Imperceptible adversarial audio steganography based on psychoacoustic model

Recently, deep learning based audio steganalysis methods have demonstrated superior performance in detecting the conventional audio steganography, which poses great chanllegnes to the conveiontional audio steganography. In this work, observed that the neural network can easily be deceived by specially perturbed inputs, i.e., adversarial examples, we propose an imperceptible audio steganography method based on psychoacoustic model. Specifically, we first add perturbation on the stego audio for constructing noise stego audio, which is delivered to the trained steganalyzer for misclassification. The perturbation is optimized in the adversarial process, aiming to seek an optimal perturbation that guarantee the imperceptibility and undetectability of stego audio. Further consider that the difficulty to optimize the threshold loss function using gradient back-progagation, we adopt two-stage optimization strategy to minimize the loss function. The first stage attempts to find a suitable perturbation to deceive the steganalyzer. The second stage concentrates on further optimizing the perturbation to make the stego imperceptible. For the practical steganography, the optimal perturbation obtained from the adversarial attack process is added on the original cover audio to construct the adversarial cover audio. Then one can use information embedding algorithm to embed the secret message on the adversarial cover to generate stego audio. Extensive experiments show that the proposed method can generate the adversarial cover audio with high perceptual quality and the undetectability performance outperforms the conventional audio steganography schemes.

Automated summary

Recently, deep learning based audio steganalysis methods have posed significant challenges to conventional audio steganography by demonstrating superior performance in detecting it. In this work, the authors propose an imperceptible audio steganography method based on a psychoacoustic model, taking into consideration the vulnerability of neural networks to adversarial examples. They use a two-stage optimization strategy to minimize the loss function and add perturbation to the stego audio to deceive the steganalyzer. Experimental results show that the proposed method outperforms conventional audio steganography schemes in terms of imperceptibility and undetectability.