ML that holds up in the real app, not just the notebook.

The problem

Fake faces are getting harder to spot by eye. People need a quick way to check a suspicious image or video, and a way to check an original file hasn't been edited since it was registered.

What I built

• Flask app for single, batch and video checks
• Video pipeline that samples up to 24 frames, skips blurry or face-less frames and combines the rest into one verdict
• Quality warnings when a face is missing, tiny or blurred, instead of a falsely confident answer
• SHA-256 hashing and a Solidity smart contract on a local Ganache blockchain (Web3.py) to register and verify files

Results

Trained on Celeb-DF v2 with a 70/15/15 split by video. In final live testing the app got 37 of 40 images and 17 of 19 videos right (a small hand-picked set, so a sanity check rather than a benchmark).

What I learned

• Cropping to the face was the single biggest improvement
• Offline scores aren't the whole story: I rejected an 88% model because it failed in the real app
• Splitting data by video stops the model "remembering" people and inflating the score
• Honest limits: one dataset, and a simple face detector that misses side-on faces