Face manipulated deepfake generation and recognition approaches: a survey

ABSTRACT

With the progression of deep-learning techniques, digital media recording and synthesis media generation have become exceptionally easy. Due to open access of user-friendly deepfaking applications generated using Artificial Intelligence methods especially Generative Adversarial Networks (GANs) and Variational Auto-encoders (VAEs), synthesis of recorded media has been made more effortless. Digital media synthesized using such applications is termed as deepfake. Generation of realistic fake audio/video content poses critical threats to an individual and society at large. To curb the threat of deepfaking, numerous deepfake detection algorithms have been proposed. This paper presents a survey on state-of-the-art deepfake generation techniques, categorized into face swap, attribute manipulation, and lip-sync manipulation. An analysis of some recently developed techniques that can generate images from text is also presented in the proposed paper. In addition, state-of-art image and video datasets released in the domain of deepfake detection are also discussed. The analysis provided in this survey reveals that every new deepfake generation technique calls for the development of novel deepfake detection techniques. Among deepfake detection techniques proposed so far, MobileNet-CNN model, multi-scale temporal CNN model, and GAN-based technique outperform good for face swap, lip-syncing, and attribute manipulation detection with an accuracy of 99.28%, 97.1% ,and 100%, respectively. This survey would help researchers to understand the literature on deepfake generation and detection which is required for future development in this field.

GRAPHICAL ABSTRACT

KEYWORDS:

• Face swap
• attribute manipulation
• lip-syncing
• deepfake Datasets

Disclosure statement

No potential conflict of interest was reported by the author(s).