Multimedia Deepfake Threats and Detection Strategies Based on Limited Data: An Exhaustive Survey

Resumen

The high-speed development of multimedia deepfake technology has fundamentally altered the digital creativity and, at the same time, increased cyber and social threats. The current state of deepfake generation is informed by recent artificial intelligence methods, such as Generative Adversarial Networks (GANs), which learn to create realistically visual and audio content via adversarial learning; autoencoder-based models, which allow manipulating identity and expression compared to images; and diffusion models, which are trained to produce high-fidelity media by refining noise to data. Even though these practices aid in innovative uses in entertainment, availability, and customized media, they also help to perpetrate malpractices like political and social manipulation, privacy invasion, identity theft, misinformation, and cybersecurity assaults. Due to the recent popularity of deepfakes, there has been rapid growth in the design of a wide range of media detection models for video and audio-based deepfakes. In this study, literature on state-of-the-art models for deep-fake detection has been compiled and summarized, which includes CNN-based Spatio-Temporal models, CNN and LSTM hybrid models, Capsule Networks, and Sharp Multiple Instance Learning models (S-MIL), which were tested on popular deep-fake databases like FaceForensics++, Face2Face, DFDC, and ASV-Spoof. Results from the study indicate a high level of efficacy for the models in crossDataset environments, with accuracy between 84.6% and 100%, precision up to 98%, and AUC reaching (1.0). Limitations still exist for deep-fake detection models relate to high computational complexity, sensitivity to attacks, and the lack of generalizability for real-time processing.