Culturally-Aware Financial Fraud Detection Using Vision-Language Models

Huangqi Jiang

Culturally-Aware Financial Fraud Detection Using Vision-Language Models

Huangqi Jiang

Published: 06 May 2025, Last Modified: 29 May 2025VLMs4All 2025 PosterEveryoneRevisionsBibTeXCC BY 4.0

Keywords: cryptocurrency scams, multilingual detection, cultural adaptation, vision-language models, financial fraud, NLLB, ViLBERT

TL;DR: We present Multilingual VLM framework detecting culturally-adapted crypto scams with 92% accuracy, introducing language-aware detection and fairness benchmarks

Abstract: We present a multilingual vision-language framework that overcomes the limitations of English-centric approaches through three innovations: (1) a language-aware pipeline combining NLLB for multilingual processing with ViLBERT’s multimodal analysis, achieving 92% scam detection accuracy (38% improvement over monolingual baselines); (2) cultural signal recognizers identifying high-risk markers (e.g., religious appeals in Arabic, unrealistic returns in Mandarin) with 0.87 F1-score; and (3) CryptoScam-18, the first multilingual benchmark covering 18 languages, enabling fairness evaluation (Δbias < 0.15). Experiments demonstrate superior performance while maintaining <100ms latency, providing critical tools for combating culturally-adapted financial fraud in decentralized ecosystems.

Submission Number: 21

Loading