There are many factors that affect the accuracy. Compression, resizing, and cropping can dramatically affect results. Low quality images cause a decrease in accuracy because the detector is given less information to make a decision. Rare corner cases concerning the image content can also cause a decrease in accuracy (e.g. different viewpoints, scenes, lighting…etc.). If a model has never seen, for example, images of deepfake cats skydiving, then it’s possible that it will misclassify images of this nature.
At Reality Defender, it’s important that we consider all cases and can generalize to novel unseen image content/files. Generalize/generalization is a term used in AI that means the ability of algorithms to perform well on unseen images (unseen during training).