An anomaly detection approach for backdoored neural networks: face recognition as a case study