Fake or Compromised? Making Sense of Malicious Clients in Federated Learning