Frequency-Aware Deepfake Detection: Improving Generalizability through Frequency Space Learning

There has been a significant effort in the field of forgery detection, with many studies focusing on leveraging spatial information from images. Rossler et al.(Rossler et al. 2019) utilize images to train the Xception (Chollet 2017) to achieve fake face image detection. Other image-based detection methods have developed specific artifact detection in distinct facial regions, such as eyes(Li et al. 2018) and lips (Haliassos et al. 2021). Chai et al.(Chai et al. 2020) adopt limited receptive fields to identify patches that render images detectable, highlighting the importance of specific local features. With the emergence of deepfake technology, efforts are being made to enhance the generalization ability of detectors, particularly for unseen data. Yu et al.(Yu et al. 2020) introduce artifacts from the camera imaging process. Furthermore, various methods (Wang et al. 2020, 2021; Chen et al. 2022; Cao et al. 2022; He et al. 2021; Shiohara et al. 2022) aim to enrich the diversity of training data, employing techniques such as data augmentation, adversarial training, reconstruction, and blending images. CDDB (Li et al. 2023) adopts incremental learning to achieve continual deepfake detection. Notably, recent works by Ojha et al.(Ojha et al. 2023) and Tan et al.(Tan et al. 2023) employ the feature map and gradients as the general representation, respectively.

The research by (Frank et al. 2020; Durall et al. 2020), which highlights the effectiveness of frequency artifacts in the domain of deepfake detection, has significantly influenced subsequent studies. These findings have led many image forgery detectors to shift their attention toward capturing unique patterns within the frequency domain. The work by Masi et al. (Masi et al. 2020) stands out as it meticulously investigates artifacts present in both the color space and the frequency domains, while $F^{3}$-Net (Qian et al. 2020) suggests using the discrepancy of frequency statistics between real and forged images as a means to differentiate face image manipulations. An adaptive frequency features learning is designed by FDFL (Li et al. 2021) to mine subtle artifacts from the frequency domain, enabling the detection of forged images. In parallel, Luo et al.(Luo et al. 2021) improve the generalization performance through the integration of multiple high-frequency features, thereby fortifying the robustness of the forgery detection process. Moreover, ADD (Woo et al. 2022) incorporates two meticulously designed distillation modules to emphasize the significance of frequency information, incorporating frequency attention distillation and multi-view attention distillation. Recently, frequency-based studies have been proposed for generalized detection. BiHPF (Jeong et al. 2022a) emphasizes amplifying artifact magnitudes through the utilization of dual high-pass filters. The FreGAN model, introduced by (Jeong et al. 2022c) ingeniously mitigates the impact of frequency-level artifacts through the deployment of frequency-level perturbation maps. (Wang et al. 2023) introduces dynamic graph learning to exploit the relation-aware features in spatial and frequency domains.

Our primary focus lies within the realm of Generalizable Deepfake Detection. Our objective is to construct a universal detector capable of accurately identifying deepfake images even when confronted with constrained training sources. In this context, let us consider a real-world image scenario denoted as $X$ sampled from $n$ different sources:

$$\begin{split}&X=\{X_{1},X_{2},...,X_{i},...,X_{n}\},\\ &X_{i}=\{x_{j}^{i},y_{j}\}_{j=1}^{N_{i}},\end{split}$$

(1)

where $N_{i}$ represents the number of images originating from the $i$th source $X_{i}$, $x_{j}^{i}$ is the $j$th image of $X_{i}$. Each image is labeled with $y$, indicating whether it belongs to the category of ”real”$(y=0)$ or ”fake” $(y=1)$. Here we train a binary classifier $D(\cdot)$, utilizing the training source $X_{i}$:

$$D^{{i}}=\mathop{\arg\min}_{\theta}\ l(D(X_{i};\theta),\ y),$$

(2)

where $l()$ denote the loss function. Our overarching goal is to design a detector that is trained on the data originating from $X_{i}$, but demonstrates strong performance when confronted with images coming from previously unseen sources, denoted as $X_{t}$. This generalizability across unseen sources is a crucial objective of our detector.

With the primary objective of bolstering generalizability to unseen sources, we have devised a frequency domain learning network to effectively enhance deepfake detection capabilities. The comprehensive architecture of the FreqNet approach is depicted in Figure 3. Within our methodology, we introduce practical and compact frequency learning plugin modules designed to compel the CNN classifier to operate within the frequency domain. These modules include the high-frequency representation and the frequency convolutional layer. The modular nature of these components allows seamless integration into the CNN classifier. This architectural innovation has culminated in the creation of a novel and lightweight detector named FreqNet, incorporating the frequency learning plugins alongside a limited number of CNN layers.

We design a lightweight CNN classifier, employing residual convolutional blocks without pretraining. During the training process, we utilize the Adam optimizer (Kingma et al. 2015) with an initial learning rate of $2\times 10^{-2}$. The batch size is set at 32, and we train the model for 100 epochs. A learning rate decay strategy is employed, reducing the learning rate by twenty percent after every ten epochs. Consistent with established baselines (Jeong et al. 2022a, c), we utilize the average precision score (A.P.) and accuracy (Acc.) as the primary evaluation metrics to gauge the effectiveness of our proposed method. These metrics provide a comprehensive assessment of the performance of our approach against the baselines. We employ the PyTorch framework (Paszke et al. 2019) for the implementation of our method, utilizing the computational power of the Nvidia GeForce RTX 3090 GPU. For the critical task of Fast Fourier Transform (FFT), we leverage the $torch.fft.fftn$ function within the PyTorch library.

In order to demonstrate the remarkable generalization ability of our FreqNet on unseen sources, we carry out evaluations on a real-world scene dataset. This dataset comprises images sourced from a total of 17 different generation models, encompassing 8 models from the ForenSynths test set and an additional 9 models from our own self-synthesis process. This evaluation setup introduces increased complexity compared to the previous experiments, as the testing sets encompass a diverse range of GANs. This challenging scenario effectively simulates an open-world scene, making it a rigorous test of our approach’s adaptability and robustness in real-world, unpredictable settings.

We compare with the previous methods: BiHPF (Jeong et al. 2022a), FreGAN (Jeong et al. 2022c), LGrad (Tan et al. 2023), Ojha (Ojha et al. 2023). To ensure fair and meaningful comparisons, we adopt the same experimental setting as the established baselines (Jeong et al. 2022a, c). Specifically, the 1-class, 2-class and 4-class settings refer to training with the (horse), (chair, horse), (car, cat, chair, horse) categories of ProGAN, respectively.

The results of the ForenSynths dataset are presented in Table 1. The proposed FreqNet surpasses its counterparts in terms of mean Acc. metric and mean A.P. metrics, except for the value of A.P. on the 1-class setting. Notably, with the 4-class setting, FreqNet achieves a mean Acc. value of 91.5%, demonstrating its strong performance. In comparison to the current state-of-the-art methods LGrad and Ojha, our FreqNet exhibits substantial improvements, surpassing these methods by 5.4% and 2.4% in mean Acc., using fewer parameters. In the 1- and 2-class settings, our FreqNet also achieves gains of 2.9% and 0.4% compared to Ojha. Furthermore, compared to FingerprintNet(Jeong et al. 2022b) tested on six unseen models, our FreqNet achieves a marked improvement in mean accuracy, rising from 82.6% to 90.6% and exhibiting a significant gain of 8.0%. Additionally, we provide results on 9 models from self-synthesis in Table 2. We adopt the 4-class setting detector to perform testing. Compared to Ojha, our FreqNet achieves a marked improvement in mean accuracy, soaring from 77.6% to an impressive 94.0%, resulting in a significant gain of 16.4%. When testing on face images, the proposed FreqNet achieves accuracy rates of 98.7%, 99.0%, and 99.5% on the ProGAN, StyleGAN, and StyleGAN2 datasets, respectively.

Furthermore, we provide a comprehensive overview of the number of parameters and the mean accuracy across all real-world scenes in Table 3. It is evident that our FreqNet, with a modest parameter count of 1.9 million, significantly outperforms the current state-of-the-art model Ojha (Ojha et al. 2023), which boasts an extensive 304 million parameters. This substantial difference in parameter count translates into a notable performance gain, with our FreqNet achieving a remarkable improvement of 9.8% in mean accuracy compared to the larger model. This result underscores the efficiency and effectiveness of our FreqNet approach, demonstrating that superior performance can be achieved with significantly fewer parameters, a crucial advantage in real-world applications.

Compared to other frequency-based methods, such as BiHPF (Jeong et al. 2022a), FrePGAN (Jeong et al. 2022c), F3Net(Qian et al. 2020), our FreqNet achieves better performance on the real-world scene. The results confirm the generalization capability of the proposed frequency domain learning to extract a general representation of artifacts, and generalize this representation across various GAN models and categories.

We perform ablation analyses on our FreqNet by individually removing the proposed modules. The results of these ablation experiments are presented in Table 4. Upon removal of the designed modules, we observe a decline in the detection performance, underscoring the efficacy of the proposed components. In the revised version, we will expound further on the specifics of the ablation analysis to provide a more comprehensive understanding.

Visualization of Class Activate Map. To visually demonstrate the discriminative regions identified by our detector, we present the Class Activation Maps (CAM) in Figure 4. The CAMs are generated using images from ProGAN, StyleGAN, StyleGAN2, and CelebA-HQ. The CAMs provide insights into the areas of focus for the detector in distinguishing real from fake images. It’s noteworthy that the CAMs for real images highlight a broader portion of the image, while the CAMs for fake images tend to emphasize localized regions. Interestingly, even though the detector is primarily trained using a dataset containing cars, cats, chairs, and horses, it showcases the ability to recognize face images effectively. This highlights the versatility and adaptability of our detector in identifying distinct deepfake characteristics, even beyond the classes it was primarily trained on.