Deepfake Generation and Detection Algorithms: A Comprehensive Survey of Multimodal Learning, Robustness, Explainability, and Trustworthy Synthetic Media Forensics

Deepfake generation now combines autoencoders, generative adversarial networks, diffusion models, transformers, neural rendering, and audio-conditioned synthesis. This survey presents a structured narrative synthesis of image, video, audio, and multimodal generation and detection, organised around the relationship between synthesis mechanisms and forensic evidence. Literature selection follows explicit source, search-term, eligibility, deduplication, and evidence-appraisal criteria. Generation methods are compared by conditioning requirements, identity preservation, temporal coherence, and computational cost. Detection methods are analysed through spatial, frequency, temporal, physiological, device, audio–visual, and provenance evidence, with attention to the assumptions that make each cue informative. Evaluation definitions distinguish reconstruction quality, distributional realism, identity and speech fidelity, classification performance, calibration, robustness, localisation, and efficiency. Results are interpreted within their dataset and protocol rather than pooled into an architecture-wide ranking. The comparative analysis separates demonstrated benchmark behaviour from proposed directions in open-set detection and multimodal synthesis. Practical recommendations prioritise held-out generators and datasets, realistic processing pipelines, explicit operating thresholds, calibrated confidence, faithful explanations, and provenance validation. Together, these requirements connect algorithmic performance to the evidence needed for trustworthy synthetic-media assessment.