This paper introduces FreqNet, a frequency-aware deepfake detection method that enhances the generalizability of deepfake detectors. The method focuses on frequency domain learning to improve the detector's ability to identify deepfake images across diverse sources and GAN models. FreqNet incorporates two key modules: high-frequency representation and frequency convolutional layers. The high-frequency representation module forces the detector to focus on high-frequency information, while the frequency convolutional layer enables the detector to learn source-agnostic features in the frequency domain. The method is evaluated on 17 different GAN models, demonstrating state-of-the-art performance with a 9.8% improvement in accuracy compared to existing methods, while requiring significantly fewer parameters. The results show that FreqNet outperforms other frequency-based methods in terms of generalization ability and robustness. The method is implemented using a lightweight CNN classifier and leverages the Fast Fourier Transform (FFT) to extract frequency information. The experiments demonstrate that FreqNet achieves high accuracy on real-world scenes and is effective in detecting deepfakes across various GAN models and categories. The paper also presents ablation studies and visualizations of class activation maps to further validate the effectiveness of FreqNet. Overall, FreqNet provides a novel and efficient approach to deepfake detection that is robust to variations in generation models and image sources.This paper introduces FreqNet, a frequency-aware deepfake detection method that enhances the generalizability of deepfake detectors. The method focuses on frequency domain learning to improve the detector's ability to identify deepfake images across diverse sources and GAN models. FreqNet incorporates two key modules: high-frequency representation and frequency convolutional layers. The high-frequency representation module forces the detector to focus on high-frequency information, while the frequency convolutional layer enables the detector to learn source-agnostic features in the frequency domain. The method is evaluated on 17 different GAN models, demonstrating state-of-the-art performance with a 9.8% improvement in accuracy compared to existing methods, while requiring significantly fewer parameters. The results show that FreqNet outperforms other frequency-based methods in terms of generalization ability and robustness. The method is implemented using a lightweight CNN classifier and leverages the Fast Fourier Transform (FFT) to extract frequency information. The experiments demonstrate that FreqNet achieves high accuracy on real-world scenes and is effective in detecting deepfakes across various GAN models and categories. The paper also presents ablation studies and visualizations of class activation maps to further validate the effectiveness of FreqNet. Overall, FreqNet provides a novel and efficient approach to deepfake detection that is robust to variations in generation models and image sources.