Recognizing the Financial Fraud Type of Telegram Advertisement Posts Based on the Multimodal Learning

The instant messaging applications such as Tele-gram, Wire, and Signal have been widely used by cyber criminals for the fast propagation of illicit advertisements, particularly financial frauds. Examples of fraud include credit card scams, bank account drops, promo abuse/discount fraud, and money transfers. This paper explores the application of the multi- modal learning technique to analyze financial fraud activities on Telegram. We collected the data from 10 channels in the Telegram which including text and image. By combining text and image data, the multimodal learning especially the combination of Swin Transformer embeddings with attention-based fusion delivered the best results, which offers a more accurate approach to detecting financial frauds, capturing nuanced details that may be missed by single-modal learning methods. Integrating multiple data types allows for a higher accurate and Fl-score classification of fraud behavior, enhancing the capability to identify and prevent various types of frauds.