A customized image editing framework for diverse prohibited and restricted products in illegal online transactions.

The circulation of prohibited and restricted goods in online transactions seriously violates consumer rights and threatens public safety. However, the lack of a dataset for prohibited and restricted goods makes it difficult to regulate such illegal online transactions. Therefore, a multimodal dataset for prohibited and restricted goods is proposed, including 38,513 images and 77,026 texts. Nevertheless, because of the diversity and potential adversarial modifications of prohibited and restricted goods, intelligent recognition of such items still faces significant challenges. Thus, an image editing framework for prohibited and restricted goods in online transactions is proposed. This framework integrates three novel components: (1) a PR-adapter that optimizes image prompts through image augmentation and compression representation techniques; (2) a text description generator combining the CLIP model and a multimodal large language model (MobileVLM) to generate more precise textual descriptions of images; and (3) an image generator, including a new loss function designed to fine-tune the stable diffusion model, enabling a better understanding of text semantics and generating images that more closely align with the textual descriptions. Experimental results show that this framework can generate diverse and accurate images of prohibited and restricted goods, effectively enhancing the development of intelligent supervision for online transactions.