In the rapidly evolving digital landscape, network security has become a paramount concern impacting national security, economic development, and social stability. Traditional network flow watermarking techniques face significant limitations in high-speed network environments, including low detection efficiency, high computational overhead, and weak resistance to network interference. This paper addresses these challenges by presenting a novel network flow watermarking segmentation detection technique based on abstract feature extraction. The proposed method divides watermark features into multiple segments, employing fast binary comparison for initial filtering and in-depth analysis for detailed detection. Key features are extracted across temporal, frequency, and statistical dimensions, with Support Vector Machine (SVM) models used for final classification. Simulated experiments under varying traffic conditions (2 to 10 Gbps) demonstrate that the system maintains high accuracy (>99.6%) and precision (100%), though recall fluctuates between 65% and 72%. The results indicate a robust detection mechanism but highlight scalability challenges due to non-linear increases in processing delay with higher traffic volumes. The proposed method shows significant potential for real-time watermark detection in high-speed network environments, offering valuable support for network attack tracing and advanced persistent threat (APT) detection.