AI Video Detection WebApp

Problem Statement 7:

As AI technologies rapidly integrate into our daily lives, concerns about privacy and security have become more urgent than ever. With the rise of powerful generative AI models, large-scale data collection, and cloud-based deployment, users face increasing risks: sensitive data leakage, identity theft, etc. This hackathon invites participants to explore solutions in the following areas:

Enhancing the privacy of AI systems themselves (Privacy of AI); and/or
Using AI to defend user privacy and security (AI for Privacy).

Inspiration:

In this day and age, AI content on social media has evolved to seamlessly replicate authentic video footages. Yet nowadays, AI detection sites only deal with deepfakes. But what about non-deepfake videos that look real. There are groups of users, particularly the elderly, who are not able to discern AI videos from real ones given how hyper-realistic some videos can be. We need to protect these groups of users so as to curb the spread of misinformation and inform them on true reality.

What it does:

This project is a machine learning-based system designed to detect whether an input video has been AI-generated or is real. By analysing sampled frames from videos, the model classifies each video with probability scores indicating the likelihood of it being AI-generated or authentic.

How we built it:

Extracts a fixed number of representative frames from any input video.
Utilizes a pretrained video transformer model fine-tuned specifically for distinguishing AI-generated videos.
Supports batch training and inference on local video datasets.
Automatically handles corrupted or missing videos during data processing.
Outputs interpretable probabilities for real vs. AI-generated content.
Saves trained model checkpoints for reuse or further fine-tuning.

Challenges we ran into:

Due to time constraints and limited hardware, we were only able to test about 1000 videos.
It was also difficult to find good video datasets as our solution is very broad and it aims to tackle AI videos in general. ## Accomplishments that we're proud of:
Given these constraints, our model managed to fine-tune to a level that is sufficient for a demo version

What we learned:

We learn what goes on in the process of training a model, including data pre-processing, which is probably the most important aspect of machine learning. We also get a glimpse of how fine-tuning the model works and we definitely gain insights on how we can improve for future AI hackathons