Recent research has uncovered that CSS style sheets can be leveraged for user fingerprinting by probing various characteristics, thereby circumventing existing detection mechanisms and mitigation strategies-posing serious privacy concerns. In contrast to the extensively studied JavaScript-based fingerprinting, no dedicated detection techniques have been developed specifically for CSS-based tracking. Furthermore, the inherent nature of CSS code, which lacks data flow and control flow structures and execution capabilities, complicates the analysis. To bridge this gap, we propose SDM-GAT, a novel detection approach that employs a Graph Attention Network (GAT) with a multi-head attention mechanism to effectively analyze graphs derived from CSS style sheets. In light of the limitations of existing datasets, we built a comprehensive dataset by collecting CSS content from Alexa’s Top 10,000 websites and embedding StylisticFP tracking code into selected sites. Then, We converted CSS code into graph representations, pruned these graphs to reduce computational complexity, and applied GAT to classify and detect tracking CSS code with high accuracy. Experimental results indicate that SDM-GAT outperforms traditional detection methods, demonstrating superior recognition accuracy and generalization capacity. The model achieves an accuracy of 98.01% and an AUC of 0.98, making it the first effective detection technique targeting CSS-based tracking. This study provides valuable insights and opens new avenues for enhancing privacy protection in web environments.