FRLD-DF: frequency ring-guided LoRA adaptation of DINOv2 vision transformer for generalizable deepfake detection

Recent deepfake synthesis techniques produce highly convincing forgeries, increasing the need for detection models that remain effective when applied to previously unseen manipulation domains. Many existing methods exploit only a single type of evidence, such as spatial information, fine spectral patterns, or coarse frequency signals, which limits their ability to handle diverse forgery processes and unconstrained realworld data. In addition, stronger performance is often achieved at the cost of increasing the total number of learnable model parameters. To overcome these challenges, a novel end-to-end framework, FRLD-DF, is proposed for generalized deepfake detection by combining parameter efficient transformer adaptation with structured spectral modeling. The framework adapts a frozen foundation model, DINOv2 Vision Transformer, using Low-Rank Adaptation to preserve robust pretrained semantic representations while reducing overfitting. Moreover, it incorporates a Hierarchical Radial Spectral Decomposition module to capture organized global and local frequency band characteristics across multiple spatial scales. Finally, a frequency spatial fusion attention module is designed to integrate spectral manipulation indicators with transformer semantic features. Extensive generalization experiments across seven evaluation settings demonstrate that FRLD-DF achieves state-of-the-art performance on cross-dataset benchmarks while maintaining a low number of trainable parameters. Under standard crossdataset evaluation, FRLD-DF achieves an average AUC improvement of 2.83%. Furthermore, on a more diverse crossdataset benchmark including the challenging WildDeepfake dataset and the heterogeneous DF40 forgery pipelines, FRLDDF obtains an average AUC improvement of 10.2%.