Deepfake image detection method and system based on frequency domain transform

Abstract

This disclosure provides a deepfake image detection method based on frequency domain transformation, comprising: receiving an image to be detected; extracting local spatial domain features and local frequency domain features of the image to be detected; filtering local frequency domain features using channel attention; fusing local spatial domain features and the filtered local frequency domain features to obtain fused features of the image to be detected; and determining the authenticity of the image to be detected based on the fused features.

Description

TECHNICAL FIELD

[0001] The present disclosure mainly relates to image detection, and particularly relates to detection of deepfake images. BACKGROUND

[0002] In recent years, deepfake technology related to images and videos has rapidly emerged, vividly illustrating that seeing is not necessarily believing, and bringing corresponding risks and challenges to individuals and society, thus it is urgent to detect deepfakes on the network.

[0003] Deep learning has strong performance in the field of image processing, so that automatic deepfake detection can be realized through model training, however, only detecting and identifying the RGB features of images, as the fake samples become more realistic, the detection performance will be severely reduced.

[0004] Therefore, there is a need for an efficient deepfake image detection scheme in the field. SUMMARY

[0005] To solve the above technical problems, the present disclosure provides a deepfake image detection scheme based on frequency domain conversion, which simultaneously analyzes the spatial domain features and frequency domain features of images, quickly finds useful local tampering information by introducing channel attention to filter important frequency domain features, and achieves high-performance and high-robust deepfake image detection.

[0006] In an embodiment of the present disclosure, a deepfake image detection method based on frequency domain conversion is provided, comprising: receiving a to-be-detected picture; extracting local spatial domain features and local frequency domain features of the to-be-detected picture; filtering the local frequency domain features by using channel attention; fusing the local spatial domain features and the filtered local frequency domain features to obtain fusion features of the to-be-detected picture; and judging the authenticity of the to-be-detected picture based on the fusion features.

[0007] In another embodiment of the present disclosure, the extraction of the local spatial domain features of the to-be-detected picture is performed by using a convolutional neural network with a small number of layer convolutions.

[0008] In yet another embodiment of the present disclosure, the extraction of the local frequency domain features of the to-be-detected picture comprises frequency domain conversion on the to-be-detected picture.

[0009] In another embodiment of the present disclosure, the filtering of the local frequency domain features by using channel attention is used to filter important local frequency domain features.

[0010] In yet another embodiment of the present disclosure, the filtering of the local frequency domain features by using channel attention is performed by SE-Net.

[0011] In another embodiment of this disclosure, the selected local frequency domain features may be further refined before being fused with the local spatial domain features.

[0012] In yet another embodiment of this disclosure, further refinement of the selected local frequency domain features includes recursively analyzing the selected local frequency domain features and extracting finer-grained features from the selected local frequency domain features.

[0013] In another embodiment of this disclosure, the model used in the method is trained using a focus loss function.

[0014] In one embodiment of this disclosure, a deepfake image detection system based on frequency domain transformation is provided, comprising: a receiving module for receiving an image to be detected; an extraction module for extracting local spatial features and local frequency domain features of the image to be detected; a filtering module for filtering the local frequency domain features using channel attention; a fusion module for fusing the local spatial features and the filtered local frequency domain features to obtain fused features of the image to be detected; and a judgment module for judging the authenticity of the image to be detected based on the fused features.

[0015] In yet another embodiment of this disclosure, the extraction module extracts the local spatial features of the image to be detected using a convolutional neural network with a small number of convolutional layers.

[0016] In another embodiment of this disclosure, the extraction module extracts local frequency domain features of the image to be detected by performing frequency domain transformation on the image to be detected.

[0017] In one embodiment of this disclosure, a computer-readable storage medium storing instructions that, when executed, cause a machine to perform the aforementioned method.

[0018] This overview is provided to introduce, in a simplified form, some of the concepts further described in the detailed description below. This overview is not intended to identify key or essential features of the claimed subject matter, nor is it intended to limit the scope of the claimed subject matter. Attached Figure Description

[0019] The above-described invention and the following detailed description will be better understood when read in conjunction with the accompanying drawings. It should be noted that the drawings are merely examples of the claimed invention. In the drawings, the same reference numerals represent the same or similar elements.

[0020] Figure 1 This is a flowchart illustrating a depth-spoofing image detection method based on frequency domain transformation according to an embodiment of the present disclosure.

[0021] Figure 2This is a schematic diagram illustrating a deep forgery image detection process based on frequency domain transformation according to an embodiment of the present disclosure.

[0022] Figure 3 This is a schematic diagram illustrating a spatial feature extraction framework according to an embodiment of the present disclosure.

[0023] Figure 4 This is a schematic diagram illustrating a frequency domain feature extraction framework according to an embodiment of the present disclosure.

[0024] Figure 5 This is a schematic diagram illustrating a fusion classification framework according to an embodiment of the present disclosure.

[0025] Figure 6 This is a block diagram illustrating a deepfake image detection system based on frequency domain transformation according to an embodiment of the present disclosure. Detailed Implementation

[0026] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, the specific embodiments of this disclosure will be described in detail below with reference to the accompanying drawings.

[0027] Numerous specific details are set forth in the following description in order to provide a full understanding of this disclosure. However, this disclosure may be practiced in other ways different from those described herein, and therefore is not limited to the specific embodiments disclosed below.

[0028] In the detection of forged images, the application of deep learning differs significantly from conventional computer vision tasks. First, the target of image authenticity detection is not the semantic content region, but rather the tampered region. Whether it's object detection or semantic segmentation, deep learning models identify the semantic content of an image, such as people, animals, or vehicles. However, image authenticity detection requires models to identify tampered regions. In different scenarios, the tampered region might be a semantically defined target, an irregular region, or even a region that has been removed.

[0029] Secondly, the features differ. Conventional computer vision tasks focus on the semantic content of images, without needing to pay attention to subtle changes between image contents. However, image authenticity detection tasks are the opposite. Image authenticity detection tasks need to focus on subtle changes in the tampering boundaries, judging the authenticity of an image based on changes in boundary artifacts and statistical features.

[0030] Secondly, image post-processing has different effects on recognition accuracy. In conventional computer vision tasks, normal image compression does not significantly affect the accuracy of the model. However, in image authenticity detection tasks, since passive image evidence collection requires attention to subtle feature changes at the boundaries, image post-processing techniques (such as blurring and compression) greatly damage the clues of image tampering. After processing, the subtle changes at the boundaries of the image will be greatly reduced or even disappeared.

[0031] In recent years, deep learning technology has developed rapidly. Among its applications, Generative Adversarial Networks (GANs) generate images that are superimposed onto the original image or video, and then stitch together personal voices, facial expressions, and body movements to create fake content that is difficult to distinguish from reality. This is because the GAN framework generates images by establishing a connection between missing regions and real regions, thereby predicting the missing regions based on the current region's feature distribution. Consequently, the altered regions and real regions often exhibit striking similarities in statistical features and edge artifacts. Therefore, extracting effective features becomes very difficult.

[0032] Furthermore, the GAN framework consists of a generator and a discriminator, both of which are deep learning networks. The generator is responsible for generating images that are difficult to distinguish from real ones, while the discriminator is responsible for identifying the images. When the discriminator cannot identify a generated image, that image is output. Therefore, learning tampering clues that are undetectable to the human eye is also challenging.

[0033] Therefore, there is a need in this field for a solution that can efficiently detect the authenticity of deepfake images.

[0034] This disclosure focuses on detecting tampered regions in images, rather than semantic content regions. It extracts local spatial features from the image rather than global spatial features. Based on the fact that GAN-generated images often exhibit artifacts in the high-frequency range, local frequency domain features are extracted and filtered. These extracted and filtered local frequency domain features are then fused to detect image authenticity. This significantly reduces the amount of data involved, thus effectively improving detection efficiency. Furthermore, the more refined acquisition of local tampering information enhances robustness.

[0035] Therefore, this disclosure provides a deepfake image detection scheme based on frequency domain transformation, which simultaneously incorporates the analysis of spatial and frequency domain features of the image. By introducing channel attention to filter out important frequency domain features, useful local tampering information can be quickly found, thereby achieving high-performance and highly robust deepfake image detection.

[0036] Figure 1 This is a flowchart illustrating a depth forgery image detection method 100 based on frequency domain transformation according to an embodiment of the present disclosure.

[0037] At 102, receive the image to be detected.

[0038] In different scenarios, the image to be detected can be a static image, a dynamic image, a short video, or a video.

[0039] In step 106, local spatial and frequency domain features of the image to be detected are extracted.

[0040] As mentioned earlier, the target of image authenticity detection is not the semantic content region, but the tampered region. Therefore, extracting the local spatial features of the image to be detected can extract the features of the tampered region in the image. To avoid deep convolutional networks (e.g., CNN, RNN, etc.) extracting deep semantic information of the image, only a few layers of convolution can be used to extract local spatial features.

[0041] In one embodiment of this disclosure, an image may be segmented into non-overlapping image blocks of the same size, and these image blocks may then be used as input to a convolutional network.

[0042] In another embodiment of this disclosure, a sliding window can be used to extract local features of an image in order to achieve a more refined acquisition of local features.

[0043] The following will refer to Figure 3 This invention provides a detailed description of a spatial feature extraction process according to an embodiment of the present disclosure.

[0044] As mentioned earlier, it is difficult to distinguish between real and fake images generated by GANs, especially in terms of spatial features. However, since the frequency domain features of GAN-generated images contain artifacts (e.g., rasterization artifacts), extracting local frequency domain features is of positive significance. The image to be detected can be converted to the frequency domain, and local frequency domain features at different frequencies / channels can be generated.

[0045] In 108, channel attention is used to filter local frequency domain features.

[0046] Shifting attention to the most important parts is called the attention mechanism. In visual systems, the attention mechanism can be viewed as a dynamic selection process, achieved by adaptively weighting features according to the importance of the input. Channel-attention networks, including SENet, CBAM, and ECANet, can adaptively predict potential key features.

[0047] In this disclosure, a channel attention mechanism is introduced to filter out local frequency domain features in important channels in order to reduce inference time. In one embodiment of this disclosure, SENet is used for filtering. Those skilled in the art will understand that different channel attention mechanisms / networks can be used to filter local frequency domain features depending on different scenarios.

[0048] Furthermore, the selected local frequency domain features can be refined. In one embodiment of this disclosure, the selected local frequency domain features are recursively analyzed, and finer-grained features are extracted from the obtained local frequency domain features.

[0049] The following will refer to Figure 4 The frequency domain feature extraction process according to an embodiment of the present disclosure is described in detail.

[0050] In step 110, local spatial domain features and selected local frequency domain features are fused to obtain the fused features of the image to be detected.

[0051] Depending on the scenario, the fusion of local spatial features and selected (or refined) local frequency domain features can be achieved using concat, add operations, or pyramid fusion. In one embodiment of this disclosure, the fusion of local spatial features and refined local frequency domain features can be further performed using an attention mechanism to automatically learn the importance of different features, including spatial and frequency domain features.

[0052] In step 112, the authenticity of the image to be detected is determined based on the fusion features.

[0053] Determining the authenticity of an image based on fused features can be done using classification networks (e.g., ResNet, Inception).

[0054] The following will refer to Figure 5 A detailed description of the fusion classification process according to an embodiment of this disclosure.

[0055] Therefore, the deepfake image detection method based on frequency domain transformation disclosed in this paper incorporates the analysis of both spatial and frequency domain features of the image. By introducing channel attention to filter out important frequency domain features, useful local tampering information can be quickly found, thereby achieving high-performance and high-robustness deepfake image detection.

[0056] Figure 2 This is a schematic diagram illustrating a deep forgery image detection process based on frequency domain transformation according to an embodiment of the present disclosure.

[0057] like Figure 2As shown in the upper right figure, the spatial / temporal domain (hereinafter collectively referred to as the spatial domain) and the frequency domain offer different perspectives in image processing. In the spatial domain, the function's independent variable (x,y) is considered a point in two-dimensional space, and the digital image f(x,y) is a discrete function defined on a rectangular region in two-dimensional space. In other words, if f(x,y) is considered as a two-dimensional signal with varying amplitude, the image can be processed in the frequency domain using various transformation techniques (e.g., Fourier transform, Discrete Cosine Transform (DCT), Walsh transform, and wavelet transform). Some characteristics are more prominent and easier to process in the frequency domain. For example, noise patterns are difficult to identify in spatial images, but are easier to find and process in the frequency domain.

[0058] Similarly, as Figure 2 As shown in the figure below, in deepfake image detection, since GAN-generated images are almost indistinguishable from real images in terms of spatial features, we consider introducing frequency domain features through frequency domain transformation. Based on the possible frequency domain artifacts of GAN-generated images, we use a channel attention mechanism to filter the local frequency domain features of multiple channels of the image, and then fuse the filtered frequency domain features with the local spatial domain features before classifying them as real or fake.

[0059] Therefore, in this disclosure as Figure 2 In the illustrated embodiment, when faced with deepfake images, local tampering features can be extracted from both the spatial and frequency domains, making it impossible for artificially forged images, machine learning-generated forged images, or a combination of both to escape detection.

[0060] Furthermore, by introducing a channel attention mechanism, the number of features that the model needs to compute is reduced, thereby improving the efficiency and robustness of the frequency domain transform-based deepfake image detection process disclosed in this paper.

[0061] Figure 3 This is a schematic diagram illustrating a spatial feature extraction framework according to an embodiment of the present disclosure.

[0062] In such Figure 3 In the spatial feature extraction framework shown, the image to be detected is first preprocessed.

[0063] In one embodiment of this disclosure, preprocessing includes segmenting the image into non-overlapping image patches of the same size, and then using these image patches as input to a convolutional network.

[0064] In another embodiment of this disclosure, the preprocessing includes determining the size of a sliding window to prepare for extracting local features of an image by overlapping sliding windows of that size, thereby achieving a more refined acquisition of local features.

[0065] like Figure 3As shown, the spatial feature extraction framework according to an embodiment of the present disclosure is mainly composed of a convolutional neural network containing an appropriate number of convolutional layers.

[0066] Convolutional neural networks (CNNs) typically consist of multiple convolutional layers, each containing multiple convolutional kernels. These kernels scan the entire image sequentially from left to right and top to bottom, producing output data called feature maps. Early convolutional layers capture local, detailed information about the image and have small receptive fields; each pixel in the output image utilizes only a small area of ​​the input image. Subsequent convolutional layers progressively increase their receptive fields, capturing more complex and abstract information. After operations across multiple convolutional layers, an abstract representation of the image at various scales is obtained. In other words, in CNNs, generally speaking, shallower networks focus more on detailed information, while deeper networks focus more on semantic information.

[0067] As mentioned earlier, in terms of spatial features, the target of image authenticity detection is not the semantic content region, but the tampered region; the focus should be on subtle changes in the tampering boundaries. To avoid deep convolutional networks (e.g., CNN, RNN, etc.) extracting deep semantic information of images, only a few layers of convolution can be used to extract local spatial features, because extracting the local spatial features of the image to be detected is sufficient to extract the tampered region features in the image.

[0068] like Figure 3 As shown, in one embodiment of this disclosure, shallow convolution and mid-level convolution are used to extract local spatial features of an image. The convolutional neural network includes convolution 1 + activation, pooling 1, convolution 2 + activation, pooling 2, and a fully connected layer. Those skilled in the art will understand that this few-layer convolutional neural network can have different appropriate numbers of convolutional layers depending on different application scenarios.

[0069] Therefore, by not using deep convolutional networks to extract semantic information from images, but only using a suitable number of convolutional layers to extract detailed forgery features, the amount of data to be processed is reduced, and the efficiency of the spatial feature extraction framework is improved.

[0070] Figure 4 This is a schematic diagram illustrating a frequency domain feature extraction framework according to an embodiment of the present disclosure.

[0071] In such Figure 4 In the frequency domain feature extraction framework shown, the image to be detected is first preprocessed. This preprocessing includes resizing, cropping, and flipping the original image, and converting the original RGB image to YCbCr format.

[0072] In YCbCr format, Y represents luminance, Cb represents blue chromaticity, and Cr represents red chromaticity. YCbCr format typically uses a 4:1:1 ratio, meaning each pixel stores an 8-bit Y value, and every 2*2 pixels store an 8-bit Cr and Cb value. Thus, what would normally require 8*3=24 bits per pixel in an RGB image becomes only 8 bits + 8bit / 4 + 8bit / 4=12 bits per pixel in YCbCr. This halves the image size.

[0073] The preprocessed image is then subjected to frequency domain transformation (e.g., Discrete Fourier Transform (DFT), Discrete Cosine Transform (DCT), Discrete Wavelet Transform (DWT), etc.), and the frequency domain transformation coefficients at the same frequency are assigned to the same channel. This generates multiple channels. In deepfake images, some channels have a greater impact on image authenticity detection than others; therefore, this disclosure retains only the important channels, requiring less communication bandwidth while achieving higher accuracy.

[0074] Retaining only important channels is achieved by introducing channel attention mechanisms (e.g., SENet, CBAM, and ECANet). Figure 4 As shown, SENet is used, where multi-channel frequency domain features are input. Given an input x, the number of feature channels is c'. After a series of convolutions and other transformations, features with c feature channels are obtained. Then, the previously obtained features are recalibrated: first, the features are squeezed (W x H x C) → 1 x 1 x C, and then excitation is used to generate weights w for each feature channel to explicitly model the correlation between feature channels. Then, the weights w output by the excitation are used to weight each channel, thereby achieving recalibration (scaling) of the original channels in the channel dimension.

[0075] like Figure 4 The frequency domain feature extraction framework shown improves accuracy by modeling the correlation between feature channels and enhancing the frequency domain features of important channels. Furthermore, the frequency domain features of important channels can be refined by recursively analyzing the selected local frequency domain features and extracting finer-grained features from them.

[0076] Figure 5 This is a schematic diagram illustrating a fusion classification framework according to an embodiment of the present disclosure.

[0077] The extracted local spatial and frequency domain features of the image are input into, for example... Figure 5 The fusion classification framework shown.

[0078] Local spatial and frequency domain features are fused to generate fused image features. Depending on the scenario, the fusion of local spatial features and refined local frequency domain features can be achieved using operations such as `concat`, `add`, or pyramid fusion. Figure 5 As shown, the fusion of local spatial features and refined local frequency features can be further fused using an attention mechanism to automatically learn the importance of different features, including spatial and frequency features.

[0079] The fused features are fed into a convolutional neural network for image authenticity classification.

[0080] In one embodiment of this disclosure, the fusion classification framework can be implemented using ResNet plus SENet. In another embodiment of this disclosure, the fusion classification framework can be implemented using Inception plus SENet. Those skilled in the art will understand that this fusion classification framework can be implemented in other ways, such as using a convolutional neural network plus a channel attention module.

[0081] The model, comprised of a spatial domain feature extraction framework, a frequency domain feature extraction framework, and a fusion classification framework, can be trained using the Focal Loss function to improve its learning ability on difficult samples. The Focal Loss function is a loss function that addresses imbalanced classification by assigning weights to the loss values ​​of samples based on their relative difficulty in classification; that is, it assigns smaller weights α1 to easily distinguishable samples and larger weights α2 to difficult-to-distinguish samples. The expression for this loss function can be written as:

[0082] L SUM =α1×L 易分辨 +α2×L 难分辨

[0083] Therefore, this loss function focuses on samples that are difficult to distinguish.

[0084] Due to the limited number of samples, its training set consists of three parts: first, publicly available fake datasets; second, samples generated using the best-performing existing fake model; and third, fake samples from real-world scenarios collected from various content publishing platforms.

[0085] Those skilled in the art will understand that this training set can be constructed differently as technology advances, and will not be elaborated upon here.

[0086] Figure 6 This is a block diagram illustrating a frequency domain transformation-based deepfake image detection system 600 according to an embodiment of the present disclosure.

[0087] A deepfake image detection system 600 based on frequency domain conversion according to an embodiment of the present disclosure includes a receiving module 602, an extraction module 606, a filtering module 608, a fusion module 610, and a judgment module 612.

[0088] The receiving module 602 receives the image to be detected. In different scenarios, the image to be detected can be a still image, a moving image, a short video, a video, etc.

[0089] The extraction module 606 extracts the local spatial domain features and local frequency domain features of the image to be detected.

[0090] Image authenticity detection targets not semantic content regions, but rather the tampered regions. Therefore, extracting local spatial features of the image to be detected is sufficient to extract the features of the tampered regions. To avoid deep convolutional networks (e.g., CNN, RNN) extracting deep semantic information of images, only a few layers of convolution can be used to extract local spatial features.

[0091] GAN-generated images are difficult to distinguish from real ones, especially in terms of spatial features. However, since the frequency domain features of GAN-generated images contain artifacts (e.g., rasterization artifacts), extraction module 606 further extracts local frequency domain features. Extraction module 606 can convert the image to be detected to the frequency domain and generate local frequency domain features at different frequencies / channels.

[0092] The filtering module 608 uses channel attention to filter local frequency domain features.

[0093] In visual systems, the attention mechanism can be viewed as a dynamic selection process, achieved by adaptively weighting features according to the importance of the input. Channel-attention networks, including SENet, CBAM, and ECANet, can adaptively predict potential key features.

[0094] In this disclosure, to reduce inference time, the filtering module 608 introduces a channel attention mechanism to filter out local frequency domain features under important channels. In one embodiment of this disclosure, the filtering module 608 uses SENet for filtering. Those skilled in the art will understand that different channel attention mechanisms / networks can be used to filter local frequency domain features depending on different scenarios.

[0095] Furthermore, the filtering module 608 can refine the filtered local frequency domain features. In one embodiment of this disclosure, the filtering module 608 recursively analyzes the filtered local frequency domain features and extracts finer-grained features from the obtained local frequency domain features.

[0096] The fusion module 610 fuses local spatial features and selected local frequency features to obtain fused features of the image to be detected.

[0097] Depending on the scenario, the fusion of local spatial features and the selected (or refined) local frequency domain features can be performed by the fusion module 610 using concat and add operations, or pyramid fusion, etc. In one embodiment of this disclosure, the fusion of local spatial features and refined local frequency domain features can be further performed by the fusion module 610 using an attention mechanism to automatically learn the importance of different features, including spatial and frequency domain features.

[0098] The judgment module 612 determines the authenticity of the image to be detected based on the fused features. This determination can be performed using a classification network (e.g., ResNet, Inception).

[0099] Therefore, the frequency domain transformation-based deepfake image detection system disclosed herein incorporates the analysis of both spatial and frequency domain features of the image. By introducing channel attention to filter out important frequency domain features, it can quickly find useful local tampering information, thereby achieving high-performance and robust deepfake image detection.

[0100] The various steps and modules of the frequency domain conversion-based deepfake image detection method and system described above can be implemented in hardware, software, or a combination thereof. If implemented in hardware, the various illustrative steps, modules, and circuits described in connection with this invention can be implemented or executed using a general-purpose processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic components, hardware components, or any combination thereof. A general-purpose processor can be a processor, microprocessor, controller, microcontroller, or state machine, etc. If implemented in software, the various illustrative steps and modules described in connection with this invention can be stored as one or more instructions or codes on a computer-readable medium or transmitted. Software modules implementing the various operations of this invention can reside in storage media such as RAM, flash memory, ROM, EPROM, EEPROM, registers, hard disks, removable disks, CD-ROMs, cloud storage, etc. The storage medium can be coupled to a processor so that the processor can read and write information from / to the storage medium and execute corresponding program modules to implement the various steps of this invention. Moreover, software-based embodiments can be uploaded, downloaded, or remotely accessed through appropriate communication means. Such appropriate means of communication include, for example, the Internet, the World Wide Web, intranets, software applications, cables (including fiber optic cables), magnetic communication, electromagnetic communication (including RF, microwave and infrared communication), electronic communication, or other such means of communication.

[0101] It should also be noted that these embodiments may be described as processes depicted as flowcharts, flow diagrams, structure diagrams, or block diagrams. Although a flowchart may describe the operations as a sequential process, many of these operations can be executed in parallel or concurrently. Furthermore, the order of these operations can be rearranged.

[0102] The disclosed systems, apparatuses, and devices should not be limited in any way. Rather, the invention encompasses all novel and non-obvious features and aspects of the various disclosed embodiments (individually and in various combinations and sub-combinations of each other). The disclosed systems, apparatuses, and devices are not limited to any particular aspect or feature or combination thereof, and no disclosed embodiment is required to have any one or more specific advantages or to solve any particular or all technical problems.

[0103] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many modifications under the guidance of the present invention without departing from the spirit and scope of the claims. All of these modifications fall within the scope of protection of the present invention.

Claims

1. A method for detecting deepfake images based on frequency domain transform, comprising: Receive the image to be detected; Extract the local spatial domain features and local frequency domain features of the image to be detected; The local frequency domain features are filtered using channel attention; The local spatial domain features and the selected local frequency domain features are fused to obtain the fused features of the image to be detected; and The authenticity of the image to be detected is determined based on the fusion features, wherein... The selected local frequency domain features can be further refined before being fused with the local spatial domain features. This further refinement of the selected local frequency domain features includes recursively analyzing the selected local frequency domain features and extracting finer-grained features from them.

2. The method as described in claim 1, wherein the local spatial features of the image to be detected are extracted using a convolutional neural network with a small number of convolutional layers.

3. The method as described in claim 1, wherein extracting local frequency domain features of the image to be detected includes performing frequency domain transformation on the image to be detected.

4. The method as described in claim 1, wherein channel attention filtering of the local frequency domain features is used to filter out important local frequency domain features.

5. The method as described in claim 4, wherein the local frequency domain features are filtered using channel attention via SENet.

6. The method as described in claim 1, wherein the model used in the method is trained using a focus loss function.

7. A deepfake image detection system based on frequency domain transform, comprising: The receiving module receives the image to be detected; The extraction module extracts the local spatial domain features and local frequency domain features of the image to be detected; The filtering module uses channel attention to filter the local frequency domain features; The fusion module fuses the local spatial features and the selected local frequency features to obtain the fusion features of the image to be detected; and The judgment module determines the authenticity of the image to be detected based on the fused features, wherein... The selected local frequency domain features can be further refined before being fused with the local spatial domain features. This further refinement of the selected local frequency domain features includes recursively analyzing the selected local frequency domain features and extracting finer-grained features from them.

8. In the system of claim 7, the extraction module extracts the local spatial features of the image to be detected using a convolutional neural network with a small number of convolutional layers.

9. The system as described in claim 7, wherein the extraction module extracts local frequency domain features of the image to be detected by performing frequency domain transformation on the image to be detected.

10. A computer-readable storage medium storing instructions that, when executed, cause a machine to perform the method as claimed in any one of claims 1-6.

Images