Publications

The emerging threat of deepfake technology allows creation of highly realistic synthetic media with major implications for digital communication security, enabling manipulation for misinformation, identity theft, and fraud. This paper provides a comprehensive and comparative overview of the latest algorithms proposed for detecting deepfakes, categorized into four types: deep-learning based, biological signals, frequency domain, and hybrid algorithms. We evaluate performance on benchmark datasets (FaceForensics++, DFDC, Celeb-DF) using metrics like AUC-ROC, computational complexity, and generalization ability. Deep-learning based methods perform well on individual datasets but degrade with cross-dataset generalization or adversarial examples. Attention-based methods show improved robustness, with DSP-FWA achieving AUC of 0.930 on FaceForensics++. This paper discusses security considerations and suggests a framework for protective structures in digital communication.