One-Class Classification for Mobile Ad Fraud Detection with Contrastive Learning

In recent years, mobile advertising has become essential in digital marketing. However, ad fraud has caused substantial economic loss in this industry. Traditional machine learning-based detection method fails to address the real-world challenge that the dataset for learning algorithm is extremely imbalanced. In this paper, we identify the ad fraud detection as a one-class classification task, and propose a novel oneclass classification method based on contrastive learning. The model is constructed with a Transformer-based dual contrastive learning algorithm. It mainly achieves the detection of fraud traffic in a large number of mixed ad traffic by learning the feature representation of the fraud ad traffic through known labeled data. Through experimental validation, the accuracy of our model reaches $96.39 \%$ and the precision rate reaches $98.05 \%$, thus demonstrating the superior performance of our method in detecting ad fraud traffic.