verilookie

Verilookie: Building a Unified, Gamified Defense Against Deepfakes and AI-Generated Media

The proliferation of deepfakes and AI-generated content presents a severe threat to digital trust, enabling widespread misinformation, fraud, and identity misuse. Verilookie was developed as a direct response to this challenge, offering users a powerful, multi-layered tool for media authenticity verification, augmented by an engaging, personalized learning experience.

Unified Detection Pipeline: A Multi-Layered Approach

The core innovation of Verilookie is its Unified Detection Pipeline, designed to provide conclusive results even when dealing with complex synthetic media. This pipeline operates in sequential layers:

1. Primary Deepfake Analysis: We utilize a pre-trained Vision Transformer (ViT) model, sourced from Hugging Face's prithivMLmods/Deep-Fake-Detector-v2-Model, as the first line of defense. If the model flags the media as a fake with high confidence (>= 0.7), the result is immediately returned. The backend handles video analysis by sampling and processing multiple frames using OpenCV.
2. Secondary AI Generation Check: If the deepfake analysis is inconclusive (confidence below 0.7), the request is forwarded to the NVIDIA AI API for AI-generated image detection. This ensures we catch content created by generative models (like Midjourney or DALL-E) that might bypass traditional deepfake models.
3. Unified Result: The system combines the scores from both layers to deliver a single, contextualized verdict (Authentic, Deepfake, or AI-Generated) with a confidence score and a clear explanation.

Adaptive Learning through Gamification

Recognizing that technology alone is insufficient to counter evolving threats, Verilookie focuses heavily on user education via gamification. The goal is to cultivate 'digital resilience' in users.

• AI-Powered Quizzes: The quiz generation system leverages the Google Gemini API. Based on a user's quiz history, the system dynamically generates new, context-aware multiple-choice questions about deepfakes and AI detection methods.
• Adaptive Difficulty: Our QuizGenerator calculates an adaptive difficulty level (1 to 5) by tracking the user's weighted average accuracy across recent quizzes. This ensures that quizzes are always challenging but tailored to the user's current knowledge base.
• XP and Progression: Users earn Experience Points (XP) for completing detections and quizzes. The XP system drives progression through defined levels (Novice, Apprentice, Expert, Master, Grandmaster) and unlocks various achievements and badges, making the learning process engaging.

Technical Architecture

Verilookie is structured as a full-stack application leveraging modern cloud technologies:

• Backend: Built with Python and FastAPI, providing a high-performance REST API layer. The backend manages the detection logic, integrates PyTorch/Hugging Face models, handles file uploads (using python-multipart and temporary file storage), and communicates with external AI services (NVIDIA and Gemini).
• Frontend: Developed using React (JavaScript), offering an intuitive user interface for file submission, result display, and interactive quiz components.
• Containerization: The entire application is containerized using Docker, ensuring consistent deployment environments and ease of setup.

Prototype Status and Future Enhancements

Currently, Verilookie functions as a highly effective prototype. While the core AI detection and gamification systems are robustly implemented, the current persistence layer uses local JSON files for storing user performance and quiz data. Future development will focus on scaling this foundation by integrating a robust SQL or NoSQL database for production scalability, implementing a full user authentication system, and optimizing the detection pipeline latency for real-time video analysis.