Original title: Benchmarking Spatial, Spectral, and Self-Supervised Cues for Face Forgery Detection under Realistic Degradation
Lucas Cunha, Lucas Sotomaior, Lucas Gasperin, Beatriz Caldas +2 · September 2026
Spotting fake face images is hard when pictures get blurry, compressed, or resized in the real world. Computer programs that spot fakes work well on clean test pictures, but real images are rarely clean. Researchers tested six detector families on the Multi-Dimensional Face Forgery Image dataset using clean and degraded test sets. They tested normal pixel pictures, frequency-based math views called spectral cues, and combinations of both. The detectors worked by learning either tiny visual errors or overall face shapes. The team found that doing well on clean images does not mean a model handles real flaws well. A model called Xception was top on clean pictures, but dropped heavily on degraded ones. Meanwhile, a frozen model called DINOv3A type of AI model trained on images without labels to learn general features. held up best on degraded pictures by looking at broad facial structure.
Imagine you're trying to spot a counterfeit bill by looking at tiny details like microprinting. But if the bill is worn or folded, those details are hard to see. In this paper, AI models are tested on fake faces in clear versus degraded images. Models relying on fine details fail when those details are obscured, while a model trained to recognize overall patterns performs better under degradation.
Face forgery detectors help spot deepfakes that can spread misinformation or fraud. If they fail when images are blurry or compressed, fake content might go unnoticed, undermining trust in digital media.
This work establishes a benchmark and evaluates existing models, but it does not propose a new forgery detection method or test on other datasets beyond MFFI. It also does not cover all possible types of image degradation.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Benchmarking Spatial, Spectral, and Self-Supervised Cues for Face Forgery Detection under Realistic Degradation — Lucas Cunha, Lucas Sotomaior, Lucas Gasperin et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.