<p>This work presents the development of robust models for classifying posts and images on Dark Web forums. Existing approaches generally process text and images separately, limiting their capacity to capture the inherently multimodal characteristics of Dark Web content. In addition, conventional deep-learning techniques often struggle with the scarcity and domain-specific nature of Dark Web datasets. To overcome these limitations, we propose a hybrid deep-learning architecture that integrates a multimodal transformer with a co-attention mechanism to fuse textual and visual information, alongside Graph Neural Networks to model the forums’ structural interconnections. Textual data are encoded with BERT embeddings, while visual features are extracted using ResNet-50. The co-attention module dynamically aligns text and image representations to strengthen classification performance. Relationships among forum users and posts are represented through Graph Attention Networks, and topic distributions derived from Latent Dirichlet Allocation enhance both interpretability and accuracy. We further explore advanced transfer-learning strategies for Dark Web imagery, including model-agnostic meta-learning (MAML) for meta-transfer and domain-adversarial neural networks (DANN) for adversarial domain adaptation. These methods enable effective adaptation to Dark Web datasets and reduce domain discrepancies. Overall, the proposed hybrid approach is expected to achieve an improvement of approximately 10–12% points over current baselines, targeting about 90% accuracy for forum posts and 88–92% for images and samples, thereby advancing multimodal, cross-domain Dark Web content classification.</p>