Gradient Disaggregation: Breaking Privacy in Federated Learning by Reconstructing the User Participant Matrix