Detecting Deepfakes Through Inconsistencies in Local Camera Surface Frames

Disinformation is always more induced by resorting to altered or full synthetically generated realistic contents generally known as deepfakes. This poses the need of reliable instruments able to reveal such phony contents to avoid the polarization of persons’ thinking. Diverse techniques have been presented, whose common denominator is to search for some inconsistencies induced by the fake generation process and evidenced by looking at specific characteristics. It seems that the possible combination of various distinctive features could provide more effective achievements. In this view and following our previous study on this matter, we propose to investigate a new feature which basically takes into account of the built-in relations existing within the entire acquisition scene at the time of image (video) capturing and then represented by the image pixel values. The idea behind is that deepfake generation process can break such intrinsic dependencies among illumination (intensity and direction), object surfaces (position and reflectance) and acquisition noise, thus evidencing a possible manipulation. According to this, a feature descriptor, named Local Surface Descriptor, has been constructed and used to train CNN-based models for deepfake detection. Experimental results conducted on different operative contexts demonstrate a significant detection accuracy up to 90%.