FDGAT-WTA: A dynamic detection model for web tracking and advertising based on improved graph attention networks

Abstract

Web tracking and advertising (WTA) have become pervasive on the Internet, presenting significant challenges to user privacy and data security. Although current defense mechanisms, such as filter list based interceptors and machine learning methods, provide a solution, they do not perform well in complex network environments with missing features, and their large size makes both performance and overhead subject to optimization. This paper introduces FDGAT-WTA (Fine-tuning Dynamics GAT for WTA Detection), a dynamic model based on an improved graph attention network, designed for efficient WTA detection. The model constructs network traffic as a Homogeneous Directed Multigraph (HDMG) and modifies the graph attention aggregation strategy, enabling deep feature extraction and dynamic graph extension through transductive and inductive learning methods. The dynamic detection phase leverages pruning techniques to reduce computational load and memory usage. The experimental results show that compared with existing machine learning based WTA detection methods, FDGAT-WTA has improved detection performance by about 5%, reduced model overhead by about 25% under the same data scale , and can adapt to real complex network environments with partially missing features with minimal performance loss, realizing lightweight and efficient dynamic detection.