Real ou Fake

Inspiration

Generative AI tools can now create hyper-realistic images with a single prompt. While this drives incredible creator expression, it also poses a massive threat to digital trust. When we looked at existing AI-detection tools, we noticed a critical flaw: they only work in laboratory conditions.

In the real world—especially on high-velocity short-form content platforms —an image is never just "raw." It gets re-encoded, compressed into WebP/JPEG, downscaled, filtered, and converted into video frames. standard detectors rely on fragile, high-frequency microscopic pixel noise that completely vanishes after being processed through social media pipelines. When these artifacts disappear, baseline detection accuracy collapses from over 90% down to a coin-flip.

What it does

Real ou Fake is a resilient content-verification engine built to detect AI-generated visuals even after aggressive post-processing, compression, and platform re-encoding.

How we built it

Tech Stack: ML Framework: PyTorch, TorchVision, ONNX Runtime (for accelerated CPU/GPU inference)

Challenges we ran into

Early in the hackathon, our initial model accuracy plummeted from 94% on raw test sets to 52% as soon as we ran an image through basic TikTok-style video encoding. Fixing this required rewriting our training loop to prioritise frequency-domain wavelets over raw spatial pixels.

Accomplishments that we're proud of

Successfully integrated visual Grad-CAM heatmaps, allowing moderation teams to see why an image was flagged rather than relying on a black-box model score.

What we learned

A model is only useful for platforms operating at TikTok's scale if it can run fast enough to handle millions of uploads per minute. Learning to optimise with ONNX and weight quantisation proved just as crucial as tuning our hyper-parameters.