Abstract: The remarkable advancements in generative artificial intelligence have ushered in deepfake technology. This technology enables high-fidelity manipulation and synthesis of multimedia content with unprecedented ease, thereby posing significant threats to digital society security. In this regard, this study systematically identifies four core challenges hindering deepfake detection: insufficient robustness in complex scenarios, limited generalization capability for novel forgery techniques, scarcity of standardized databases and evaluation benchmarks, and the inherent trade-off between computational efficiency and detection accuracy. In response to these challenges, this paper proposes a multifaceted technical framework, including frequency-domain decoupled dynamic feature extraction, multimodal contrastive learning, adversarial robustness augmentation, and lightweight model deployment. Lastly, this paper explores the establishment of a synergistic governance paradigm that facilitates the coordinated evolution of legal norms and detection technologies.