easily tamper the aggregated result in the over-the-air computation by launching a simple jamming signal. Thus, it is significant to improve the jamming resilience of the over-the-air computation so that it can be adopted in a wider range of wireless scenarios. In this paper, we propose the first jamming-resilient over-the-air computation algorithm for the federated learning via the multiple channel technique. Specifically, in our jamming-resilient algorithm, the learning agents randomly and uniformly choose one of the multi-channels to update their parameters, and the server drop a fraction of results according to a greedy screening rule. A theoretical analysis with some rigorous proofs is presented to show that the expected aggregated result by our algorithm is asymptotically close to the authentic value when the data on the learning agents has a uniform distribution. An extensive simulation is also conducted to show the correctness of our algorithm.