Targeted pooled latent-space steganalysis and remediation scheme for generative steganography

The researchers propose a method for steganalysis in latent space rather than image space, demonstrating that the steganographic scheme proposed by Hu et al. for latent diffusion models, which is practically undetectable in image space, can be detected in latent space. Such schemes embed messages in generated images by modifying latent seed vectors; the paper finds that the latent vectors after message embedding are distributed on a hypersphere, whereas the unembedded cover vectors follow an independent and identically distributed Gaussian distribution, based on which the vector norms in the two cases can be modeled as Gaussian distributions with different variances, and a likelihood ratio test is derived to implement pooled steganalysis. The paper also studies the influence of whether the prompt is known and the number of diffusion steps on detection, and gives a remediation method: randomly sampling the latent vector norm before generation can restore the original scheme to undetectable in latent space.