Detecting “DeepFakes” in H.264 Video Data Using Compression Ghost Artifacts

Abstract

Fast track article for IS&T International Symposium on Electronic Imaging 2020: Media Watermarking, Security, and Forensics proceedings.

Detecting Deepfakes with Haralick’s Texture Properties

“…However, the classifier improves upon our previous approach [28] based on ghost artefacts by great extent.…”

“…The performance of our proposed approach and several existing methods, which where evaluated on the FaceForensics++ dataset, are displayed in Table 1. We also compare the proposed method of this paper with our previous approach based on compression ghost artefacts, see [28].…”

A survey of deepfakes in terms of deep learning and multimedia forensics

“…The accuracy of this method reaches about 99.68% in detecting GAN images. A new approach to detect deepfake is proposed by [59] based on the "JPEG ghost" algorithm. This algorithm recognizes tampered faces from real ones by analyzing incompatible compression errors.…”

Deepfakes Generation and Detection: A Short Survey

“…For instance, Koopman et al [ 62 ] analyzed photo response non-uniformity (PRNU) for detection. Also, warping artifacts [ 63 ], eye blinking [ 64 ], optical flow with CNNs [ 65 ], heart rate [ 66 ], image quality [ 28 ], local image textures [ 37 ], long short-term memory (LSTM) and recurrent neural network (RNN) [ 67 ], multi-LSTM and blockchain [ 68 ], clustering [ 69 ], context [ 70 ], compression artifacts [ 71 ], metric learning [ 72 ], CNN ensemble [ 73 ], Identity-aware [ 74 ], transformers [ 75 ], audio-visual dissonance [ 76 ], and multi-attentional [ 77 ] features were used. Very few works have been focused on deepfake detection method’s explainability (e.g., [ 78 ]) and generalization capability (e.g., work of Bekci et al in [ 38 ] and Aneja et al [ 79 ] work using zero-shot learning).…”