On Model Protection in Federated Learning against Eavesdropping Attacks