The survey tracks the evolution of deepfake generation from early autoencoders and GANs to diffusion models and autoregressive transformers, noting steady gains in realism and temporal coherence.

It frames broad societal risks—non-consensual imagery, financial fraud, political misinformation, and disruptions to legal processes—calling for platform governance and technical defenses working in tandem.

A hybrid defense strategy is urged, combining artifact, temporal, biological, and cross-modal cues, with continual benchmarking against unseen generators to stay ahead.

Generalization gaps across datasets, sensitivity to compression, and deliberate adversarial evasion highlight an ongoing arms race between forgery methods and detectors.

The authors push for explainable detection, robustness to adversarial perturbations, privacy-preserving federated learning, efficiency-focused neural architecture search, and provenance standards to boost digital trust.

The article serves as a comprehensive reference linking generation paradigms to forensic signatures and outlines a path toward trustworthy, generalizable deepfake detection systems.

Early detection relied on low-level artifacts and biometric signals, while modern methods deploy deep multimodal learning, Vision Transformers, cross-modal analysis, and temporal coherence checks for robustness.

The field has shifted from static images to video and full-body reenactment, with milestones like Face2Face and 3D-aware head synthesis enabling fidelity across poses and angles.

Diffusion models now dominate image and video synthesis, with extensions such as stable video diffusion, AnimateDiff, and Lumiere enabling text- and reference-driven generation with coherent motion.

Benchmarking relies on datasets like FaceForensics++, Deepfake Detection Challenge, Celeb-DF, and AV-Deepfake1M, though cross-dataset generalization remains a persistent weakness.