Deep support vector data description of anomaly detection with positive class edge outlier exposure and maximum double hypersphere interval

The existing deep support vector data description (DSVDD) models for anomaly detection still have room for improvement. These models cannot guarantee that the features of abnormal samples will fall outside the decision hypersphere, thus posing a risk in anomaly detection decisions. To eliminate this risk, this paper proposes a deep support vector data description of anomaly detection model with positive class edge outlier exposure and maximum double hypersphere interval (DSVDD-OEDH). The model comprises an algorithm for constructing a positive class margin outlier exposure set to assist in controlling the decision hypersphere for anomaly detection, as well as an algorithm for maximizing the distance between the concentric double hyperspheres that isolate the positive class training samples and the positive class margin outlier exposure set. The paper also presents one theorem that theoretically prove that the anomaly detection SVDD model with an appropriate outlier exposure set achieves a higher correct detection rate than the one without an outlier exposure set. Experimental results demonstrate that the anomaly detection model proposed in this paper significantly outperforms other deep anomaly detection SVDD models.