High capacity speech steganography for the G723.1 coder based on quantised line spectral pairs interpolation and CNN auto-encoding

In this paper, a novel steganographic method for Voice over IP applications —called Steganography-based Interpolation and Auto-Encoding (SIAE)— is proposed. The aim of the proposed scheme is to securely transmit a secret speech hidden within another (cover) speech coded with a G723.1 coder. SIAE embeds the steganograms in four interpolated and quantised line spectral pairs (QLSP) vectors. In order to minimize the changes in the cover speech, the proposed approach uses a 1D auto-encoder to compress the payload, and this scheme only requires embedding eight bits in about 30% of the packets. At the receiver side, the secret data can be successfully expanded to its original size upon decoding. This represents a significant reduction in the number of modified bits compared to state-of-the-art schemes, and results in enhanced undetectability and decreased steganographic quality loss. The results show that the proposed auto-encoder scheme has a very high performance since it can compress the embedded data up to 80 times from its original size, leading to a steganographic capacity that exceeds one kilobit per second (kpbs). In terms of imperceptibility, which is a relevant property for speech-in-speech steganography, the proposed SIAE method entails a very imperceptible distortion, with an average steganographic quality loss not greater than 0.19 in terms of mean opinion scores (MOS). Last but not least, the proposed method evades steganalysis specifically targeted at speech steganography. The tested steganalytic methods fail in detecting the steganographic content produced with the proposed SIAE method, yielding classification results that are indistinguishable from random guessing.