Graph based anomaly detection is an important task in many real-world domains such as health care, insurance, finance, and cyber-security. Even if existing semi-supervised models have proven to be efficient in identifying anomalies, they assume however that a labeled sample of the network is available but without taking into account the real-world problem of the unreliability of such a sample. In this paper we consider attributed networks and, we propose a new framework based on two graph convolutional (GCN) auto-encodeurs trained following a suspicion mechanism: the first GCN is trained on a sample suspected of being composed of normal entities while the second one on a sample suspected of containing anomalies. The final classification is done by coupling the result of both auto-encodeurs. We demonstrate that our approach obtains at least equivalent performances as state-of-the-art methods in the perfect sample case while being more resilient to the introduction of mistakes in these labeled samples.