From curated datasets to social-media reality: In-the-wild, multimodal, and generalizable deepfake detection

Deepfake detection has advanced rapidly alongside progress in generative adversarial networks, diffusion models, neural rendering, and multimodal synthesis. However, strong performance on curated benchmark datasets does not necessarily translate into reliable performance in real-world environments, where manipulated media are affected by platform compression, reposting, post-processing, multimodal inconsistency, temporal drift, and evolving generation pipelines. This review examines deepfake detection from the perspective of in-the-wild robustness, multimodal reasoning, and cross-domain generalization. It first clarifies the conceptual foundations of in-the-wild and generalizable deepfake detection, then reviews the data ecosystem, including benchmark datasets, real-world media conditions, multimodal data gaps, annotation quality, and benchmark aging. The review further analyzes the major causes of detector failure under realistic deployment conditions, including generator shift, platform-induced degradation, modality mismatch, language and speaker variation, and broader contextual diversity. It then compares major methodological families, including artifact-based methods, deep representation-based approaches, transfer-based learning, domain adaptation, domain generalization, and multimodal detection strategies, with emphasis on their relevance to real-world robustness. Finally, the review discusses evaluation practices, external validation, realistic benchmarking, and future directions such as continually updated benchmarks, missing-modality-robust systems, uncertainty-aware detection, provenance integration, and foundation-model-based approaches. Overall, the review argues that reliable deepfake detection requires a shift from benchmark-centered optimization toward robust, multimodal, and deployment-oriented evaluation and design.