A Multi-Expert Transformer Model for Click Fraud Detection

Click fraud is a persistent issue in the online advertising industry, resulting in considerable loss of revenue and engagement metrics. Existing fraud-detection techniques such as decision trees and gradient boosting do not cope well with new patterns of fraud and adversarial manipulation. The primary objective of this study is to create an adaptive and scalable detection framework that effectively identifies emerging patterns of click fraud without compromising user data privacy. To this end, we propose a transformer-based framework with expert networks that uses multi-head attention, BiLSTMs, GRUs, and fusion embeddings for fraud detection. Classification Feature Fusion is used to improve fraud detection accuracy in the multi-expert transformer model, while protecting sensitive data using differential privacy techniques such as noise and gradient clipping. The model was tested on the Avazu dataset [1] and surpassed XGBoost and Cascaded Forests with an AUC-ROC of 0.9830 and higher Precision, Recall, and F1 scores. These metrics demonstrate the scalability, robustness, and efficiency of the proposed model in mitigating click fraud, while improving online marketing analytics. This study highlights the potential of multi-expert transformer models for adaptive privacy-aware click fraud detection and their application in real-time advertisement bidding systems for comprehensive ad fraud detection and prevention.