This paper presents the first application of a graph neural network (GNN) utilizing the GraphSAGE architecture to VoIP voice stream steganography to address the computational complexity and generalization challenges of deep learning-based steganography analysis. By constructing a simple graph from VoIP streams and using GraphSAGE to capture both fine-grained information and high-dimensional patterns, we achieve high detection accuracy. Experimental results demonstrate a detection accuracy of over 98% even with samples as short as 0.5 seconds, and even under challenging conditions with low embedding rates, we achieve 95.17% accuracy, a 2.8% improvement over the best-performing existing methods. Furthermore, our proposed method demonstrates efficiency with an average detection time of 0.016 seconds for 0.5-second samples, demonstrating its suitability for online steganography tasks.
