MSSN: Multi-Stream Steganalysis Network for Detection of QIM-Based Steganography in VoIP Streams

Quantization index modulation (QIM) based steganography can hide secret information in voice-over-IP streams. Misuse of QIM-based steganography is a potential threat to network security. Although deep learning models have achieved great success in detecting QIM-based VoIP steganography, the detection performance still has a large room for improvement. In this paper, we propose a novel and efficient QIM-based steganography detection network named the multistream steganalysis network (MSSN). The architecture of MSSN is designed based on the characteristics of the codewords to capture the correlation among them. Unlike previous studies that only take the LSP codewords as input, MSSN utilizes all the encoding codewords to obtain more discriminative features with the multi-stream architecture. Meanwhile, a dual-pathway strategy is adopted to extract and enhance multi-level latent features, as well as suppress the noise signals. The multistream architecture with dual-pathway (MSDP) allows MSSN to capture both the sequential and functional correlations among the encoding codewords. Experimental results show that MSSN has superb detection performance that outperforms other stateof-the-art QIM-based VoIP steganalysis models. Also, MSSN's response time is only slightly longer than that of some lightweight models, which indicates that MSSN is practical for performing real-time steganalysis.