Anti-fraud hackathon winner showcases multi-dimensional tech-nology framework
In Taiwan, scams have evolved from isolated tactics into cross-channel, multi-step attack schemes. Most victims are aware of fraud; rather, they are often driven to make poor decisions under conditions of high pressure, limited information, and tight time constraints. Enabling individuals to access reliable risk assessments and timely guidance before taking critical actions has therefore become an urgent priority.At the same time, rapid advances in generative AI have accelerated the proliferation of sophisticated misinformation, deepfakes, and fraud schemes, posing growing threats to social trust and public safety. Strengthening information verification and risk mitigation through AI has thus emerged as a pressing global challenge.In the "Agent for Truth: Disinformation Defense Hackathon," Team (1), formed by members from the Department of Electrical Engineering at National Taiwan University, was awarded the "Excellent Award" in the "Fraud Identification and Prevention" category, sponsored by Gogolook, for its solution "FakeOff."Fraud tactics continue to evolve rapidly, often leveraging current events to lower public vigilance. For instance, during the tax filing season, deceptive SMS messages tend to surge. Traditional anti-fraud models, which rely heavily on historical datasets, often lack the ability to proactively detect newly emerging fraud patterns.Team (1) explained that "FakeOff" addresses this limitation through a continuous learning and data alignment mechanism. By leveraging web scraping technologies to monitor major news platforms, the system captures real-time events and applies AI to identify content that could potentially be exploited by malicious actors. This approach enables model training and deployment prior to the large-scale spread of fraudulent messages, thereby strengthening early-stage fraud prevention.The competition was powered by Amazon Web Services (AWS), providing cloud infrastructure that enabled teams to build and deploy advanced AI solutions at scale. Team (1) utilized AWS infrastructure to build a comprehensive AI-driven anti-fraud system. The solution was deployed on Amazon Elastic Compute Cloud (Amazon EC2), with large-scale fraud datasets and training data stored in Amazon Simple Storage Service (Amazon S3). It also leveraged Amazon Titan Text Embeddings V2 on Amazon Bedrock for efficient text vectorization, highlighting the breadth and scalability of AWS cloud services in supporting advanced AI application development.From a technical architecture perspective, "FakeOff" integrates the visual language model Claude Sonnet 4.6 to analyze screenshots and textual content. The system adopts a multi-agent collaborative framework built on Claude Haiku 4.5, in which a Function Calling Agent dynamically invokes fraud detection tools, blacklists, and number lookup APIs based on visual cues. A Conclusion Agent then consolidates these outputs to generate interpretable assessment reports and actionable prevention recommendations.To address the core challenge of fraudulent message detection, "FakeOff" incorporates a cross-model alignment and voting mechanism. By integrating leading large language models-including GPT-4o, Claude Sonnet 4.5, Llama 3 70B, Mistral, and Qwen3-30B-the system performs cross-validation to identify models that best capture the nuances of the Chinese language. It further enhances performance by localizing and training on global datasets, effectively addressing the limited availability of fraud-related data in Taiwan.The solution also adopts a continuous learning architecture, combining Amazon Titan Text Embeddings V2 for text vectorization with a neural network classifier. This design enables ongoing model refinement through real-world data and human feedback, ensuring sustained accuracy and adaptability in an evolving fraud landscape.To counter emerging fraud tactics, "FakeOff" autonomously scans recent news across multiple platforms to extract high-risk keywords, enabling early detection of new scam patterns without requiring full model retraining. This forward-looking technical approach was a key factor in earning strong recognition from the judging panel.Team (1) noted that, with on-site support from AWS Solution Architects and mentorship from Gogolook, the competition enabled the team to expand its perspective beyond purely technical development to encompass real-world application scenarios. The team also expressed its aspiration to contribute to fraud prevention and strengthen information security.
