ABSTRACT The contemporary world is fuelled by the social web, with approximately 504.17 h of video content and 1,050,000 photos going up online every minute. This open‐access media is swiftly exploited to facilitate deepfake generation that can disseminate misleading information and tarnish reputations. Furthermore, the subtlety of deepfakes has escalated to the point where deep learning‐based techniques yield deepfakes that are perceptibly flawless and indistinguishable from reality. Moreover, the existing detection models suffer severe affliction when in the crosshairs of adversarial attacks. Therefore, it is paramount to promote the development of resilient evasion strategies against this societal bane, which “takes the bull by the horns” while shielding them from hostile adversaries. In this direction, the presented article outlines the neoteric progression in deepfake topography, assimilating generation, detection, and adversarial attack‐based techniques. It delineates an all‐inclusive taxonomy based on topical research stemming from the deepfake landscape. Based on the taxonomical classification, a systematic and exhaustive review of the avant‐garde deepfake detection techniques is conducted. Furthermore, this article reviews state‐of‐the‐art active defense techniques against audio/video deepfakes, incorporating both generator‐ and detector‐based adversarial attacks, and discusses issues with deepfake datasets. Lastly, the trends in deepfake detection and dataset utilization have been highlighted, bringing to the fore the all‐important spheres requiring futuristic investigation. This article is categorized under: Technologies > Artificial Intelligence Technologies > Machine Learning Commercial, Legal, and Ethical Issues > Security and Privacy
Gupta et al. (Thu,) studied this question.