Deepfake Detection Beyond Benchmark Accuracy: Generalization, Robustness, and Reproducibility

Deepfake detection has become a central problem in computer vision and multimedia security as generative systems make synthetic images, video, and audio increasingly accessible. This systematic literature review presents a large-scale, per-paper full-text analysis of 1,410 papers published from January 2018 through the January 2026 cutoff. Across 11 research questions, we examine methods, benchmark resources, generator-detector adaptation, robustness, code evidence, evaluation protocols, metrics, adversarial research, terminology, open challenges, and research-community structure. The results reveal rapid architectural progress but weak comparability: 1,024 papers (72.6%) report no robustness testing beyond their training distribution; 798 of 1,091 papers with evaluation metadata (73.1%) omit the train-test split strategy; and 667 papers (47.3%) explicitly claim state-of-the-art performance while protocols remain heterogeneous. CNNs are the largest single method category (266 papers), followed by transformers (207) and foundation-model approaches (121); the year-annotated series contains 28 foundation-model papers in 2024 and 50 in 2025. FaceForensics++ remains the dominant benchmark family after alias normalization (330 paper-family mentions), but the resource audit shows that benchmarks must be separated from source/pretraining datasets and generator models or services. Code evidence is similarly tiered: 721 papers contain a code/model link, 334 satisfy a conservative single-GitHub proxy for likely self-released implementation code, and 15 pass the strict README/title-match audit available in the pipeline. Generalization is the most frequently coded open challenge (436 mentions). Adversarial coding yields 137 attack-oriented and 390 defense-oriented paper-label assignments, which may overlap; adversarial training appears in 11 defense papers (2.8%). The central conclusion is that deepfake detection is no longer constrained primarily by model capacity: the binding constraints are evaluation validity, cross-generator generalization, and reproducible evidence.