The rapid advancement of generative artificial intelligence (GenAI) has significantly enhanced the efficiency of content creation and dissemination. At the same time, it has accelerated the proliferation of misinformation, manipulated content, and digital fraud, posing increasing challenges to democratic governance, social stability, and the integrity of digital trust and information ecosystems. In this context, achieving a balance between technological innovation and risk governance, while strengthening a trusted information environment, has become a key priority for both government and industry in Taiwan.
Concurrently, the global expansion of virtual asset investment has prompted jurisdictions worldwide to strengthen regulatory frameworks governing digital asset transactions. Despite these efforts, fraudulent activities involving virtual assets continue to evolve in both scale and complexity. Malicious actors frequently exploit nominee accounts to conduct layered money laundering schemes, resulting in delayed fraud detection and challenges in asset recovery. Conventional approaches -including static blacklists and rule-based controls-are increasingly insufficient to address these evolving threats. As a result, the development of highly interpretable, AI-enabled early warning mechanisms to support timely risk identification and mitigation has emerged as a critical focus for both policymakers and industry.
Against this backdrop, Team Bibi Lab, comprising students from National Tsing Hua University (NTHU), was awarded top honors in the "Virtual Asset Transaction Security" category-sponsored by BitoPro, a cryptocurrency exchange under BitoGroup-at the "Agent for Truth: Disinformation Defense Hackathon." Their project, "AI Anti-Fraud Guardian for Virtual Currency Transactions," demonstrated a high degree of technical innovation and real-world applicability, earning strong recognition from the judging panel.
Team Bibi Lab extracted 140 features across 18 categories from KYC, fiat currency, and cryptocurrency transaction data provided by BitoPro. At the core of their approach is an innovative "three-layer risk tracking mechanism," which analyzes shared wallets, shared IP addresses, and indirectly connected users to trace how risk propagates through the transaction network-forming one of the system's most critical signal sources.
To ensure high interpretability and support compliance requirements, the team deliberately avoided complex ensemble models that could obscure decision logic. Instead, they adopted LightGBM combined with Focal Loss, enabling the model to focus on hard-to-detect minority cases and effectively address severe data imbalance.
In addition, the team constructed a network graph of internal transfers, shared wallets, and common IP addresses using NetworkX. By applying algorithms such as PageRank and community detection, they identified critical relay nodes within fund flows, further strengthening the system's ability to uncover hidden fraud patterns.
The competition was powered by Amazon Web Services (AWS), whose cloud infrastructure enabled Team Bibi Lab to build a comprehensive AI-driven anti-fraud system. Their solution features a four-layer interpretability architecture: "Continuous Risk Scoring and Tiering," which classifies users into four risk levels from low to extremely high; "Feature Deviation Analysis," which visualizes how user behavior diverges from baseline norms; "Rule-Based Interpretation," which generates automated textual explanations; and an "AI Risk Diagnosis Report," which leverages the Claude 3.5 Haiku on Amazon Bedrock to generate professional compliance analysis reports.
The system is built on a robust AWS cloud stack. The solution is deployed on Amazon Elastic Compute Cloud (Amazon EC2), with raw data and feature matrices stored in Amazon Simple Storage Service (Amazon S3), and AWS Glue handling ETL and feature engineering. AWS Lambda supports batch risk scoring, while AWS Step Functions orchestrates the end-to-end workflow. For high-risk cases, Amazon Simple Notification Service (Amazon SNS) sends real-time alerts to compliance teams, and Amazon CloudWatch ensures system monitoring and alerting-demonstrating the breadth and scalability of AWS cloud services in supporting advanced AI application development.
Team Bibi Lab noted that, despite their background in AI development, they had no prior experience with AWS cloud services. Prior to the competition, AWS organized a series of technical workshops-covering AI applications and enterprise data-enabling the team to quickly build proficiency with relevant tools and services.
With guidance from BitoPro mentors and industry experts, the team deepened its understanding of cryptocurrency operations. They also incorporated key design principles-including "auditability for compliance personnel" and "automated risk alerting" -into their solution, ultimately delivering a practical, deployable anti-fraud system that contributes to a more resilient and trustworthy digital financial ecosystem.
