This paper introduces a novel deepfake detection method using Binary Neural Networks (BNNs) for efficient real-time inference with minimal accuracy loss. The approach incorporates Fast Fourier Transform (FFT) and Local Binary Pattern (LBP) as additional channel features to detect manipulation traces in frequency and texture domains. The method is evaluated on COCOFake, DFFD, and CIFAAE datasets, demonstrating state-of-the-art performance with a significant efficiency gain of up to 20× reduction in FLOPs during inference. The proposed method uses BNext as a backbone, which is pre-trained on ImageNet, and augments input images with FFT magnitude and LBP channels to enhance detection capabilities. The model is trained on three benchmark datasets, achieving competitive results while reducing computational consumption. The method also includes an ablation study to highlight the impact of design choices. The results show that the BNN-based approach can match the performance of full-precision models with significantly lower computational requirements. The study highlights the potential of BNNs in deepfake detection, emphasizing the need for efficient methods to combat the spread of deepfake media. The research contributes to the field by proposing the first BNN-based deepfake detection method, demonstrating the effectiveness of BNNs in identifying generated images, and providing quantitative results to support further investigation. The code is available at https://github.com/fedeIoper/binary_deepfake_detection.This paper introduces a novel deepfake detection method using Binary Neural Networks (BNNs) for efficient real-time inference with minimal accuracy loss. The approach incorporates Fast Fourier Transform (FFT) and Local Binary Pattern (LBP) as additional channel features to detect manipulation traces in frequency and texture domains. The method is evaluated on COCOFake, DFFD, and CIFAAE datasets, demonstrating state-of-the-art performance with a significant efficiency gain of up to 20× reduction in FLOPs during inference. The proposed method uses BNext as a backbone, which is pre-trained on ImageNet, and augments input images with FFT magnitude and LBP channels to enhance detection capabilities. The model is trained on three benchmark datasets, achieving competitive results while reducing computational consumption. The method also includes an ablation study to highlight the impact of design choices. The results show that the BNN-based approach can match the performance of full-precision models with significantly lower computational requirements. The study highlights the potential of BNNs in deepfake detection, emphasizing the need for efficient methods to combat the spread of deepfake media. The research contributes to the field by proposing the first BNN-based deepfake detection method, demonstrating the effectiveness of BNNs in identifying generated images, and providing quantitative results to support further investigation. The code is available at https://github.com/fedeIoper/binary_deepfake_detection.