A forensic evaluation method for DeepFake detection using DCNN-based facial similarity scores

Detecting DeepFake videos has become a central task in modern multimedia forensics applications. This article presents a method to detect face swapped videos when the portrayed person in the video is known. We propose using a threshold classifier based on similarity scores obtained from a Deep Convolutional Neural Network (DCNN) trained for facial recognition. We compute a set of similarity scores between faces extracted from questioned videos and reference materials of the person depicted. We use the highest score to classify the questioned videos as authentic or fake, depending on the threshold chosen. We validate our method on the Celeb-DF (v2) dataset (Li et al., 2020) [13]. Using the training and testing splits specified on the dataset, we obtained an HTER of 0.020 and an AUC of 0.994, surpassing the most robust approaches against this dataset (Tran et al., 2021) [37]. Additionally, a logistic regression model was used to convert the highest score into a likelihood ratio for greater applicability in forensic analyses.