We propose DictPFL, a framework that ensures efficient and private federated learning (FL) by encrypting shared gradients and keeping most gradients local, while still preserving the performance of global gradient aggregation.