Project · computer vision · open source
xdfdet
Explainable deepfake detection
Eschatia Labs built xdfdet to check each frame of a video for AI-altered faces. It uses EfficientNet-B4, an image recognition model, and Grad-CAM heat maps that show which parts of the face each decision relied on. Its best model reaches 0.897 AUC on FaceForensics++, where 0.5 is guessing and 1 is perfect.
What it does
xdfdet sorts video frames into real and altered and explains each decision with a Grad-CAM heat map over the face. The study behind it compares nine training methods, including augmentation, which varies the training images, and cutout, which blacks out parts of them. It measures accuracy, how closely confidence matches results, and which parts of the face each model relies on.
On FaceForensics++, a public collection of real and altered face videos, EfficientNet-B4 reaches 0.897 AUC with varied and partly blacked-out training images, versus 0.868 without them. The repository holds the training and evaluation code and eight trained models.
Materials
Eight EfficientNet-B4 models on Hugging Face and Kaggle, the training and evaluation code, the UBMK 2026 paper and the MSc thesis it grew from.