GenAI-AntiFraudX

Inspiration

The inspiration for GenAI-AntiFraudX stemmed from a critical observation: traditional anti-fraud education—static pamphlets, passive seminars, and text-heavy websites—is failing to keep pace with AI-driven cybercrime. In an era where scammers use GenAI to create hyper-realistic deepfakes and personalized phishing scripts, passive learning is obsolete.

We wanted to shift the paradigm from passive reading to active simulation. Our vision was to create a "digital vaccine": a system where users can safely experience and fight off simulated scams in a gamified environment, building "active immunity" before facing real threats.

What it does

GenAI-AntiFraudX is an interactive RPG platform that places users in realistic fraud scenarios. It features:

Real-Time Battle Simulation: Users engage in unscripted conversations with an AI "Scammer" who adapts tactics based on the user's responses.

Dual-Mode Operation: It runs Online for high-fidelity simulation and Offline for privacy and accessibility in remote areas.

Expert Intervention: An AI "Expert Agent" monitors the chat, providing real-time warnings and analyzing the user's vulnerability.

Post-Battle Analysis: A "Recorder Agent" evaluates the conversation, scoring the user's defense skills and providing actionable feedback.

How we built it

We engineered a Hybrid Multi-Agent System that bridges cloud power with edge privacy.

Hybrid Dual-Engine Architecture:

Online Mode (Gemini API): We utilized the Google Gemini API for our cloud-based simulation. Gemini's superior reasoning capabilities allow the "Expert Agent" to dissect complex, multi-layered fraud scenarios and enable the "Scammer Agent" to improvise highly psychological manipulation tactics.

Offline Mode (Ollama): To ensure accessibility, we implemented a local mode using Ollama running models like gemma3:4b and mistral:7b. This allows the system to function without internet, ensuring data privacy and zero latency.

The Multi-Agent Core: The backend, built with FastAPI, orchestrates four distinct agents:

Scammer Agent: Simulates 10 specific fraud tactics (e.g., Crypto, Romance).

Victim Agent: Simulates 4 personas (e.g., Elderly, Student) for demonstration purposes.

Expert Agent: Provides defense logic.

Recorder Agent: Handles scoring and analysis.

Retrieval-Augmented Generation (RAG):We use ChromaDB to store real-world fraud alerts.

Challenges we ran into

Harmonizing Two "Brains": Ensuring consistent behavior between the creative Gemini API and the more rigid local Ollama models was difficult. We built a RoleEnforcer utility to standardize outputs and prevent the AI from breaking character (hallucination).

Data Contamination: We needed to distinguish between synthetic data (AI vs. AI) and high-value human data. We designed a Data Classification System to separate auto_training data from player_mode data, ensuring future model fine-tuning relies on genuine human interactions.

Quantifying "Trust": mathematically defining how "tricked" a user is was complex. We derived a formula where trust fluctuates based on Scammer persuasion and Expert intervention .

Accomplishments that we're proud of

The "Arms Race" System: We built an automated evolution system where the Scammer and Expert agents train against each other, continuously upgrading their tactics based on successful/failed attempts.

Seamless RPG Integration: We successfully integrated complex LLM agents into RPG Maker, creating a user-friendly frontend that hides the complex backend logic.

Accessibility: By implementing the offline mode with Ollama, we made advanced AI education available to communities with poor internet connectivity or strict privacy concerns.

What we learned

Hybrid AI is the Future: Combining the raw reasoning power of cloud models like Gemini with the privacy of local models like Ollama creates the most robust solution for sensitive applications.

Engagement = Retention: Users retained anti-fraud knowledge significantly better when emotionally invested in a "battle" rather than passively reading warnings.

Guardrails are Essential: Prompt engineering alone is insufficient; robust structural guardrails (like our RoleEnforcer) are necessary for production-ready agents.

What's next for GenAI-AntiFraudX

Public Web Launch: We are moving from local prototypes to a hosted web platform using the Gemini API for broader public access.

Human-in-the-Loop Fine-Tuning: We plan to use the high-quality data collected from our "Player Mode" to fine-tune specialized small language models (SLMs) for even better offline performance.

Expanded "Arms Race": We will automate the feedback loop further, allowing the system to auto-generate new scam scenarios based on emerging real-world news reports.

Built With

• chromadb
• docker-&-docker-compose
• fastapi
• gemma-3-(4b)
• github-codespaces
• google-adk
• google-gemini-api
• javascript-(es6+)
• mistral-7b
• ollama
• pixi.js
• python-3.10+
• rpg-maker-mz
• sqlite
• uvicorn
• websocket