Unified unsupervised depth forgery detection method based on prototype guidance and double hyperspheres

Abstract

The invention discloses a unified unsupervised depth forgery detection method based on prototype guidance and double hyperspheres, and the method comprises the following steps: S1, extracting visual artifact features generated by depth forgery, and achieving the generation of pseudo labels through the clustering of a Gaussian mixture model; s2, performing comparative learning through a category prototype of momentum updating; and S3, realizing effective fusion of a feature space and a geometric decision by respectively constructing independent hyper-spheres for real and forged samples, and constructing a dual-depth support vector data description framework. The system has the beneficial effects that the system is composed of three core modules: a visual artifact feature-based pseudo label generator provides a reliable supervision signal, and the visual artifact feature-based pseudo label generator provides a visual artifact feature-based pseudo label description framework; a prototype guided contrast learning (PGCL) module enhances the discrimination capability through a prototype of momentum update, and a dual-depth support vector data description (Dual-DeepSVDD) module constructs a dual-hyperspherical decision boundary of true and false samples, thereby realizing effective integration of feature learning and geometric decision.

Description

Technical Field

[0001] This invention relates to the fields of prototype guidance and dual hypersphere technology, specifically to a unified unsupervised deepfake detection method based on prototype guidance and dual hypersphere. Background Technology

[0002] To address the labeling dependency problem of supervised methods, researchers have proposed using unsupervised learning-based single-class anomaly detection models, treating deepfake detection as an anomaly detection task. These models treat real samples as normal and fake samples as anomalies. Existing anomaly detection methods mainly include the following categories: (1) anomaly detection methods based on reconstruction error; (2) anomaly detection frameworks based on deep features; and (3) self-supervised anomaly detection mechanisms. The core idea of ​​these anomaly detection methods is to learn the feature distribution of "normal" data from a large number of real samples, and then mark samples that significantly deviate from this distribution as anomalies.

[0003] Although this method effectively reduces the dependence on labeled samples and avoids the expensive data labeling process, it also brings several key limitations, which seriously restrict its application effect in deepfake detection. The problems include the following: (1) The fundamental limitation of feature representation ability: existing methods only learn the feature distribution of real samples and lack effective discriminative features to distinguish between real and fake samples. Since it only deals with real data, the model cannot understand the feature patterns of fake samples and performs poorly when faced with high-quality deepfake content; (2) The problem of fuzzy decision boundary: since it only models a single normal class, the anomaly detection method cannot capture the discriminative features between real and fake samples. In particular, when the feature distributions of fake samples and real samples highly overlap, it is difficult to establish a clear classification boundary, which leads to fake samples near the decision boundary being misclassified as real samples, thus limiting the detection performance. Summary of the Invention

[0004] The purpose of this invention is to provide a unified unsupervised deepfake detection method based on prototype guidance and dual hyperspheres to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a unified unsupervised depth forgery detection method based on prototype guidance and dual hyperspheres, comprising the following:

[0006] S1: Extract the visual artifact features generated by deepfakes and use Gaussian mixture model clustering to generate pseudo-labels;

[0007] S2: Comparative learning through class prototypes updated by momentum;

[0008] S3: By constructing independent hyperspheres for real and fake samples respectively, the feature space and geometric decision are effectively integrated to build a dual-depth support vector data description framework;

[0009] S4: Effectively integrates pseudo-label generation, feature learning, and decision boundary construction into the overall model architecture.

[0010] Preferably, step S1 includes the following:

[0011] S11: In the process of generating deepfake technology, subtle visual artifacts inevitably occur in specific facial areas. Although these artifacts are difficult to detect visually, they exhibit obvious statistical differences in the texture frequency domain. Reliable supervision signals are then generated through multi-region texture analysis and unsupervised clustering.

[0012] S12: Deepfake algorithms exhibit significant differences in their processing capabilities across different facial regions, leading to regionalized artifact distribution. Based on the inherent limitations of facial geometry and forgery techniques, three artifact-sensitive regions were identified: eyes, mouth, and facial boundary regions. First, a dlib face detector was used to extract 68 facial key points. Based on these key points, precise geometric masks for each region were constructed. The eye region was generated using convex hull operations on the left and right eye key points, primarily capturing blink coordination and edge sharpness anomalies. The mouth region combined with clustering algorithms achieved precise segmentation of teeth and lips, focusing on detecting anomalies in tooth texture consistency. The facial boundary region was constructed using inner and outer boundary masks to identify trace features in face-swapping fusion.

[0013] S13: Considering that visual artifacts mainly manifest as subtle perturbations in local texture patterns, a Laws texture filter is used for multi-scale feature extraction. The Laws filter effectively responds to changes in texture directionality, roughness, and contrast, and is highly sensitive to texture inconsistencies introduced by depth sensing. Let the preprocessed image be Iprep, and the Laws filter be... Texture features of each region are extracted using the mean of the filtered energy: in ⊙ represents convolution operation, Mregion represents the mask of the corresponding region, and μ(·) represents the mean operation within the mask region. The final VAF feature vector is constructed by connecting the texture features of the three regions:

[0014] Preferably, S14: Due to the high dimensionality of the extracted features and the scale differences between different regions, the VAF features are first standardized, and then Principal Component Analysis (PCA) is used to retain 98% of the variance information for dimensionality reduction. In the dimensionality-reduced feature space, Gaussian Mixture Model (GMM) is used for unsupervised binary clustering. GMM can effectively model the probability structure of the feature distribution, and its density function is expressed as: Where πk is the mixing weight, μ k and ∑k These are the mean and covariance matrix of the k-th Gaussian component, respectively;

[0015] S15: GMM clustering generates two clusters and assigns a corresponding pseudo-label to each sample. Due to the lack of prior knowledge, it is impossible to directly determine which cluster represents the real sample or the fake sample. However, this label uncertainty is within the acceptable range of the framework because the core function of the pseudo-label is to achieve binary partitioning of the data to provide a supervisory signal, rather than an absolutely accurate category label. In addition, the discriminant mechanism based on the difference in hypersphere radius during the inference stage can effectively solve this problem. Although the clustering process introduces a certain proportion of label noise, the subsequent prototype contrastive learning has a strong noise tolerance. It naturally corrects the misassignment in the iterative optimization through momentum-updated category prototypes, prompting samples with similar intrinsic properties to converge to the correct feature space. Through carefully designed visual artifact feature extraction and clustering strategies, high-quality pseudo-labels are generated for the unlabeled data, providing key supervisory signals for subsequent prototype contrastive learning and decision boundary construction.

[0016] Preferably, step S2 includes the following:

[0017] S21: Comparative learning through momentum-updated class prototypes promotes intra-class tightness and inter-class separation, thereby enabling the encoder to learn more discriminative feature representations;

[0018] S22: The PGCL module has the ability to tolerate noise from erroneous pseudo-labels. Firstly, the category prototype is formed by aggregating features from a large number of samples. Therefore, even if there are a small number of erroneously labeled samples, they will not significantly deviate from the prototype center, thus ensuring statistical stability. Secondly, during the global optimization process, a large number of correct samples dominate the feature learning direction, making the impact of a small number of erroneously labeled samples on the overall discrimination ability relatively small.

[0019] S23: In deepfake detection, traditional contrastive learning data augmentation methods are mainly designed for image classification tasks and cannot be directly applied to detection tasks based on subtle feature differences between real and fake images. Since real and fake samples are visually highly similar but have subtle differences in local features, an asymmetric data augmentation strategy specifically adapted to face forgery detection is designed. The core design principle of this strategy is to achieve a balance between semantic integrity and sample diversity. The weak augmentation view retains more original semantic information through gentle transformation to ensure stable updates of the category prototype; conversely, the strong augmentation view enhances the generalization ability of the model through aggressive transformation.

[0020] S25: Five targeted data augmentation methods are designed: First, random erasure simulates occlusion interference in real-world scenes by selectively occluding local facial areas; second, random cropping adapts to different field of view and imaging conditions through multi-scale cropping operations; third, horizontal flipping simulates changes in facial orientation under different shooting angles; fourth, Gaussian noise enhances the model's robustness to image quality degradation by injecting random noise; and fifth, color jitter simulates diverse lighting conditions and shooting environments by adjusting brightness, contrast, saturation, and hue parameters.

[0021] S26: The strong enhancement strategy employs all five enhancement techniques to maximize sample diversity through aggressive transformations, thereby improving the model's adaptability to real-world interference. In contrast, the weak enhancement strategy employs four techniques, excluding random erasure, to increase sample diversity while maintaining the integrity of facial semantic information through gentle transformations.

[0022] Preferably, S27: The InfoNCE loss function, as a classic contrastive learning method, has achieved significant success in multiple computer vision tasks such as image classification and object detection. This method learns discriminative representations by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs. Its mathematical expression is: Where hi and hi+ represent the feature representations of the sample and its positive sample, respectively, hj- represents the negative sample feature, i is the temperature parameter, and N is the batch size;

[0023] S28: The InfoNCE loss function faces significant challenges in deepfake detection. In deepfake detection, the feature distributions of real and fake samples severely overlap in the high-dimensional feature space, resulting in weak distinguishing signals between positive and negative sample pairs. This hinders InfoNCE from effectively learning discriminative feature representations through instance-level comparative learning. To address this issue, prototype-based contrastive learning is introduced, transforming traditional inter-instance comparisons into comparisons between instances and class prototypes. By aggregating samples of the same category, a class prototype with stronger discriminative power is constructed. The positive and negative sample comparison signals during the contrastive learning process become clearer and more stable, effectively alleviating the difficulty in distinguishing due to overlapping feature distributions. Considering the different functions of weak and strong augmentation views in feature learning, a weighted fusion strategy is adopted to calculate the class prototype, given by the following formula: Pk = a·fkiweak + (1-a)·fkistrong, where μ is the momentum coefficient. This represents the weighted average of the features of category k in the current batch. This momentum update mechanism maintains the continuity of historical feature information while adaptively incorporating the feature distribution changes of the current batch, thereby effectively improving the category representativeness of the prototype.

[0024] S29: Based on dynamically updated category prototypes, a prototype contrast loss function is designed. Unlike the traditional InfoNCE method, it uses a simpler and more efficient cross-entropy loss to achieve prototype-level contrast. Its mathematical expression is: Where fi is the L2 normalized feature vector of the sample, Pci represents the prototype vector of the category to which sample i belongs, Pk represents the prototype vector of category k, and the number β is the temperature parameter. This loss function achieves the optimization goal of clustering sample features toward the correct category prototype and moving away from the incorrect category prototype through the softmax normalization mechanism. Through the above design, the PGCL module can effectively utilize the pseudo-label information generated in the first stage, and then optimize the discriminative ability of the feature representation through prototype-guided contrastive learning.

[0025] Preferably, step S3 includes the following:

[0026] S31: The pseudo-labels in the first stage provide reliable supervision signals for its training, while the discriminative feature representations obtained by the PGCL module training in the second stage lay a solid foundation for the construction of the decision space.

[0027] S32: Next, a Dual-DeepSVDD model with the same structure but independent parameters is constructed. Each model contains a feature adaptation layer and an orthogonal transformation layer. The feature adaptation layer maps the features of the pre-trained encoder to a compact representation space suitable for hypersphere modeling through batch normalization and the ReLU activation function. The orthogonal transformation layer uses an orthogonal initialization strategy to preserve the geometric properties of the features. By employing an alternating training mechanism, the parameters of the two models are updated sequentially in each epoch, avoiding model degradation problems that may be caused by joint training. This promotes the formation of mutually separate and compact representations of the two hyperspheres in the feature space. The hypersphere center is updated using a momentum-based strategy, the formula of which is: Where c(t) represents the normalized hypersphere center after iteration t, and μ is the momentum coefficient. Indicates the center of the current batch feature;

[0028] S33: In deepfake detection, real samples originate from real physical imaging processes, therefore their feature representations exhibit a relatively stable and concentrated distribution pattern in the depth feature space. Conversely, fake samples originate from the generation processes of various artificial synthesis algorithms. Due to the lack of unified physical imaging constraints and the involvement of multiple generation strategies and forgery techniques, their feature representations exhibit significant instability and dispersion in the depth space. Based on this difference in distribution characteristics, we determine the category assignment by comparing the hypersphere radii: smaller hyperspheres correspond to real samples, and larger hyperspheres correspond to fake samples (e.g., ...). Figure 6 As shown in the figure, this hypothesis is verified by t-SNE visualization results.

[0029] Preferred, S34: Figure 6 The blue area (real samples) exhibits a relatively compact distribution, while the red area (fake samples) shows a more dispersed feature distribution pattern. During the inference phase, classification decisions are made by calculating the cosine distance from the test sample to the centers of the two hyperspheres. The calculation formula is as follows:

[0030] S35: where φ(x) represents the L2 normalized feature vector, c represents the center of the hypersphere, and samples are assigned to the category corresponding to the closer hypersphere, thereby achieving effective geometric decision-making in the high-dimensional feature space.

[0031] 8. The unified unsupervised deepfake detection method based on prototype guidance and dual hyperspheres according to claim 1, characterized in that: step S4 includes the following:

[0032] S41: Based on the above steps, extract texture features from key facial regions and apply Gaussian mixture model (GMM) clustering to generate reliable pseudo-labels. These pseudo-labels provide key supervision signals for subsequent training stages.

[0033] S42: Next, the PGCL module is introduced. This module generates two augmented views for each sample through weak data augmentation and strong data augmentation. This module learns by comparing these augmented views with the class prototypes updated with momentum. The samples are brought closer to the corresponding class prototypes while moving away from other class prototypes. Through this process, the PGCL module promotes intra-class compactness and inter-class separability, enabling the encoder to learn highly discriminative feature representations.

[0034] S43: Finally, we further designed the Dual-DeepSVDD module to construct the decision space. This module establishes two independent hyperspheres for real data and fake data respectively. The module learns a compact feature distribution by dynamically updating the center and radius of the hypersphere. During inference, the sample is classified by calculating its distance from the center of the two hyperspheres, thereby achieving accurate detection.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] The system consists of three core modules: a pseudo-label generator based on visual artifact features provides reliable supervision signals; a prototype-guided contrastive learning (PGCL) module enhances the discrimination ability through momentum-updated class prototypes; and a dual-deep support vector data description (Dual-DeepSVDD) module constructs a dual hyperspherical decision boundary for true and false samples, realizing the effective integration of feature learning and geometric decision-making.

[0037] 1. In the in-dataset evaluation, PGDD-Net achieved an accuracy of 88.22% and an AUC of 92.23% on the FF++C23 dataset, which is the best performance among all unsupervised methods. Compared with existing unsupervised methods: DFUCL achieves 75.67% accuracy and 82.64% AUC, UNTAG achieves 78.52% accuracy and 86.71% AUC, FUDD achieves 85.14% accuracy and 89.27% ​​AUC, and OSDD achieves 84.21% accuracy and 90.45% AUC, all of which are significantly lower than our method.

[0038] 2. In cross-dataset generalization evaluation, PGDD-Net demonstrates excellent generalization ability, achieving an average performance of 69.96% accuracy and 80.62% AUC across four datasets. This significantly outperforms all existing unsupervised methods. Specifically, it achieves 65.83% accuracy and 71.01% AUC on the DFDC dataset, 53.15% accuracy and 80.02% AUC on the DFR dataset, and 72.65% accuracy and 79.21% AUC on the CelebDF dataset. These are the best results among unsupervised methods. In contrast, the second-best unsupervised method, FUDD, only achieves an average performance of 67.71% accuracy and 77.92% AUC.

[0039] 3. Based on the above results, the three-stage framework is verified to have strong adaptability in the face of unknown forgery techniques. Among them, the dual hypersphere modeling effectively solves the limitations of the traditional single hypersphere method in deepfake detection. By constructing independent real and fake sample hyperspheres, clearer inter-class boundary division is achieved. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0041] Figure 2 This is a comparative analysis diagram between the present invention and existing methods;

[0042] Figure 3 This is a flowchart of the pseudo-label generation process of the present invention;

[0043] Figure 4 This is a comparative schematic diagram illustrating the prototype guidance of the present invention;

[0044] Figure 5 This is a schematic diagram illustrating the dual-depth support vector data description of the present invention;

[0045] Figure 6 This diagram illustrates the experimental verification of the distribution assumptions of the real sample concentration and the fake sample dispersion in this invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Please see Figures 1-6 This invention provides a technical solution: a unified unsupervised depth forgery detection method based on prototype guidance and dual hyperspheres, comprising the following:

[0048] S1: Extract the visual artifact features generated by deepfakes and use Gaussian mixture model clustering to generate pseudo-labels;

[0049] S2: Comparative learning through class prototypes updated by momentum;

[0050] S3: By constructing independent hyperspheres for real and fake samples respectively, the feature space and geometric decision are effectively integrated to build a dual-depth support vector data description framework;

[0051] S4: Effectively integrates pseudo-label generation, feature learning, and decision boundary construction into the overall model architecture.

[0052] Step S1 includes the following:

[0053] S11: In the process of generating deepfake technology, subtle visual artifacts inevitably occur in specific facial areas. Although these artifacts are difficult to detect visually, they exhibit obvious statistical differences in the texture frequency domain. Reliable supervision signals are then generated through multi-region texture analysis and unsupervised clustering.

[0054] S12: Deepfake algorithms exhibit significant differences in their processing capabilities across different facial regions, leading to regionalized artifact distribution. Based on the inherent limitations of facial geometry and forgery techniques, three artifact-sensitive regions were identified: eyes, mouth, and facial boundary regions. First, a dlib face detector was used to extract 68 facial key points. Based on these key points, precise geometric masks for each region were constructed. The eye region was generated using convex hull operations on the left and right eye key points, primarily capturing blink coordination and edge sharpness anomalies. The mouth region combined with clustering algorithms achieved precise segmentation of teeth and lips, focusing on detecting anomalies in tooth texture consistency. The facial boundary region was constructed using inner and outer boundary masks to identify trace features in face-swapping fusion.

[0055] S13: Considering that visual artifacts mainly manifest as subtle perturbations in local texture patterns, a Laws texture filter is used for multi-scale feature extraction. The Laws filter effectively responds to changes in texture directionality, roughness, and contrast, and is highly sensitive to texture inconsistencies introduced by depth sensing. Let the preprocessed image be Iprep, and the Laws filter be... Texture features of each region are extracted using the mean of the filtered energy: in ⊙ represents convolution operation, Mregion represents the mask of the corresponding region, and μ(·) represents the mean operation within the mask region. The final VAF feature vector is constructed by connecting the texture features of the three regions:

[0056] In S14: Due to the high dimensionality of the extracted features and the scale differences between different regions, the visual artifact features are first standardized. Then, Principal Component Analysis (PCA) is used to retain 98% of the variance information for dimensionality reduction. In the dimensionality-reduced feature space, Gaussian Mixture Model (GMM) is used for unsupervised binary clustering. GMM can effectively model the probability structure of the feature distribution, and its density function is expressed as: Where πk is the mixing weight, μ k and Σ k These are the mean and covariance matrix of the k-th Gaussian component, respectively;

[0057] S15: GMM clustering generates two clusters and assigns a corresponding pseudo-label to each sample. Due to the lack of prior knowledge, it is impossible to directly determine which cluster represents the real sample or the fake sample. However, this label uncertainty is within the acceptable range of the framework because the core function of the pseudo-label is to achieve binary partitioning of the data to provide a supervisory signal, rather than an absolutely accurate category label. In addition, the discriminant mechanism based on the difference in hypersphere radius during the inference stage can effectively solve this problem. Although the clustering process introduces a certain proportion of label noise, the subsequent prototype contrastive learning has a strong noise tolerance. It naturally corrects the misassignment in the iterative optimization through momentum-updated category prototypes, prompting samples with similar intrinsic properties to converge to the correct feature space. Through carefully designed visual artifact feature extraction and clustering strategies, high-quality pseudo-labels are generated for the unlabeled data, providing key supervisory signals for subsequent prototype contrastive learning and decision boundary construction.

[0058] Step S2 includes the following:

[0059] S21: Comparative learning through momentum-updated class prototypes promotes intra-class tightness and inter-class separation, thereby enabling the encoder to learn more discriminative feature representations;

[0060] S22: The PGCL module has the ability to tolerate noise from erroneous pseudo-labels. Firstly, the category prototype is formed by aggregating features from a large number of samples. Therefore, even if there are a small number of erroneously labeled samples, they will not significantly deviate from the prototype center, thus ensuring statistical stability. Secondly, during the global optimization process, a large number of correct samples dominate the feature learning direction, making the impact of a small number of erroneously labeled samples on the overall discrimination ability relatively small.

[0061] S23: In deepfake detection, traditional contrastive learning data augmentation methods are mainly designed for image classification tasks and cannot be directly applied to detection tasks based on subtle feature differences between real and fake images. Since real and fake samples are visually highly similar but have subtle differences in local features, an asymmetric data augmentation strategy specifically adapted to face forgery detection is designed. The core design principle of this strategy is to achieve a balance between semantic integrity and sample diversity. The weak augmentation view retains more original semantic information through gentle transformation to ensure stable updates of the category prototype; conversely, the strong augmentation view enhances the generalization ability of the model through aggressive transformation.

[0062] S25: Five targeted data augmentation methods are designed: First, random erasure simulates occlusion interference in real-world scenes by selectively occluding local facial areas; second, random cropping adapts to different field of view and imaging conditions through multi-scale cropping operations; third, horizontal flipping simulates changes in facial orientation under different shooting angles; fourth, Gaussian noise enhances the model's robustness to image quality degradation by injecting random noise; and fifth, color jitter simulates diverse lighting conditions and shooting environments by adjusting brightness, contrast, saturation, and hue parameters.

[0063] S26: The strong enhancement strategy employs all five enhancement techniques to maximize sample diversity through aggressive transformations, thereby improving the model's adaptability to real-world interference. In contrast, the weak enhancement strategy employs four techniques, excluding random erasure, to increase sample diversity while maintaining the integrity of facial semantic information through gentle transformations.

[0064] S27: Utilizing the InfoNCE loss function as a classic contrastive learning method, it has achieved significant success in multiple computer vision tasks such as image classification and object detection. This method learns discriminative representations by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs. Its mathematical expression is: Where hi and hi+ represent the feature representations of the sample and its positive sample, respectively, hj- represents the negative sample feature, i is the temperature parameter, and N is the batch size;

[0065] S28: The InfoNCE loss function faces significant challenges in deepfake detection. In deepfake detection, the feature distributions of real and fake samples severely overlap in the high-dimensional feature space, resulting in weak distinguishing signals between positive and negative sample pairs. This hinders InfoNCE from effectively learning discriminative feature representations through instance-level comparative learning. To address this issue, prototype-based contrastive learning is introduced, transforming traditional inter-instance comparisons into comparisons between instances and class prototypes. By aggregating samples of the same category, a class prototype with stronger discriminative power is constructed. The positive and negative sample comparison signals during the contrastive learning process become clearer and more stable, effectively alleviating the difficulty in distinguishing due to overlapping feature distributions. Considering the different functions of weak and strong augmentation views in feature learning, a weighted fusion strategy is adopted to calculate the class prototype, given by the following formula: Pk = a·fkiweak + (1-a)·fkistrong, where μ is the momentum coefficient. This represents the weighted average of the features of category k in the current batch. This momentum update mechanism maintains the continuity of historical feature information while adaptively incorporating the feature distribution changes of the current batch, thereby effectively improving the category representativeness of the prototype.

[0066] S29: Based on dynamically updated category prototypes, a prototype contrast loss function is designed. Unlike the traditional InfoNCE method, it uses a simpler and more efficient cross-entropy loss to achieve prototype-level contrast. Its mathematical expression is: Where fi is the L2 normalized feature vector of the sample, Pci represents the prototype vector of the category to which sample i belongs, Pk represents the prototype vector of category k, and the number β is the temperature parameter. This loss function achieves the optimization goal of clustering sample features toward the correct category prototype and moving away from the incorrect category prototype through the softmax normalization mechanism. Through the above design, the PGCL module can effectively utilize the pseudo-label information generated in the first stage, and then optimize the discriminative ability of the feature representation through prototype-guided contrastive learning.

[0067] Step S3 includes the following:

[0068] S31: The pseudo-labels in the first stage provide reliable supervision signals for its training, while the discriminative feature representations obtained by the PGCL module training in the second stage lay a solid foundation for the construction of the decision space.

[0069] S32: Next, a Dual-DeepSVDD model with the same structure but independent parameters is constructed. Each model contains a feature adaptation layer and an orthogonal transformation layer. The feature adaptation layer maps the features of the pre-trained encoder to a compact representation space suitable for hypersphere modeling through batch normalization and the ReLU activation function. The orthogonal transformation layer uses an orthogonal initialization strategy to preserve the geometric properties of the features. By employing an alternating training mechanism, the parameters of the two models are updated sequentially in each epoch, avoiding model degradation problems that may be caused by joint training. This promotes the formation of mutually separate and compact representations of the two hyperspheres in the feature space. The hypersphere center is updated using a momentum-based strategy, the formula of which is: Where c(t) represents the normalized hypersphere center after iteration t, and μ is the momentum coefficient. Indicates the center of the current batch feature;

[0070] S33: In deepfake detection, real samples originate from real physical imaging processes, therefore their feature representations exhibit a relatively stable and concentrated distribution pattern in the depth feature space. Conversely, fake samples originate from the generation processes of various artificial synthesis algorithms. Due to the lack of unified physical imaging constraints and the involvement of multiple generation strategies and forgery techniques, their feature representations exhibit significant instability and dispersion in the depth space. Based on this difference in distribution characteristics, we determine the category assignment by comparing the hypersphere radii: smaller hyperspheres correspond to real samples, and larger hyperspheres correspond to fake samples (e.g., ...). Figure 6 As shown in the figure, this hypothesis is verified by t-SNE visualization results.

[0071] Among them, S34: Figure 6 The blue area (real samples) exhibits a relatively compact distribution, while the red area (fake samples) shows a more dispersed feature distribution pattern. During the inference phase, classification decisions are made by calculating the cosine distance from the test sample to the centers of the two hyperspheres. The calculation formula is as follows:

[0072] S35: where φ(x) represents the L2 normalized feature vector, c represents the center of the hypersphere, and samples are assigned to the category corresponding to the closer hypersphere, thereby achieving effective geometric decision-making in the high-dimensional feature space.

[0073] 8. The unified unsupervised deepfake detection method based on prototype guidance and dual hyperspheres according to claim 1, characterized in that: step S4 includes the following:

[0074] S41: Based on the above steps, extract texture features from key facial regions and apply Gaussian mixture model (GMM) clustering to generate reliable pseudo-labels for unlabeled data. These pseudo-labels provide key supervision signals for subsequent training stages.

[0075] S42: Next, the PGCL module is introduced. This module generates two augmented views for each sample through weak data augmentation and strong data augmentation. This module learns by comparing these augmented views with the class prototypes updated with momentum. The samples are brought closer to the corresponding class prototypes while moving away from other class prototypes. Through this process, the PGCL module promotes intra-class compactness and inter-class separability, enabling the encoder to learn highly discriminative feature representations.

[0076] S43: Finally, we further designed the Dual-DeepSVDD module to construct the decision space. This module establishes two independent hyperspheres for real data and fake data respectively. The module learns a compact feature distribution by dynamically updating the center and radius of the hypersphere. During inference, the sample is classified by calculating its distance from the center of the two hyperspheres, thereby achieving accurate detection.

[0077] Specifically, deepfake technology inevitably produces subtle visual artifacts in specific facial regions during the generation process. Although these artifacts are difficult to detect visually, they exhibit significant statistical differences in the texture frequency domain. Therefore, reliable supervision signals are generated through multi-region texture analysis and unsupervised clustering. The deepfake algorithm's processing capabilities for different facial regions vary significantly, leading to regionalized artifact distribution. Based on the inherent limitations of facial geometry and deepfake technology, three artifact-sensitive regions are identified: the eyes, mouth, and facial boundary regions. First, a dlib face detector is used to extract 68 facial key points. Based on these key points, precise geometric masks for each region are constructed. The eye region is generated using convex hull operations on key points of both eyes, primarily capturing anomalies in blink coordination and edge sharpness. The mouth region utilizes clustering algorithms for precise segmentation of teeth and lips, focusing on detecting anomalies in tooth texture consistency. The facial boundary region is constructed using inner and outer boundary masks to identify trace features in face-swapping fusion. Considering that visual artifacts mainly manifest as subtle perturbations in local texture patterns, a Laws texture filter is used for multi-scale feature extraction. The Laws filter effectively responds to changes in texture directionality, roughness, and contrast, and is highly sensitive to texture inconsistencies introduced by deepfakes. Let the preprocessed image be Iprep, and the Laws filter be... Texture features of each region are extracted using the mean of the filtered energy: in ⊙ represents convolution operation, Mregion represents the mask of the corresponding region, and μ(·) represents the mean operation within the mask region. The final visual artifact feature vector is constructed by connecting the texture features of the three regions: Due to the high dimensionality of the extracted features and the scale differences between different regions, the VAF features are first standardized. Then, Principal Component Analysis (PCA) is used to reduce the dimensionality while retaining 98% of the variance information. In the dimensionality-reduced feature space, Gaussian Mixture Model (GMM) is used for unsupervised binary clustering. GMM can effectively model the probability structure of the feature distribution, and its density function is expressed as: Where πk is the mixing weight, μ k and ∑ kThese are the mean and covariance matrices of the k-th Gaussian component, respectively. GMM clustering generates two clusters, and each sample is assigned a corresponding pseudo-label. Due to the lack of prior knowledge, it's impossible to directly determine which cluster represents a real or fake sample. However, this label uncertainty is within the acceptable range of the framework, because the core function of the pseudo-label is to achieve binary partitioning of the data to provide a supervisory signal, rather than absolutely accurate category labeling. Furthermore, a discrimination mechanism based on the difference in hypersphere radius can effectively solve this problem. Although the clustering process introduces a certain proportion of label noise, the subsequent prototype contrastive learning has strong noise tolerance. It further utilizes momentum... The new category prototypes naturally correct misassignments during iterative optimization, prompting samples with similar intrinsic attributes to converge to the correct feature space. Through carefully designed visual artifact feature extraction and clustering strategies, high-quality pseudo-labels are generated, providing crucial supervision signals for subsequent prototype contrastive learning and decision boundary construction. Contrastive learning is performed using category prototypes updated with momentum to promote intra-class compactness and inter-class separability, enabling the encoder to learn more discriminative feature representations. The PGCL module possesses noise tolerance for erroneous pseudo-labels. Firstly, the category prototypes are formed through the aggregation of features from a large number of samples, so even with a small number of mislabeled samples, they will not significantly deviate from the prototype center, thus ensuring statistical stability. Secondly, during global optimization, a large number of correct samples dominate the feature learning direction, making the impact of a small number of mislabeled samples on the overall discriminative ability relatively small. In deepfake detection, traditional contrastive learning data augmentation methods are mainly designed for image classification tasks and cannot be directly applied to detection tasks based on subtle feature differences between real and fake images. Since real and fake samples are visually highly similar but have subtle differences in local features, an asymmetric data augmentation strategy specifically adapted for face forgery detection is designed. The core design principle of this strategy lies in achieving… Balancing semantic integrity and sample diversity, weak augmentation views preserve more original semantic information through gentle transformations to ensure stable updates of category prototypes; conversely, strong augmentation views enhance the model's generalization ability through aggressive transformations. Five targeted data augmentation methods are designed: first, random erasure simulates occlusion interference in real-world scenes by selectively occluding local facial regions; second, random cropping adapts to different field of view and imaging conditions through multi-scale cropping operations; third, horizontal flipping simulates changes in facial orientation under different shooting angles; and fourth, Gaussian noise enhances the model's robustness to image quality degradation by injecting random noise.Fifth, color jitter simulates diverse lighting conditions and shooting environments by adjusting brightness, contrast, saturation, and hue parameters. The strong enhancement strategy employs all five enhancement techniques, maximizing sample diversity through aggressive transformations to improve the model's adaptability to real-world interference. Correspondingly, the weak enhancement strategy employs four techniques, excluding random erasure, increasing sample diversity while maintaining the integrity of facial semantic information through gentle transformations. Utilizing the InfoNCE loss function as a classic contrastive learning method, it has achieved significant success in multiple computer vision tasks such as image classification and object detection. This method learns discriminative representations by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs. Its mathematical expression is: ; Where hi and hi+ represent the feature representations of the sample and its positive sample, respectively, hj- represents the negative sample feature, i is the temperature parameter, and N is the batch size. The InfoNCE loss function faces significant challenges in deepfake detection. In deepfake detection, the feature distributions of real and fake samples severely overlap in the high-dimensional feature space, resulting in weak distinguishing signals between positive and negative sample pairs. This hinders InfoNCE from effectively learning discriminative feature representations through instance-level comparative learning. To solve this problem, prototype-based contrastive learning is introduced, transforming the traditional inter-instance comparison into a comparison between an instance and a class prototype. By aggregating samples of the same category, a class prototype with stronger discriminative power is constructed. The positive and negative sample comparison signals in the contrastive learning process become clearer and more stable, effectively alleviating the difficulty of distinction caused by overlapping feature distributions. Considering the different functions of weak and strong augmented views in feature learning, a weighted fusion strategy is adopted to calculate the class prototype, given by the following formula: Pk=a·fkiweak+(1-a)·fkistrong, where μ is the momentum coefficient. This represents the weighted feature mean of category k in the current batch. This momentum update mechanism maintains the continuity of historical feature information while adaptively incorporating changes in the feature distribution of the current batch, thereby effectively improving the category representativeness of the prototype. Based on the dynamically updated category prototypes, a prototype comparison loss function is designed. Unlike the traditional InfoNCE method, it uses a simpler and more efficient cross-entropy loss to achieve prototype-level comparison. Its mathematical expression is: Where fi is the L2 normalized feature vector of the sample, Pci represents the prototype vector of the class to which sample i belongs, Pk represents the prototype vector of class k, and β is the temperature parameter. This loss function, through the softmax normalization mechanism, achieves the optimization goal of clustering sample features towards the correct class prototype and moving away from the incorrect class prototype. Through the above design, the PGCL module can effectively utilize the pseudo-label information generated in the first stage, and then optimize the discriminative ability of the feature representation through prototype-guided contrastive learning. The pseudo-labels in the first stage provide reliable supervision signals for its training, while the discriminative feature representations obtained by the PGCL module training in the second stage lay a solid foundation for the construction of the decision space. Based on this foundation, a Dual-DeepSVDD model with the same structure but independent parameters is constructed. Each model contains a feature adaptation layer and an orthogonal transformation layer. The feature adaptation layer maps the features of the pre-trained encoder to a compact representation space suitable for hypersphere modeling through batch normalization and ReLU activation. The orthogonal transformation layer uses an orthogonal initialization strategy to preserve the geometric properties of the features. By employing an alternating training mechanism, the parameters of the two models are updated sequentially in each epoch, avoiding model degradation problems that may be caused by joint training. This promotes the formation of mutually separate and compact representations of the two hyperspheres in the feature space. The hypersphere center is updated using a momentum-based strategy, with the following formula: Where c(t) represents the normalized hypersphere center after iteration t, and μ is the momentum coefficient. The center of the current batch of features is represented by the hypersphere. In deepfake detection, real samples originate from real physical imaging processes, thus their feature representations exhibit a relatively stable and concentrated distribution pattern in the deep feature space. Conversely, fake samples originate from various artificial synthesis algorithms. Due to the lack of unified physical imaging constraints and the involvement of multiple generation strategies and forgery techniques, their feature representations exhibit significant instability and dispersion in the deep feature space. Based on this difference in distribution characteristics, we determine the category assignment by comparing the hypersphere radii: the hypersphere with the smaller radius corresponds to real samples, and the hypersphere with the larger radius corresponds to fake samples (e.g., ...). Figure 6 As shown in the figure, this hypothesis is verified through t-SNE visualization results. Figure 6 The blue area (real samples) exhibits a relatively compact distribution, while the red area (fake samples) shows a more dispersed feature distribution pattern. During the inference phase, classification decisions are made by calculating the cosine distance from the test sample to the centers of the two hyperspheres. The calculation formula is as follows: Where φ(x) represents the L2-normalized feature vector, and c represents the center of the hypersphere. Samples are assigned to the category corresponding to the nearest hypersphere, thus achieving effective geometric decision-making in the high-dimensional feature space. Based on the above steps, texture features are extracted from key facial regions, and Gaussian Mixture Model (GMM) clustering is applied to generate reliable pseudo-labels for unlabeled data. These pseudo-labels provide key supervision signals for subsequent training stages. Then, the PGCL module is introduced, which generates two augmented views for each sample through weak and strong data augmentation. This module updates these augmented views with momentum-updated class prototypes. Through contrastive learning, samples are brought closer to their corresponding class prototypes while moving away from other class prototypes. This process promotes intra-class compactness and inter-class separability, enabling the encoder to learn highly discriminative feature representations. Finally, a Dual-DeepSVDD module is designed to construct the decision space. This module establishes two independent hyperspheres for real data and fake data, respectively. The module learns a compact feature distribution by dynamically updating the center and radius of the hyperspheres. During inference, samples are classified by calculating their distances to the centers of the two hyperspheres, thereby achieving accurate detection.

[0078] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A unified unsupervised depth forgery detection method based on prototype guidance and dual hyperspheres, characterized in that, Includes the following: S1: Extract the visual artifact features generated by deepfakes and use Gaussian mixture model clustering to generate pseudo-labels; S2: Comparative learning through class prototypes updated by momentum; S3: By constructing independent hyperspheres for real and fake samples respectively, the feature space and geometric decision are effectively integrated to build a dual-depth support vector data description framework; S4: Effectively integrates pseudo-label generation, feature learning, and decision boundary construction into the overall model architecture.

2. The unified unsupervised depth forgery detection method based on prototype guidance and dual hyperspheres as described in claim 1, characterized in that: Step S1 includes the following: S11: In the process of generating deepfake technology, subtle visual artifacts inevitably occur in specific facial areas. Although these artifacts are difficult to detect visually, they exhibit obvious statistical differences in the texture frequency domain. Reliable supervision signals are then generated through multi-region texture analysis and unsupervised clustering. S12: Deepfake algorithms exhibit significant differences in their processing capabilities across different facial regions, leading to regionalized artifact distribution. Based on the inherent limitations of facial geometry and forgery techniques, three artifact-sensitive regions were identified: eyes, mouth, and facial boundary regions. First, a dlib face detector was used to extract 68 facial key points. Based on these key points, precise geometric masks for each region were constructed. The eye region was generated using convex hull operations on the left and right eye key points, primarily capturing blink coordination and edge sharpness anomalies. The mouth region combined with clustering algorithms achieved precise segmentation of teeth and lips, focusing on detecting anomalies in tooth texture consistency. The facial boundary region was constructed using inner and outer boundary masks to identify trace features in face-swapping fusion. S13: Considering that visual artifacts mainly manifest as subtle perturbations in local texture patterns, a Laws texture filter is used for multi-scale feature extraction. The Laws filter can effectively respond to changes in texture directionality, roughness, and contrast, and is highly sensitive to texture inconsistencies introduced by depth forgery. Let the preprocessed image be I. prep The Laws filter is Texture features of each region are extracted using the mean of the filtered energy: in The symbol represents convolution, ⊙ represents element-wise multiplication, and M represents convolution. region This represents the mask for the corresponding region, and μ(·) represents the mean operation within the masked region. The final VAF feature vector is constructed by concatenating the texture features of the three regions:

3. The unified unsupervised depth forgery detection method based on prototype guidance and dual hyperspheres as described in claim 2, characterized in that: S14: Due to the high dimensionality of the extracted features and the scale differences between different regions, the visual artifact features are first standardized. Then, Principal Component Analysis (PCA) is used to reduce the dimensionality while retaining 98% of the variance information. In the dimensionality-reduced feature space, Gaussian Mixture Model (GMM) is used for unsupervised binary clustering. GMM can effectively model the probability structure of the feature distribution, and its density function is expressed as: Where π k For mixed weights, μ k and ∑ k These are the mean and covariance matrix of the k-th Gaussian component, respectively; S15: GMM clustering generates two clusters and assigns a corresponding pseudo-label to each sample. Due to the lack of prior knowledge, it is impossible to directly determine which cluster represents the real sample or the fake sample. However, this label uncertainty is within the acceptable range of the framework because the core function of the pseudo-label is to achieve binary partitioning of the data to provide a supervisory signal, rather than an absolutely accurate category label. In addition, the discriminant mechanism based on the difference in hypersphere radius during the inference stage can effectively solve this problem. Although the clustering process introduces a certain proportion of label noise, the subsequent prototype contrastive learning has a strong noise tolerance. It naturally corrects the misassignment in the iterative optimization through momentum-updated category prototypes, prompting samples with similar intrinsic properties to converge to the correct feature space. Through carefully designed visual artifact feature extraction and clustering strategies, high-quality pseudo-labels are generated for the unlabeled data, providing key supervisory signals for subsequent prototype contrastive learning and decision boundary construction.

4. The unified unsupervised depth forgery detection method based on prototype guidance and dual hyperspheres as described in claim 1, characterized in that: Step S2 includes the following: S21: Comparative learning through momentum-updated class prototypes promotes intra-class tightness and inter-class separation, thereby enabling the encoder to learn more discriminative feature representations; S22: The PGCL module has the ability to tolerate noise from erroneous pseudo-labels. Firstly, the category prototype is formed by aggregating features from a large number of samples. Therefore, even if there are a small number of erroneously labeled samples, they will not significantly deviate from the prototype center, thus ensuring statistical stability. Secondly, during the global optimization process, a large number of correct samples dominate the feature learning direction, making the impact of a small number of erroneously labeled samples on the overall discrimination ability relatively small. S23: In deepfake detection, traditional contrastive learning data augmentation methods are mainly designed for image classification tasks and cannot be directly applied to detection tasks based on subtle feature differences between real and fake images. Since real and fake samples are visually highly similar but have subtle differences in local features, an asymmetric data augmentation strategy specifically adapted to face forgery detection is designed. The core design principle of this strategy is to achieve a balance between semantic integrity and sample diversity. The weak augmentation view retains more original semantic information through gentle transformation to ensure stable updates of the category prototype; conversely, the strong augmentation view enhances the generalization ability of the model through aggressive transformation. S25: Five targeted data augmentation methods are designed: First, random erasure simulates occlusion interference in real-world scenes by selectively occluding local facial areas; second, random cropping adapts to different field of view and imaging conditions through multi-scale cropping operations; third, horizontal flipping simulates changes in facial orientation under different shooting angles; and fourth, Gaussian noise enhances the model's robustness to image quality degradation by injecting random noise. Fifth, color jitter simulates diverse lighting conditions and shooting environments by adjusting brightness, contrast, saturation, and hue parameters; S26: The strong enhancement strategy employs all five enhancement techniques to maximize sample diversity through aggressive transformations, thereby improving the model's adaptability to real-world interference. In contrast, the weak enhancement strategy employs four techniques, excluding random erasure, to increase sample diversity while maintaining the integrity of facial semantic information through gentle transformations.

5. The unified unsupervised depth forgery detection method based on prototype guidance and dual hyperspheres as described in claim 4, characterized in that: S27: Utilizing the InfoNCE loss function as a classic contrastive learning method, significant success has been achieved in multiple computer vision tasks such as image classification and object detection. This method learns discriminative representations by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs. Its mathematical expression is: Where h i and h i + represents the feature representation of the sample and its positive sample, respectively, h j - indicates negative sample features, i is the temperature parameter, and N is the batch size; S28: The InfoNCE loss function faces significant challenges in deepfake detection. In deepfake detection, the feature distributions of real and fake samples severely overlap in the high-dimensional feature space, resulting in weak discriminative signals between positive and negative sample pairs. This hinders InfoNCE from effectively learning discriminative feature representations through instance-level comparative learning. To address this issue, prototype-based contrastive learning is introduced, transforming traditional inter-instance comparisons into comparisons between instances and class prototypes. By aggregating samples of the same category, a class prototype with stronger discriminative power is constructed. The positive and negative sample comparison signals during the contrastive learning process become clearer and more stable, effectively alleviating the discrimination difficulties caused by overlapping feature distributions. Considering the different functions of weak and strong augmentation views in feature learning, a weighted fusion strategy is adopted to calculate the class prototype, given by the following formula: P k =a·f kiweak +(1-a)·f kistrong Where μ is the momentum coefficient, This represents the weighted average of the features of category k in the current batch. This momentum update mechanism maintains the continuity of historical feature information while adaptively incorporating the feature distribution changes of the current batch, thereby effectively improving the category representativeness of the prototype. S29: Based on dynamically updated category prototypes, a prototype contrast loss function is designed. Unlike the traditional InfoNCE method, it uses a simpler and more efficient cross-entropy loss to achieve prototype-level contrast. Its mathematical expression is: Where f i P is the L2 normalized feature vector of the sample. ci P represents the prototype vector of the category to which sample i belongs. k The prototype vector β of category k is a temperature parameter. This loss function achieves the optimization goal of clustering sample features toward the correct category prototype and moving away from the incorrect category prototype through the softmax normalization mechanism. Through the above design, the PGCL module can effectively utilize the pseudo-label information generated in the first stage, and then optimize the discriminative ability of the feature representation through prototype-guided contrastive learning.

6. The unified unsupervised depth forgery detection method based on prototype guidance and dual hyperspheres according to claim 1, characterized in that: Step S3 includes the following: S31: The pseudo-labels in the first stage provide reliable supervision signals for its training, while the discriminative feature representations obtained by the PGCL module training in the second stage lay a solid foundation for the construction of the decision space. S32: Next, a Dual-DeepSVDD model with the same structure but independent parameters is constructed. Each model contains a feature adaptation layer and an orthogonal transformation layer. The feature adaptation layer maps the features of the pre-trained encoder to a compact representation space suitable for hypersphere modeling through batch normalization and the ReLU activation function. The orthogonal transformation layer uses an orthogonal initialization strategy to preserve the geometric properties of the features. By employing an alternating training mechanism, the parameters of the two models are updated sequentially in each epoch, avoiding model degradation problems that may be caused by joint training. This promotes the formation of mutually separate and compact representations of the two hyperspheres in the feature space. The hypersphere center is updated using a momentum-based strategy, with the following formula: Where c (t)表示 The normalized hypersphere center after t iterations, where μ is the momentum coefficient. Indicates the center of the current batch feature; S33: In deepfake detection, real samples originate from real physical imaging processes, so their feature representations exhibit a relatively stable and concentrated distribution pattern in the depth feature space. Conversely, fake samples originate from the generation process of various artificial synthesis algorithms. Due to the lack of unified physical imaging constraints and the involvement of multiple generation strategies and forgery techniques, their feature representations exhibit significant instability and dispersion in the depth space. Based on this difference in distribution characteristics, we determine the category assignment by comparing the size of the hypersphere radius: the hypersphere with a smaller radius corresponds to real samples, and the hypersphere with a larger radius corresponds to fake samples (as shown in Figure 6). This hypothesis is verified by t-SNE visualization results.

7. The unified unsupervised depth forgery detection method based on prototype guidance and dual hyperspheres as described in claim 6, characterized in that: S34: In Figure 6, the blue area (real samples) shows a relatively compact distribution, while the red area (fake samples) exhibits a more dispersed feature distribution pattern. During the inference phase, classification decisions are made by calculating the cosine distance from the test sample to the centers of the two hyperspheres. The calculation formula is: S35: where φ(x) represents the L2 normalized feature vector, c represents the center of the hypersphere, and samples are assigned to the category corresponding to the closer hypersphere, thereby achieving effective geometric decision-making in the high-dimensional feature space.

8. The unified unsupervised depth forgery detection method based on prototype guidance and dual hyperspheres according to claim 1, characterized in that: Step S4 includes the following: S41: Based on the above steps, extract texture features from key facial regions and apply Gaussian mixture model to generate reliable pseudo-labels. These pseudo-labels provide key supervision signals for subsequent training stages. S42: Next, the PGCL module is introduced. This module generates two augmented views for each sample through weak data augmentation and strong data augmentation. This module learns by comparing these augmented views with the class prototypes updated with momentum. The samples are brought closer to the corresponding class prototypes while moving away from other class prototypes. Through this process, the PGCL module promotes intra-class compactness and inter-class separability, enabling the encoder to learn highly discriminative feature representations. S43: Finally, we further designed the Dual-DeepSVDD module to construct the decision space. This module establishes two independent hyperspheres for real data and fake data respectively. The module learns a compact feature distribution by dynamically updating the center and radius of the hypersphere. During inference, the sample is classified by calculating its distance from the center of the two hyperspheres, thereby achieving accurate detection.

Images