Semi-supervised anomaly detection in dynamic communication networks

To ensure the security and stabilization of communication networks, anomaly detection is first line defense. However, their learning process suffers two major issues: (1) inadequate labels: there are many different kinds attacks but rare abnormal nodes in mt these atstacks; (2) inaccurate considering heavy network flows new emerging attacks, providing accurate labels for all very expensive. The label problem challenges existing methods because majority normal result a biased classifier while noisy will further degrade performance classifier. tackle issues, we propose SemiADC, Semi-supervised Anomaly Detection framework dynamic Communication networks. SemiADC approximately learns feature distribution with regularization from ones. It then cleans datasets extracts sasainaccurate by learned structure-based temporal correlations. These self-learning processes run iteratively mutual promotion, finally help increase accuracy detection. Experimental evaluations on real-world demonstrate effectiveness our which performs substantially better than state-of-art approaches without demand adequate supervision.