This paper tackles the problem of detecting deepfakes of Slovene politicians in the real-world environment of social media. For the purpose of empirically evaluating deepfake detection methods, we constructed a dataset containing 8000 facial images of eight Slovene politicians. The dataset consists of 4000 images extracted from publicly available deepfake videos posted on Facebook and Instagram, and 4000 images extracted from authentic videos published by different national medias. Using this dataset, we evaluated seven established deepfake detection methods: CapsuleNet, CORE, FFD, MesoNet, RECCE, SRM, and Xception. The results show that the performance of the evaluated methods significantly degrades when applied to real-world deepfakes instead of standardized benchmark datasets. Furthermore, several methods systematically confused the authentic and manipulated image classes. Qualitative analysis revealed that artefacts in modern political deepfakes are often local, short-lived, and subtle. The obtained results confirm a substantial gap between laboratory evaluation of deepfake detectors and their practical applicability to real political content distributed through social media.
|
