Cross-Modal Steganography Detection Using Multimodal Large Language Models: A Comparative Study of Zero-Shot Prompting, Classical Steganalysis, and Parameter-Efficient Fine-Tuning

Classical steganalysis tools such as the Spatial Rich Model (SRM) assume stationary natural-image sensor noise, an assumption violated by synthetic imagery from GANs and diffusion models. We investigate whether a multimodal large language model (MLLM), LLaVA-1.6-Mistral-7B, can detect least significant bit (LSB) steganography by reasoning about texture and noise rather than a fixed statistical model. On a 16,000-record dataset built from CIFAR-10 and CelebA covers with LSB embedding at 10%-30% capacity, plus JPEG and Gaussian adversarial variants, we compare zero-shot chain-of-thought prompting, an SRM baseline, and LoRA fine-tuned LLaVA with answer-only label masking on a stratified 200-image test set. Zero shot LLaVA is near-chance (AUC = 0.526). SRM reaches AUC = 0.730; however, its F1 collapses from 0.850 to 0.546 at 5% FPR. LoRA fine-tuning reaches AUC = 0.807 and F1@5%FPR = 0.743, meeting both targets. We show that the decisive factor is loss masking: unmasked training loss is unstable (=4.8), while masking non-answer tokens to −100 yields convergence from 0.117 to 0.022 over 10 epochs. The pipeline runs as a self-correcting SLURM job on four A100 GPUs.