A Review of AI-Generated Content Detection Technologies: From DeepFake Recognition to Multimodal Provenance

Abstract

Generative adversarial networks (GANs), diffusion models, and multimodal synthesis pipelines have reached a level where machine-produced images, audio-visual clips, and cross-platform narratives can closely mimic authentic content. Detection strategies built around isolated visual artifacts are, as a consequence, losing much of their former effectiveness. We survey representative work from IEEE Xplore, Web of Science, and CNKI covering the period 2018–2025, drawing on 38 references—28 core studies plus 10 supplementary sources on benchmarks, methodological norms, and thematic surveys—to build an analytical framework organized by detection target, forensic feature, model architecture, and deployment context. A clear trajectory emerges from the literature: early detectors that relied on local texture descriptors and frequency-domain cues have gradually given way to more integrated forensic approaches, including real-face representation learning, cross-dataset robustness enhancement, audio-visual consistency verification, and generation-source attribution. While multimodal fusion and open-set provenance analysis broaden the reach of detection to complex real-world settings, practical obstacles—platform compression, adversarial perturbation, real-time throughput limits, and the disconnect between algorithmic outputs and governance workflows—are far from resolved. Building on these findings, we map current technical routes, trace their evolutionary logic, and outline considerations for detection system design in social media governance contexts.