Steganographic schemes dedicated to generated images modify the seed vector in the latent space to embed a message. Whereas most steganalysis methods attempt to detect the embedding in the image space, this paper proposes to perform steganalysis in the latent space by modeling the statistical distribution of the norm of the latent vector. Specifically, we analyze the practical security of a scheme proposed by Hu et al. for latent diffusion models, which is both robust and practically undetectable when steganalysis is performed on generated images. We show that after embedding, the Stego (latent) vector is distributed on a hypersphere while the Cover vector is i.i.d. Gaussian. By going from the image space to the latent space, we show that it is possible to model the norm of the vector in the latent space under the Cover or Stego hypothesis as Gaussian distributions with different variances. A Likelihood Ratio Test is then derived to perform pooled steganalysis. The impact of the potential knowledge of the prompt and the number of diffusion steps is also studied. Additionally, we show how, by randomly sampling the norm of the latent vector before generation, the initial Stego scheme becomes undetectable in the latent space.

In targeted adversarial attacks on vision models, the selection of the target label is a critical yet often overlooked determinant of attack success. This target label corresponds to the class that the attacker aims to force the model to predict. Now, existing strategies typically rely on randomness, model predictions, or static semantic resources, limiting interpretability, reproducibility, or flexibility. This paper then proposes a semantics-guided framework for adversarial target selection using the cross-modal knowledge transfer from pretrained language and vision-language models. We evaluate several state-of-the-art models (BERT, TinyLLAMA, and CLIP) as similarity sources to select the most and least semantically related labels with respect to the ground truth, forming best- and worst-case adversarial scenarios. Our experiments on three vision models and five attack methods reveal that these models consistently render practical adversarial targets and surpass static lexical databases, such as WordNet, particularly for distant class relationships. We also observe that static testing of target labels offers a preliminary assessment of the effectiveness of similarity sources, \textit{a priori} testing. Our results corroborate the suitability of pretrained models for constructing interpretable, standardized, and scalable adversarial benchmarks across architectures and datasets.