Inspiration
In an era where AI-generated misinformation is becoming indistinguishable from reality, the erosion of digital trust is a major societal challenge. We were inspired to build ReelTrust to give everyday users a "digital shield"—a way to verify if the person they are seeing on screen is a real, living human being or a sophisticated deepfake.
What it does
ReelTrust is an video authenticity verifier. It allows users to: • Live Verify: Use the front-facing camera to perform real-time liveness checks. • Analyze Uploads: Process existing videos from local storage to check for "AI signatures." • Authenticity Scoring: Instead of a simple "yes/no," the app provides an AI Authenticity Score. By analyzing facial micro-expressions, landmark stability, and physiological markers, ReelTrust calculates a confidence average across the entire video to classify it as "Authenticated" or a "Potential Deepfake."
How we built it
• Frontend: Built with Jetpack Compose for a modern, scalable, and responsive UI. • AI Engine: Integrated the Presage SDK to leverage high-fidelity facial mesh and micro-movement analysis. • Video Processing: Used CameraX for live capture and MediaMetadataRetriever for frame-by-frame extraction of uploaded content. • Architecture: Developed using a modular approach, separating reusable UI components from core AI logic to ensure the project can scale beyond the hackathon.
Challenges we ran into
The Presage SDK is not compatible with our machines, since we are running x64 Windows. We tried WSL Ubuntu but since the core architecture is wrong, and the dependencies required are outdated versions, we pivoted to using the Android SDK instead. Setting up the project with Gradle took a lot of time. One of the biggest hurdles was the "Black Box" nature of many AI models. Initially, real faces were getting low confidence scores. We had to develop a custom Heuristic Multi-Factor Analysis algorithm that balanced base face detection with more subtle "signs of life" like facial landmark stability and micro-expression frequency. Managing complex multi-module Gradle configurations and secure API key injection via BuildConfig also tested our technical limits.
Accomplishments that we're proud of
We are incredibly proud of building a functional deepfake detector that runs locally on a mobile device. Moving from a static design to a fully integrated system that can extract frames from a 4K video and process them through an AI graph in real-time was a huge technical win for our team.
What we learned
We gained deep insights into facial liveness detection and the physiological markers that AI still struggles to replicate perfectly. We also learned how to bridge the gap between low-level native AI SDKs and high-level modern UI frameworks like Jetpack Compose.
What's next for ReelTrust
• Audio Authenticity: Integrating synthetic voice detection to catch audio deepfakes. • AI face shape detection: use of AI to detect face shapes to not consider non-facelike objects • rPPG Integration: Leveraging the SDK's ability to detect heart rate via skin color changes (rPPG) for even higher liveness confidence.
Log in or sign up for Devpost to join the conversation.