In this paper, a frame-level steganalysis of Quantization Index Modulation (QIM) steganography in compressed speech streams is proposed for the first time. The proposed method builds a neural network classification framework based on multi-dimensional perspective of codeword correlations, which is inspired by cognitive biology. Four dimensions are employed: global-to-local, local-to-global, forward and backward. First, the codeword embedding method is utilized to map each codeword into a compact representation. Next, Bi-LSTM is used to consider the steganographic features in time sequence and reverse time sequence. Subsequently, a dual-thread attention mechanism is designed to extract local and global features at the same time. Finally, a channel attention mechanism is employed to increase the weight that contributes the most to the current task and the convolution and fully connected layers are used to generate the frame-level steganographic label. Experimental results show that the proposed method is effective and practical in frame-level detection tasks.
