Gradient Leakage Attacks in Federated Learning

Question

May I ask if I use top-K to compress the gradient, can the attacker recover the original information of the data from the compressed gradient?