A deep fake face detection method based on dual domain

Abstract

The present application belongs to the field of computer vision and image processing, aiming at the double bottleneck of "semantic overfitting" and "space-frequency feature mislocation" existing in the existing deep fake detection, the present application proposes a deep fake face detection method based on dual domain. The present application constructs a dual domain complementary architecture of dual domain deep fake detection network DDSF-Net based on semantic suppression and space-frequency alignment, the micro mark difference module in the network significantly enhances the residual subtle fake artifacts between pixels by eliminating redundant face attributes. And the space-frequency mapping module uses the block transformation strategy to adaptively mine the abnormal distribution of frequency domain, realizes the multi-dimensional enhancement of the fake features. At the same time, through the complementary fusion of dual domain features, the detection accuracy and generalization are effectively balanced. Extensive experiments on multiple benchmark datasets show that the performance of the method in the same dataset and cross dataset scenarios is significantly better than that of the existing advanced technology.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and image processing, and particularly relates to a method for detecting deep fake faces based on dual domains. Background Technology

[0002] With the deep integration of deep generative models and computer vision, high-fidelity face generation for media is no longer science fiction. Thanks to breakthroughs in generative adversarial networks and implicit neural representation techniques, deepfakes have evolved from early coarse face replacement to sophisticated attribute editing and dynamic expression replay. However, the misuse of this technology has triggered a serious crisis of social trust, posing a severe threat to personal reputation and national information security, making the development of highly reliable face forgery detection technology urgently needed. Although a large number of forgery detection algorithms have emerged in academia, with the rapid iteration of generative technologies, existing models are facing a severe "generalization crisis." Models often perform well on the source training dataset, but their performance drops sharply when used across different datasets.

[0003] To overcome this generalization bottleneck, researchers have focused on uncovering forgery traces with cross-dataset invariance across different signal domains. Based on different feature analysis perspectives, existing methods can be categorized into three types: spatial domain, frequency domain, and dual-domain. First, mainstream spatial domain methods typically input RGB images directly into deep convolutional neural networks (CNNs). However, the inherent inductive bias of CNNs makes them extremely sensitive to high-level semantic content such as shape and texture. This leads the network to tend to establish spurious correlations between forgery labels and specific identities or background textures, rather than truly focusing on subtle tampering fingerprints between pixels. This "semantic overfitting" is the core reason why models fail in cross-dataset tests. Second, while frequency domain methods can effectively capture periodic artifacts left by upsampling operations, traditional global transformations (such as DCT / FFT) destroy the spatial topology of the image, causing the model to be unable to accurately locate local forgery regions. Furthermore, existing methods often rely on pre-set fixed filters and lack the ability to adaptively capture novel and heterogeneous generated artifacts. Finally, although existing dual-stream architectures have attempted to combine the advantages of both, current methods lack in-depth exploration of the complementarity and mapping between spatial details and frequency domain statistical features. They fail to block the interference of semantic noise on micro-trace features at the source, and also fail to effectively correct the misalignment of frequency domain features in the spatial dimension. Summary of the Invention

[0004] To address the dual bottlenecks of semantic overfitting and spatiotemporal feature misalignment in existing deep forgery detection methods, this invention proposes a dual-domain-based deep forgery face detection method. This invention constructs a dual-domain complementary architecture and overcomes the generalization bottleneck by suppressing semantic features and employing spatiotemporal mapping strategies.

[0005] The technical solution protected by this invention is: a deepfake face detection method based on dual domains, specifically performed according to the following steps:

[0006] Step S1: Construct a dual-domain deep forgery detection network DDSF-Net based on semantic suppression and space-frequency alignment. The network adopts a dual-stream architecture with parallel and complementary micro-trace differential stream S and frequency domain mapping stream F.

[0007] Step S2: Micro-trace differential stream S channel: First, the input image is passed through the micro-trace differential module MDM. MDM extracts residual features in parallel from two complementary perspectives: lateral adjacent difference and block-level reference difference. Micro-trace cues are fused through Euclidean distance metric. Then, the heterogeneous features of spatial micro-trace features are output sequentially through N standard residual blocks and M downsampled residual blocks.

[0008] Step S3: Frequency Domain Mapping Stream F Channel: First, the input image is passed through the Spatial Frequency Mapping Module (SFM). SFM uses a block adaptive filtering and inverse reconstruction high-frequency mining mechanism to dynamically capture frequency domain artifacts left by heterogeneous algorithms. Then, the comprehensive micro-trace map is passed through N standard residual blocks and M downsampled residual blocks in sequence to output heterogeneous features of frequency domain mapping features.

[0009] Step S4: The spatial micro-trace features and frequency domain mapping features obtained in steps S2 and S3 are fed into the heterogeneous features extracted by the dual-stream backbone network and then fed into the dual-domain fusion and classification module DFC. The cross-modal channel attention mechanism is used to realize the dynamic recalibration and adaptive aggregation of features and output accurate true and false classification probabilities.

[0010] Furthermore, the specific process of the Micro-Trace Differential Module (MDM) is as follows:

[0011] Step S21, Horizontal Adjacent Difference Branch: For any pixel position in the image Calculate the intensity difference between it and its left neighboring pixels to generate a horizontal micro-trace feature map. ,

[0012] (1)

[0013] in, ;

[0014] Step S22, Block-level reference difference branch: This branch first divides the image into non-overlapping blocks. Local blocks of varying sizes, within each local block In the middle, taking the top-left pixel as the local reference point, for coordinates... The pixel, and the local reference point coordinates of its block. Defined as:

[0015] (2)

[0016] For any pixel within this block Calculate its strength deviation relative to the reference point, then the block-level reference differential characteristic map The calculation is as follows:

[0017] (3)

[0018] Step S23, Micro-trace Feature Fusion: Pixel-level feature fusion is performed using Euclidean distance to construct the final comprehensive micro-trace map. :

[0019] (4)

[0020] in, To prevent numerically unstable minimal constants.

[0021] Furthermore, the specific process of the space-frequency mapping module SFM in step S3 is as follows:

[0022] Step S31, Local Spectrum Mapping and Adaptive Filtering: First, the input RGB image... Convert to grayscale Then, a block discrete cosine transform is used to... Divide into non-overlapping 8×8 local blocks, for the first... Local blocks Its frequency domain corresponding The following calculations were performed using 2D-DCT:

[0023] (5)

[0024] in, These are the spatial pixel coordinates within the block. Frequency coordinates in the frequency domain These are the orthogonal normalization coefficients; finally, a set of learnable adaptive frequency domain filter banks is used to dynamically select key frequency bands. Let be a set of learnable frequency domain filter weights, where This represents the number of target frequency band channels. The filtering process is implemented using the Hadamard product of the frequency domain plot and the weights.

[0025] (6)

[0026] in This represents element-wise multiplication, and then the K filtered responses of all blocks are rearranged according to the original spatial layout to construct the frequency domain map of the mapping. ;

[0027] Step S32, Inverse Reconstruction High-Frequency Residual Mining: First, define the reconstruction operator. For cascaded bilinear downsampling Bilinear upsampling :

[0028] (7)

[0029] Then, high-frequency forged fingerprints are derived in reverse through residual difference analysis:

[0030] (8)

[0031] Finally, through Convolutional layers perform channel fusion to obtain frequency domain fingerprint features with high generalization.

[0032] Furthermore, the specific process of the DFC module in step S4 is as follows:

[0033] First, the dual-stream features are concatenated along the channel dimension, and then passed through a 1×1 convolutional layer for preliminary feature dimensionality reduction and interaction to obtain the fused features. Subsequently, a dual-path pooling strategy was adopted, separately for... Global average pooling and global max pooling are applied, and these two statistical vectors are input into a convolutional transformation layer with shared weights to map the channel features to the latent space. Then, the transformed features of the two branches are added element-wise to fuse different statistical perspectives, and the final channel attention weight map is generated by passing a sigmoid activation function. :

[0034] (9)

[0035] in, This represents the Sigmoid function.

[0036] Finally, residual learning is introduced to apply the generated attention weights. With original features After performing element-wise multiplication, it is then combined with the input features. Add;

[0037] (10)

[0038] Ultimately, the enhanced features The data is fed into a fully connected layer and a Softmax classifier, which outputs true / false prediction results.

[0039] Compared with the prior art, the present invention has the following advantages:

[0040] 1. This invention addresses the "semantic overfitting" problem in deep forgery detection by proposing a Micro-trace Difference Module (MDM). This module introduces a dense residual mining strategy, which effectively suppresses identity semantic information by calculating local gradients and relative texture fluctuations, highlighting subtle forgery traces in residual anomalies, thereby providing a cleaner feature representation for cross-dataset detection.

[0041] 2. To address the issue of missing spatial location information caused by traditional global frequency domain transformation, this invention designs a Spatial-Frequency Mapping Module (SFM). This module utilizes a block-based spectrum transformation strategy to establish a precise mapping between frequency domain response and spatial location, uncovering frequency domain anomalies while preserving local spatial clues. Furthermore, by introducing adaptive filtering and inverse reconstruction high-frequency mining mechanisms, adaptive feature selection for heterogeneous forged spectra is achieved, significantly improving the model's ability to capture frequency domain artifacts from unknown generation algorithms.

[0042] 3. Extensive evaluation experiments were conducted on several mainstream benchmark datasets. Experimental results show that the proposed method exhibits superior detection performance in both intra-dataset and cross-dataset scenarios. In particular, in cross-dataset tests, the AUC on the DFD dataset reached 97.81%; on the DFDC dataset, the AUC was superior to other comparative models, reaching 81.28%, confirming the effectiveness of the proposed dual-domain architecture in solving the generalization problem. Attached Figure Description

[0043] The present invention will now be described in further detail with reference to the accompanying drawings.

[0044] Figure 1 This is the overall network architecture of the solution proposed in this invention.

[0045] Figure 2 This is a structural diagram of the SFM module.

[0046] Figure 3 This is a structural diagram of the dual-domain fusion and classification module.

[0047] Figure 4 This is a schematic diagram illustrating a dual-stream network architecture used only in this embodiment.

[0048] Figure 5 This is an ablation demonstration of the number of filters.

[0049] Figure 6The AUC variation curves are shown under perturbations of different intensities.

[0050] Figure 7 This is a visual analysis diagram of the present invention and Xception. Detailed Implementation

[0051] To make the objectives, features, and advantages of the present invention readily apparent, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0052] A dual-domain deepfake face detection method is proposed. This invention constructs a dual-domain deepfake detection network, DDSF-Net, based on semantic suppression and spatial-frequency alignment, as follows: Figure 1 As shown, the network adopts a parallel and complementary two-stream architecture of micro-trace differential stream S and frequency domain mapped stream F.

[0053] Micro-Trace Differential Flow S: This branch aims to construct a pure micro-trace feature space. Traditional methods often employ high-pass filtering or first-order gradient operators to suppress high-level semantic information. However, these standard operators use linear convolution operations, which are difficult to eliminate spurious correlations in complex feature spaces, thus enhancing forgery traces. To address this issue, this invention proposes a Micro-Trace Differential Module (MDM). This module extracts residual features in parallel from two complementary perspectives: lateral adjacent difference and block-level reference difference. It also fuses micro-trace cues using Euclidean distance, thereby effectively suppressing semantic content with low-frequency characteristics and forcing the model to focus on mining high-frequency anomalies between pixels.

[0054] Lateral Adjacent Differential Branch

[0055] When generative models generate images line by line, they often leave behind discontinuous artifacts or high-frequency noise in the horizontal scanning direction. In order to capture such anisotropic edge abrupt changes, this invention designs a first-order lateral gradient operator.

[0056] For any pixel position in the image Calculate the intensity difference between it and its left neighboring pixels to generate a horizontal micro-trace feature map. ,

[0057] (1)

[0058] in, This operator essentially acts as a horizontal high-pass filter. In the smoothing region of a face, adjacent pixels have similar gray levels. The response approaches zero, thus effectively suppressing low-frequency skin color semantics; while at fake edges or texture abrupt changes, the operator can produce significant residual responses.

[0059] Block-level reference difference branch

[0060] To further eliminate local illumination unevenness and capture longer-range texture dependencies, this invention introduces an anchor-based local difference strategy. This branch first divides the image into non-overlapping segments. Local blocks of varying sizes. Within each local block... In the middle, the top-left pixel is used as the local reference point. For coordinates... The pixel, and the local reference point coordinates of its block. Defined as:

[0061] (2)

[0062] For any pixel within this block Calculate its strength deviation relative to the reference point, then the block-level reference differential characteristic map The calculation is as follows:

[0063] (3)

[0064] Unlike the sliding convolution operation used in traditional difference operators, the block-level reference difference strategy proposed in this invention employs a block-based subtraction method based on fixed anchor points. All pixels within a block are relative to a local reference point. Residual calculations are performed to preserve the relative texture fluctuations of each pixel within a block relative to the reference point. In the smooth areas of a real face, these relative fluctuations are minimal; however, in fake areas, due to generation inconsistencies, the deviations of pixels relative to the anchor point exhibit an abnormal statistical distribution.

[0065] Micro-trace feature fusion

[0066] and Potential forgery traces were revealed from two perspectives: one-dimensional linear correlation and two-dimensional local correlation. To comprehensively utilize these two clues and further enhance the micro-trace signal, this invention employs Euclidean distance metric for pixel-level feature fusion to construct the final comprehensive micro-trace map. :

[0067] (4)

[0068] in, To prevent numerically unstable minimal constants. This comprehensively reflects the magnitude of anomaly at each pixel in the image. Unlike traditional linear feature fusion, the Euclidean distance metric introduced in this invention can utilize the nonlinear mapping properties of the l2 norm to adaptively amplify abrupt anomaly signals between pixels and nonlinearly suppress low-frequency redundant information in smooth regions, thereby enhancing the model's ability to detect weak forgery traces.

[0069] Frequency Domain Mapping Flow F: This branch focuses on reconstructing the correspondence between frequency domain statistical features and spatial topology. While frequency domain analysis excels at capturing synthetic artifacts, traditional global frequency domain transformations disrupt the spatial topology of images, causing the model to lose its ability to locate forged regions. The SFM module proposes a "block-based spectrum mapping" strategy while preserving the spatial layout, and combines it with inverse reconstruction residual learning to achieve accurate capture and spatial localization of frequency domain anomalies. Its structure is as follows: Figure 2 As shown.

[0070] Local Spectrum Mapping and Adaptive Filtering

[0071] First, we will input an RGB image. Convert to grayscale The physical significance of this operation lies in the fact that the traces left by deepfake algorithms are mainly manifested as abnormal fluctuations in pixel intensity, while grayscale images, by integrating brightness information, effectively eliminate redundant interference in the chroma channel that is unrelated to the tampering traces, thereby improving the model's ability to accurately capture the essential texture structure features. Secondly, in order to achieve accurate mapping between frequency domain features and spatial location, this invention abandons global transformation and instead adopts block discrete cosine transform.

[0072] Will Divide into non-overlapping 8×8 local blocks, for the first... Local blocks Its frequency domain corresponding The following calculations were performed using 2D-DCT:

[0073] (5)

[0074] in, These are the spatial pixel coordinates within the block. Frequency coordinates in the frequency domain These are the orthogonal normalization coefficients. This transformation guarantees that each Local spectral statistics for the corresponding spatial location were independently encoded.

[0075] Considering the heterogeneity in frequency band distribution of artifacts left by different generation models, this invention designs a set of learnable adaptive frequency domain filter banks to dynamically select key frequency bands. Let... Let be a set of learnable frequency domain filter weights, where This represents the number of target frequency band channels. The filtering process is implemented using the Hadamard product of the frequency domain plot and the weights.

[0076] (6)

[0077] in This represents element-wise multiplication. Subsequently, the K filtered responses of all blocks are rearranged according to their original spatial layout to construct the mapped frequency domain diagram. This step expands the single-channel grayscale space into a K-channel high-dimensional frequency domain representation, which not only preserves... The spatial resolution also explicitly encodes the specific frequency band intensity at each location.

[0078] Inverse reconstruction high-frequency residual mining

[0079] although While capturing rich spectral information, it still contains low-frequency background signals strongly correlated with semantics (such as smooth transitions in facial contours). To further refine forgery traces, this invention introduces a high-frequency mining strategy based on an inverse reconstruction mechanism. This strategy is based on multi-resolution analysis theory in signal processing: the downsampling process of an image is essentially a low-pass filter, causing irreversible loss of high-frequency details. This property is utilized to estimate the low-frequency components in the features. A reconstruction operator is defined. For cascaded bilinear downsampling Bilinear upsampling :

[0080] (7)

[0081] because It is a feature recovered through lossy compression, which retains only the low-frequency components of the original feature. Therefore, this invention can derive high-frequency forged fingerprints through residual differential analysis.

[0082] (8)

[0083] Among them, the difference term Background redundancy was explicitly eliminated, highlighting high-frequency anomalies. Finally, through Convolutional layers perform channel fusion to obtain frequency domain fingerprint features with high generalization.

[0084] like Figure 1As shown, to balance high-fidelity details of micro-traces with high-level abstract semantics during the feature extraction stage, this invention deploys a cascaded hierarchical perceptual strategy behind the MDM and SFM of the backbone networks in both branches. This strategy sequentially stacks N standard residual blocks and M downsampled residual blocks. The NRB aims to maintain high-resolution feature maps to preserve fine-grained tampering cues, while the DRB is responsible for expanding the receptive field and aggregating deep semantic representations with long-range dependencies. Finally, the heterogeneous features extracted by the dual-stream backbone network are fed into the Dual Domain Fusion and Classification (DFC) module to output true / false classification probabilities.

[0085] To balance high-fidelity details of micro-traces with high-level abstract semantics during the feature extraction stage, this invention separately extracts spatial micro-traces features... Frequency domain mapping characteristics Heterogeneous features extracted via a dual-stream backbone network are fed into the dual-domain fusion and classification module (DFC). A cross-modal channel attention mechanism is used to achieve dynamic recalibration and adaptive aggregation of features, resulting in accurate true / false classification probabilities.

[0086] The calculation process of the DFC module is as follows: Figure 3 As shown, firstly, the dual-stream features are concatenated along the channel dimension, and then passed through a 1×1 convolutional layer for preliminary feature dimensionality reduction and interaction to obtain the fused features. .

[0087] Subsequently, in order to capture the global dependencies between channels, this invention employs a dual-path pooling strategy, separately... Global average pooling and global max pooling are applied. The former aggregates background statistics, while the latter focuses on extracting salient texture features. These two statistical vectors are then fed into a shared-weight convolutional transform layer, mapping the channel features to the latent space.

[0088] Next, the transformation features of the two branches are added element-wise to fuse different statistical perspectives, and the final channel attention weight map is generated by passing the sigmoid activation function. :

[0089] (9)

[0090] in, This represents the Sigmoid function.

[0091] Finally, this invention introduces residual learning to generate attention weights. With original features After performing element-wise multiplication, it is then combined with the input features. Adding them together. This design not only enables dynamic recalibration of features, enhancing the expressive power of key channels, but also effectively guarantees gradient propagation:

[0092] (10)

[0093] Ultimately, the enhanced features The data is fed into a fully connected layer and a Softmax classifier, which outputs true / false prediction results.

[0094] The above provides a detailed description of the dual-domain deep forgery detection network DDSF-Ne based on semantic suppression and spatial-frequency alignment of the present invention. In order to verify the effectiveness of the network structure of the present invention, the present invention and the prior art are experimentally verified below.

[0095] To comprehensively evaluate the detection performance and generalization ability of the proposed model, this invention selected five mainstream public face forgery datasets for experiments, including: FaceForensics++ (FF++), Celeb-DF, DFDC, DFD and DeeperForensics-1.0 (DF1.0).

[0096] FF++ contains 1,000 real videos from YouTube. Each video is manipulated using four widely used face spoofing techniques: DF (DeepFake), FS (FaceSwap), F2F (Face2Face), and NT (NerualTexture), resulting in 4,000 spoofed videos. Each spoofing technique constitutes a subset of the dataset, which is then divided into training, validation, and test sets in a ratio of 720:140:140.

[0097] Other datasets: The Celeb-DF dataset contains Real: 590, Fake: 5639; the DFDC dataset contains Real: 1133, Fake: 4080; the DFD dataset contains Real: 364, Fake: 3068; and the DF1.0 dataset contains Real: 5318, Fake: 1000. These datasets are only used when evaluating cross-dataset performance.

[0098] The proposed model is built on the PyTorch framework and trained and inferred on a single NVIDIA GeForce RTX 4080 SUPER GPU. During data preprocessing, the dlib library is used to detect and crop facial regions, and the resolution is uniformly adjusted to 224×224. The training process adopts an end-to-end approach, using the Adam optimizer, with a batch size of 64 and a total of 100 training epochs. The initial learning rate is set to 0.001, and a cosine annealing strategy is introduced to dynamically adjust the learning rate to ensure convergence stability. The experiments primarily use accuracy (ACC) and area under the receiver operating characteristic (AUC) as the core evaluation metrics.

[0099] To verify the effectiveness of the model architecture in handling different forgery techniques, this embodiment conducted rigorous comparative experiments on the FF++ dataset. The principle of "single-class training - multi-class testing" was adopted, meaning the model was trained on a single subset of forgeries and evaluated on test sets of all four subsets. Table 1 shows the comparison results with the current state-of-the-art (SOTA) method, with bold text indicating the best and underlined text indicating the second best. The data in the table shows that although the proposed model's AUC is slightly lower than that of specific comparative models in some homologous tests, it demonstrates superior overall performance in the more challenging cross-forgery type tests. Notably, the significant performance improvement of the proposed model on the Neural Textures (NT) subset strongly validates the effectiveness of the proposed "spatial-frequency mapping information analysis" strategy. Since NT technology is based on neural rendering, it mainly preserves high-frequency facial textures but introduces specific subtle artifacts, which the frequency domain mapping strategy of this invention can accurately capture. By combining spatial micro-trace difference mechanism, this dual-domain collaborative learning strategy not only enhances the model's ability to represent details, but also gives it a stronger generalization boundary under unknown forgery types.

[0100]

[0101] In practical applications, detection models often face unknown generation methods or significant domain distribution shifts. Therefore, generalization ability is a key indicator for evaluating the practical value of a model. This embodiment trains the model using the FF++ training set and directly tests it on four previously unseen datasets: Celeb-DF, DFDC, DF-1.0, and DFD. The test results are shown in Table 2, where bold text indicates the best performance and underlined text indicates the second best. In cross-dataset testing, the proposed model achieved an AUC of 97.81% on the DFD dataset; on the DFDC dataset, its AUC outperformed other comparative models by 81.28%, even surpassing the latest DTN method by 1.27%. Experimental results demonstrate that the proposed model can learn richer, more subtle traces to reveal forgeries.

[0102]

[0103] To analyze the contributions of each module, this embodiment uses a network consisting of a single-channel Normal Residual Block (NRB) and a Down-sample Residual Block (DRB) as a baseline and designs the following variants for ablation studies: a) Dual line: using only a dual-stream network structure (e.g., Figure 4 (a) As shown); b) Dual+SFM: Add a Spatial Frequency Mapping (SFM) module to the dual-stream network. c) Dual+MDM: Add a Micro-Trace Differential (MDM) module to the dual-stream network.

[0104]

[0105] Table 3 clearly reveals the performance gains of each core module. The MDM module has a significant redundancy removal effect; compared to the SFM module, the average ACC and AUC of the model improved by 5.47% and 3.62%, respectively, after introducing the MDM module. This result proves that the spatial redundancy removal mechanism of MDM can more accurately guide attention to forgery traces, which is particularly outstanding in cross-domain scenarios. The dual-domain features exhibit significant complementarity; compared to the Dual+MDM model, the model proposed in this invention further improves the average ACC and AUC by 5.44% and 5.03%, respectively. This improvement fully confirms that the frequency domain mapping features provided by the SFM module and the spatial features extracted by MDM are highly complementary, and the synergistic effect of the dual domains can significantly enhance the model's ability to classify and distinguish between genuine and fake images.

[0106] Figure 5The impact of the number of filters in the SFM module on detection performance is demonstrated. Experiments show that when the number of filters increases from 6 to 8, the AUC on all datasets shows an upward trend, indicating that moderately increasing the number of parameters helps with feature extraction. However, when the number increases to 9, the accuracy slightly decreases. This indicates that too many filters introduce parameter redundancy, increasing the risk of overfitting. Based on the balance between performance and computational complexity, this embodiment uses 8 filters as the default configuration.

[0107] To evaluate the application value of the proposed model in real-world scenarios, this invention further analyzes the anti-interference capability and model complexity of DDFS-Net. Figure 6 The AUC curves of all comparative methods are presented under four common perturbations of varying intensities. As shown in the figure, the comparative methods exhibit strong stability under color contrast and color satiation perturbations, while the detection performance of the proposed method decreases significantly with increasing perturbation intensity, indicating a high sensitivity of the model to changes in global color distribution. For Gaussian blur attacks, all models show varying degrees of performance degradation, reflecting that the blurring operation severely disrupts the subtle spatial and frequency domain fingerprints upon which the model relies for discrimination. Under JPEG compression perturbations, the proposed method and DNT are significantly affected by block artifacts, and their robustness is weaker than Xception and TwoStream. However, combined with the complexity evaluation analysis in Table 4, it can be seen that the proposed DDFS-Net achieves a good balance between performance and overhead. The proposed method significantly reduces hardware resource dependence while maintaining efficient detection. Although there is a gap compared to the comparative models in robustness tests for specific image degradation, considering the significant reduction in parameter size, FLOPs, and times, this performance loss is completely acceptable in resource-constrained practical applications.

[0108]

[0109] Feature distribution t-SNE visualization analysis

[0110] To visually demonstrate the distribution of the proposed model in the feature space, the model was trained on the FS subset of FF++, and the t-SNE tool was used for visualization analysis on four sub-test sets. The experimental results are as follows: Figure 7 As shown in the figure, compared to the feature aliasing exhibited by Xception, the real samples (blue) and fake samples (red) of the proposed model are clearly separated in the feature space, and the inter-class boundaries are clear. This indicates that the proposed model has strong robustness in distinguishing between real and fake faces.

[0111] Extensive experiments on multiple benchmark datasets demonstrate that the proposed method significantly outperforms state-of-the-art techniques in both intra-dataset and cross-dataset scenarios. Furthermore, ablation studies strongly confirm the effectiveness of the two-stream architecture comprised of MDM and SFM in semantic suppression and spatial-frequency mapping. Although the model exhibits slight deficiencies in robustness, the proposed method significantly reduces hardware resource dependence while maintaining efficient detection. Future work will focus on combining image restoration techniques or self-supervised learning to further enhance the model's robustness under extreme degradation conditions.

[0112] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made without departing from the spirit of the present invention within the scope of knowledge known to those skilled in the art.

Claims

1. A method for detecting deepfake faces based on a dual-domain approach, characterized in that, Please follow these steps: Step S1: Construct a dual-domain deep forgery detection network DDSF-Net based on semantic suppression and space-frequency alignment. The network adopts a dual-stream architecture with parallel and complementary micro-trace differential stream S and frequency domain mapping stream F. Step S2: Micro-trace differential stream S channel: First, the input image is passed through the micro-trace differential module MDM. MDM extracts residual features in parallel from two complementary perspectives: lateral adjacent difference and block-level reference difference. Micro-trace cues are fused through Euclidean distance metric. Then, the heterogeneous features of spatial micro-trace features are output sequentially through N standard residual blocks and M downsampled residual blocks. Step S3: Frequency Domain Mapping Stream F Channel: First, the input image is passed through the Spatial Frequency Mapping Module (SFM). SFM uses a block adaptive filtering and inverse reconstruction high-frequency mining mechanism to dynamically capture frequency domain artifacts left by heterogeneous algorithms. Then, the comprehensive micro-trace map is passed through N standard residual blocks and M downsampled residual blocks in sequence to output heterogeneous features of frequency domain mapping features. Step S4: The spatial micro-trace features and frequency domain mapping features obtained in steps S2 and S3 are fed into the heterogeneous features extracted by the dual-stream backbone network and then fed into the dual-domain fusion and classification module DFC. The cross-modal channel attention mechanism is used to realize the dynamic recalibration and adaptive aggregation of features and output accurate true and false classification probabilities.

2. The method for detecting deepfake faces based on a dual-domain approach according to claim 1, characterized in that: The specific process of the Micro-Trace Differential Module (MDM) is as follows: Step S21, Horizontal Adjacent Difference Branch: For any pixel position in the image Calculate the intensity difference between it and its left neighboring pixels to generate a horizontal micro-trace feature map. , （1） in, ; Step S22, Block-level reference difference branch: This branch first divides the image into non-overlapping blocks. Local blocks of varying sizes, within each local block In the middle, taking the top-left pixel as the local reference point, for coordinates... The pixel, and the local reference point coordinates of its block. Defined as: （2） For any pixel within this block Calculate its strength deviation relative to the reference point, then the block-level reference differential characteristic map The calculation is as follows: （3） Step S23, Micro-trace Feature Fusion: Pixel-level feature fusion is performed using Euclidean distance to construct the final comprehensive micro-trace map. : （4） in, To prevent numerically unstable minimal constants.

3. The method for detecting deepfake faces based on a dual-domain approach according to claim 2, characterized in that, The specific process of the space-frequency mapping module SFM in step S3 is as follows: Step S31, Local Spectrum Mapping and Adaptive Filtering: First, the input RGB image... Convert to grayscale Then, a block discrete cosine transform is used to... Divide into non-overlapping 8×8 local blocks, for the first... Local blocks Its frequency domain corresponding The following calculations were performed using 2D-DCT: （5） in, These are the spatial pixel coordinates within the block. Frequency coordinates in the frequency domain These are the orthogonal normalization coefficients; finally, a set of learnable adaptive frequency domain filter banks is used to dynamically select key frequency bands. Let be a set of learnable frequency domain filter weights, where This represents the number of target frequency band channels. The filtering process is implemented using the Hadamard Product of the frequency domain plot and the weights. （6） in This represents element-wise multiplication, and then the K filtered responses of all blocks are rearranged according to the original spatial layout to construct the frequency domain map of the mapping. ; Step S32, Inverse Reconstruction High-Frequency Residual Mining: First, define the reconstruction operator. For cascaded bilinear downsampling Bilinear upsampling : （7） Then, high-frequency forged fingerprints are derived in reverse through residual difference analysis: （8） Finally, through Convolutional layers perform channel fusion to obtain frequency domain fingerprint features with high generalization.

4. The method for detecting deepfake faces based on a dual-domain approach according to claim 3, characterized in that, The specific process of the DFC module in step S4 is as follows: First, the dual-stream features are concatenated along the channel dimension, and then passed through a 1×1 convolutional layer for preliminary feature dimensionality reduction and interaction to obtain the fused features. ; Subsequently, a dual-pooling strategy was adopted to separately pool the pooling pools. Global average pooling and global max pooling are applied, and these two statistical vectors are input into a convolutional transformation layer with shared weights to map the channel features to the latent space. Then, the transformed features of the two branches are added element-wise to fuse different statistical perspectives, and the final channel attention weight map is generated by passing a sigmoid activation function. : （9） in, Represents the Sigmoid function; Finally, residual learning is introduced to apply the generated attention weights. With original features After performing element-wise multiplication, it is then combined with the input features. Add; （10） Ultimately, the enhanced features The data is fed into a fully connected layer and a Softmax classifier, which outputs true / false prediction results.

Images