The problem of detecting anomalies in graph streams arises in many applications such as cybersecurity and finance. Several methods are proposed in the literature to deal with this problem. However, most of these methods are vulnerable to poisoning attacks, which consist in compromising the learning process by injecting corrupted data, during the initialization or training phases, to alter the model representing the normal behavior of the system. In this work, we extend one of the most recent and effective anomaly detection methods to deal with this attack. We proceed by hybridization by considering another method of anomaly detection as a filter which eliminates empoisoned data.