Steganalysis of VoIP Streams via Bidirectional Correlation Extract Network

Voice over Internet Protocol (VoIP) steganography based on Quantization Index Modulation (QIM) has become one of the mainstream audio steganography methods due to its imperceptibility. But currently, existing steganalysis methods for QIM-based steganography are inadequate in detection performance, especially in cases of low embedding rates and short sample lengths. To address this issue, this paper puts forward a steganalysis method based on Bidirectional Temporal Convolutional Network (Bi-TCN) and Bidirectional Gated Recurrent Unit (Bi-GRU), combined with a multi-head attention mechanism. Its feature extraction framework can efficiently extract codeword correlations in VoIP speeches and detect whether the speeches contain secret messages. Additionally, compared to traditional Recurrent Neural Network (RNN) models, the proposed model accelerates the training speed and avoids the gradient vanishing. Experimental results show that the proposed model outperforms the state-of-the-art models in terms of detection accuracy, particularly when the embedding rate is lower than 50% and the speech sample length is shorter than 1.0 s.