As criminal organizations expand their use of information technology, the frequency of global financial fraud rises to historic levels each year, posing significant challenges for governments and the financial system. According to a Nasdaq Verafin research report, global financial and bank crime generated US$485.6 billion in losses in 2023, with total unlawful money transfers totaling US$3.1 trillion.
The "Otter Coding" team discovered during the "2025 AI Wave: Taiwan Generative AI Applications Hackathon" that conventional financial institutions frequently employ manual review and outmoded model screening to prevent such incidents from occurring. This method is both time-consuming and difficult to address in a timely manner. The team proposed a financial fraud detection system that enhances financial institutions' anti-fraud response capabilities by integrating AI and other advanced technologies.
The "Otter Coding" team ultimately triumphed in the highly competitive "Financial Innovation" group of Taishin Financial Holding.
Otter Coding claims victory in the Financial Innovation category with an AI-driven fraud detection system built to outsmart global financial crime. Credit: Company
AI-powered anti-fraud detection targets suspicious accounts
The "Otter Coding" team initiated a two-pronged approach to address the challenges associated with conventional financial fraud prevention. The first step is to employ the bank transaction records and fundamental account information supplied by Taishin Financial Holding to implement AI models for the identification of potential warning accounts. These accounts should be swiftly designated as warning accounts to enhance anti-fraud awareness and alleviate the staffing strain. The secondary purpose is to perform a comprehensive reverse analysis of the existing detection system utilizing large-scale language models to pinpoint possible areas for model enhancement. This will facilitate the ongoing development and enhancement of the anti-fraud system's accuracy.
Following the completion of the Generative AI Workshop offered by Amazon Web Services (AWS) and the analysis of the data content provided by Taishin Financial Holding, the "Otter Coding" team members utilized their after-work hours a week prior to the official competition to convene and contemplate the direction of AI model design. They proposed the concepts of data-driven, model optimization, and automation, and ultimately resolved to design and construct the model from five major stages.
During the model training phase, the team employed Amazon SageMaker Data Wrangler to integrate transaction records, essential account information, and alert data, perform feature engineering, and train and optimize the XGBoost classification model to achieve high precision and recall. Secondly, in the fraud prediction phase, transaction data is converted into feature vectors and input into the model for risk assessment, enabling the swift identification of high-risk accounts.
During the third phase of result evaluation, the "Otter Coding" team computed Precision, Recall, and F1-score by juxtaposing the predictions with the actual list of fraudulent accounts and performed a comprehensive study of the error types.
In the fourth round of AI analysis, the team utilized Anthropic Claude 3.5 Sonnet v2 model on Amazon Bedrock, a fully managed service that makes high-performing foundation models, to conduct second-level risk analysis, provide model optimization recommendations, and assist financial specialists in identifying deception patterns.
Ultimately, the team employed AWS Amplify, the frontend development tool, to deploy the frontend UI and stored the data in Amazon S3, resulting in an immediate and user-friendly visual analysis interface, during the cloud deployment and presentation phase.
"Otter Coding" commended the comprehensiveness of both AWS's Generative AI Workshop and DIGITIMES' event planning and venue preparation. Building on this success, the team aims to achieve similar outstanding results in future competitions.
Article edited by Sherri Wang
