With the rapid advancement of deepfake technology, associated security threats have become increasingly severe, leading to significant interest in deepfake detection techniques. Despite achieving excellent performance on in-dataset evaluations, existing detectors experience substantial performance drops when tested on cross-datasets due to overfitting specific data distributions. To address this issue, we propose a novel Ambiguous Data Generation (ADG) method to generate challenging ambiguous samples that enhance model generalization during training. Specifically, our approach uses maximum mean discrepancy to minimize the distribution gap between natural and synthetic data, ensuring statistical consistency and producing more challenging samples. Additionally, we introduce a Multi-Feature Fusion (MFF) module that combines features before and after perturbation to ensure robust learning and suppress performance degradation in the testing phase. Experimental results across multiple datasets demonstrate that our method significantly improves cross-dataset generalization and overall detection performance.