LightSecAgg: Rethinking Secure Aggregation in Federated Learning