S človeško percepcijo podkrepljen razpoznavalnik globokih ponaredkov

The main challenge in deepfake detection is poor generalization to new and evolving manipulation methods. This thesis addresses this by integrating human visual perceptual cues as auxiliary supervision to guide more robust feature learning. To support this, a dataset combining Self-Blended Images (SBIs) with crowdsourced saliency maps and categorical inconsistency annotations was created. An EfficientNet-based multi-task model is trained to jointly perform deepfake classification and perception-driven tasks. Experiments on multiple benchmark datasets show that the model with a consecutive saliency map decoder trained on our dataset generalizes better than baseline detectors, indicating that localized human attention helps the network focus on stable, perceptually meaningful cues rather than method-specific artifacts.