Describe your project

Our project addresses the challenge of verifying video authenticity and detecting tampering to build trust in digital media. With the rise of misinformation, especially through video content, ensuring integrity is essential. Our solution focuses on three key objectives:

User Authentication: We use Google OAuth 2.0 to authenticate users uploading videos, ensuring content creators are verified. Credentials are stored in Firestore for an audit trail.

Metadata Signing: After upload, video metadata (e.g., frames, credentials) is extracted, and a public-private key pair is generated. The video is signed with the private key, embedding the signature into the metadata for future verification using the public key, also stored in Firestore.

Tampering Detection: APIs and models detect tampering. Google Gemini AI and Perlin Noise analyze audio modifications, while Google Speech-to-Text checks transcription discrepancies. Google Video Intelligence flags sudden scene changes or object insertions. A Frame Splicing Model identifies frame-level tampering, and our deepfake detection model analyzes facial and speech mismatches.

Out of Scope:

Real-time tampering detection during live streams.
Analysis of non-video media (audio-only or text).
Content creation and editing tools.
Offline video validation.
Future Opportunities:

Real-time detection for live streams.
Integration with platforms like YouTube and TikTok.
Blockchain-based video verification.
Automated flagging of manipulated videos.
Cross-media tampering detection (images, audio).
Enhanced deepfake detection models.
User-friendly apps for verifying video integrity.
In conclusion, our solution offers robust video tampering detection and content integrity assurance, with future potential for real-time detection, cross-media support, and blockchain integration.

Challenges we ran into

One of the significant challenges we faced while building this project was our limited experience with video manipulation and processing. Most of the team had not worked extensively in this domain before, so we had to learn the techniques from scratch. Additionally, we were using several unfamiliar technologies, such as Supabase and the Video Intelligence API. The learning curve for these technologies was steep, but with constant reference to documentation, online tutorials, and the help of Gemini, we gradually mastered the tools and applied them to our project.

A unique hurdle was the lack of similar implementations or research papers we could draw from. Since we were working in a relatively unexplored area, we found it difficult to locate resources or examples to guide us. Understanding and adapting complex research papers to fit our needs was especially challenging in a domain we weren’t familiar with. However, through collective effort and persistence, we developed a solid understanding of the concepts, allowing us to integrate them into our project as envisioned.

Another technical challenge involved hosting the models and dependencies. Our project required substantial memory and computational power, making it impossible to run on free hosting services. As students with limited financial resources, we couldn’t afford paid hosting services. This forced us to explore alternative solutions, optimizing parts of our project to reduce resource demands.

Finally, balancing the academic workload with project deadlines proved to be a tough task. Between classes and assignments, we had to carve out time for coding, which meant working late nights and, occasionally, during lectures. Coffee, teamwork, and a shared determination helped us push through this balancing act and stay on track.