Meta-CAD: Few-shot anomaly detection for online social networks with meta-learning

Abstract

Online social networks are now an integral component of our daily life, yet they pose several security risks, notably including fraudulent activities. Promptly detecting anomalous behaviors within these networks is crucial for effective platform management. Existing unsupervised anomaly detection methods in social networks struggle to effectively distinguish anomalies from noise, leading to a high false alarm rate. It is also data-hungry like semi-supervised methods, making it challenging to cope with data scarcity in practice. To tackle these difficulties, we propose Meta-CAD, a Contrastive learning-based Anomaly Detection method in Meta-learning framework. It leverages a meta-learning framework to learn common and essential information from multiple auxiliary graphs, enabling efficient knowledge transfer and excelling in scenarios with limited data. Additionally, we design an anomaly-sensitive loss function inspired by contrastive learning, which allows the model to concentrate more on the characteristics of anomalous data by constructing positive and negative sample pairs, thereby enhancing the performance of anomaly detection. The experimental results show that Meta-CAD demonstrates superior performance, with its anomaly detection capabilities surpassing existing methods by up to 10%.