Deepfake DetectionDeepfake Detection

Edited faces are now hard to see. Their compression history is not.Edited faces are now hard to see. Their compression history is not.

When part of a JPEG is edited, that region has been compressed fewer times than the rest. Error Level Analysis resaves the image and measures the difference, so edited regions glow. A fine-tuned EfficientNet-B0 then learns to read those heatmaps.When part of a JPEG is edited, that region has been compressed fewer times than the rest. Error Level Analysis resaves the image and measures the difference, so edited regions glow. A fine-tuned EfficientNet-B0 then learns to read those heatmaps.

1. Feed the model forensics, not pixels. Classifying ELA heatmaps instead of raw images makes the signal the model learns explicit.

   Feed the model forensics, not pixels. Classifying ELA heatmaps instead of raw images makes the signal the model learns explicit.
2. Transfer, don't train from scratch. EfficientNet-B0 starts from ImageNet weights and is fine-tuned end to end.

   Transfer, don't train from scratch. EfficientNet-B0 starts from ImageNet weights and is fine-tuned end to end.
3. Different speeds for different layers. The backbone learns at 5e-5 to keep its edge filters; the head learns at 1e-3 with cosine annealing.

   Different speeds for different layers. The backbone learns at 5e-5 to keep its edge filters; the head learns at 1e-3 with cosine annealing.

accuracy, 2,000 unseen validation imagesaccuracy, 2,000 unseen validation images

86.4%86.4%

86.4%86.4%

precisionprecision

88.0%88.0%

88.0%88.0%

recallrecall

84.3%84.3%

84.3%84.3%

• ELA depends on JPEG history; PNGs and heavily recompressed images weaken the signal.

  ELA depends on JPEG history; PNGs and heavily recompressed images weaken the signal.
• Trained and validated on a balanced 10,000-frame subset of FaceForensics++ (80/20 split). The 2,000 validation images also chose the best epoch, so the figures are slightly optimistic; a separate test split is the next measurement.

  Trained and validated on a balanced 10,000-frame subset of FaceForensics++ (80/20 split). The 2,000 validation images also chose the best epoch, so the figures are slightly optimistic; a separate test split is the next measurement.
• One benchmark: other generators may behave differently.

  One benchmark: other generators may behave differently.
• Next: combine ELA with frequency-domain features and test on newer diffusion models.

  Next: combine ELA with frequency-domain features and test on newer diffusion models.