Adversarial GAN-Driven Coverless Audio-Video Steganography via Multi-Modal Feature Banks

Abstract

In this paper, a new audio-video steganography system will be presented incorporating adversarial generative models and Multi-Modal Feature Banks (MMFB) to be used as a secure and invisible channel to transmit multimedia messages. Compared to conventional embedding techniques, the proposed Adversarial GAN-Driven Coverless Audio-Video Steganography (AG-CAVS) do not require any carrier changes by embedding messages on the generative latent space of synchronized audio-video sequences. The system uses dual-stream Generative Adversarial Networks (GANs), in which common feature banks are used to support dynamic correspondence between temporal (audio) and spatial (video) encodings. An integrated optimization objective includes adversarial realism, cross-modal matching and message recovery constraints which ensure resistance to both statistical and adversarial steganalysis. Experimental results on the UCF101, VoxCeleb2, and Kinetics datasets show that AG-CAVS can be very imperceptible (Peak Signal-to-Noise Ratio (PSNR) = 41.8 dB, Structural Similarity Index Measure (SSIM) = 0.987), State-of-the-art steganographic models have a bit error rate of 0.07% and resistance to both Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) attacks which is nearly double that of their counterparts based on GAN, and the minimal error rate (= 0.987). The given solution proves the feasibility of coverless, adversarial-resistant, and multimodal secure communication to higher-level intelligent multimedia systems.