| Abstract [eng] |
In this thesis, two deepfake detection models - SBI [SY22] and FSBI [HLK+25] were researched. A custom dataset of the most popular deepfakes was collected from social networks and used to validate the trained models. The original SBI and FSBI models trained as specified by the authors, but the resulting accuracies fell short of those reported by authors. When individual architectural components were replaced, accuracy improved across all experiments on the custom dataset, whereas on the CDF [LYS+20] dataset accuracy improved in only one of the three experiments; in every case, however, inference speed increased substantially. Experiments with additional or modified data augmentations produced substantial accuracy gains on both CDF and the custom dataset. The best overall performer was the FSBI25E+EfficientNetV2_S model, which reached 92.00% accuracy on CDF and 74.6% on the custom dataset. Nonetheless, every model trained in this work detected deepfakes considerably less reliably on the custom dataset than on CDF. |
