Hi HN,

I built VeridisQuo to tackle a problem with most deepfake detectors: they tell you something is fake, but not why or where.

The approach:

- Spatial analysis (EfficientNet-B4) catches visual artifacts like texture inconsistencies and blending boundaries

- Frequency analysis (DCT + FFT) catches compression artifacts that are invisible to the human eye but characteristic of deepfake generation

- GradCAM generates heatmaps showing which regions of each frame triggered the detection

The result is a video where you can see exactly which parts of someone's face looked manipulated.

Trained on 716k faces from FaceForensics++ (Face2Face, FaceShifter, FaceSwap, NeuralTextures). ~4 hours on an RTX 3090.

Live demo: https://huggingface.co/spaces/Gazeux33/veridisquo-deepfake-d...

I'd love feedback on: - The frequency fusion approach (DCT + FFT → MLP) - Ideas for improving generalization to newer deepfake methods - Whether temporal modeling (LSTM/Transformer) would help

Happy to answer questions about the implementation.