Depth forgery detection method and system capable of resisting depth image watermark interference

Abstract

The invention belongs to the technical field of depth forgery, and particularly discloses a depth image watermark interference-resistant depth forgery detection method and system, and the method comprises the steps: constructing a depth forgery detection model which comprises a feature discarding module and a classification module; the feature discarding module eliminates redundant features from the input image to obtain an image after the redundant features are eliminated; an efficient multi-scale attention mechanism is integrated in the classification module, the classification module receives the images after the redundant features are eliminated, deep counterfeiting detection is carried out on the images after the redundant features are eliminated through multi-scale feature extraction and the attention mechanism, and a detection result is output and obtained. According to the method, the problem that the performance is remarkably degraded when an existing deep counterfeiting detection model processes an image containing an unknown deep watermark is solved.

Description

Technical Field

[0001] This invention belongs to the field of deepfake technology, specifically relating to a deepfake detection method and system that resists interference from deep image watermarks. Background Technology

[0002] Currently, deepfake technology poses a significant challenge to the credibility of social media content due to its low cost, ease of use, and high-quality generation capabilities. Therefore, research into countermeasures against deepfake technology has become crucial, and identifying the authenticity of facial images on the internet remains a significant challenge. Given the rapid development of deepfake technology, current countermeasures can be broadly categorized into three types: active forensics, passive forensics, and active defense techniques. Active forensics involves preprocessing the image using an encoder. After the image is altered, these techniques employ a corresponding decoder to assess its authenticity. Active defense methods focus on introducing adversarial perturbations into the image to ensure that it does not convey misleading information when altered. Passive forensics utilizes the powerful feature extraction capabilities of deep neural networks to learn the characteristics of forged images, thereby detecting forged content. In the current field of deepfake detection, passive forensics does not require image preprocessing and can directly detect altered content. Due to its high adaptability and wide applicability, passive forensics has become a primary method for identifying potentially forged content. In contrast, proactive forensics employs deep learning-based digital watermarking, embedding a specific watermark into an image before it is distributed to protect specific objects. However, previous research has often discussed these two techniques independently, neglecting the potential impact of deep watermarking in proactive forensics on the performance of passive detection algorithms. To address this issue, some researchers have proposed optimization modules to mitigate the negative impact of proactive forensics on passive forensics; however, their implementation requires retraining already trained and potentially deployed models. Summary of the Invention

[0003] The purpose of this invention is to address the problem that existing deepfake detection models suffer significant performance degradation when processing images with unknown depth watermarks, and to propose a deepfake detection method and system that resists interference from depth image watermarks.

[0004] The technical solution of the present invention is as follows: Firstly, a visualization method for evaluating power communication networks based on multi-source data fusion, comprising the following steps:

[0005] Build a deepfake detection model;

[0006] The image data to be detected is input into the deepfake detection model for deepfake detection, and the detection result is output.

[0007] Preferably, the deepfake detection model includes a feature discarding module and a classification module;

[0008] The feature discarding module is used to remove redundant features from the input image to obtain an image after removing redundant features;

[0009] The classification module receives the image after removing redundant features, and performs deepfake detection on the image after removing redundant features through multi-scale feature extraction and attention mechanism, and outputs the detection result.

[0010] Preferably, the feature discarding module includes a first encoding unit, a max pooling unit, a second encoding unit, a transposed convolution unit, an element-wise addition unit, a decoding unit, a final convolution unit, and a Sigmoid activation unit connected in sequence.

[0011] The other output of the first encoding unit is connected to the other input of the element-by-element addition unit.

[0012] Preferably, the first coding unit and the second coding unit have the same structure; the first coding unit is composed of a first convolutional layer, a first ReLU layer, a second convolutional layer and a second ReLU layer connected in sequence.

[0013] Preferably, the classification module is an EMA-Xception network obtained by introducing an efficient multi-scale attention mechanism into the Xception network; the EMA-Xception network includes a first entry flow unit, a second entry flow unit, a first efficient multi-scale attention unit, an eight-layer intermediate flow unit, a Conv-BN-ReLU unit, a second efficient multi-scale attention unit, and a Conv-BN unit connected in sequence.

[0014] The Conv-BN-ReLU unit comprises a third convolutional layer, a first batch normalization layer, and a third ReLU layer connected in sequence.

[0015] The Conv-BN unit consists of a fourth convolutional layer and a second batch normalization layer connected in sequence.

[0016] Preferably, the construction of the deepfake detection model specifically includes the following steps:

[0017] Obtain the raw image data;

[0018] A depth image watermarking pool module is constructed. The original image data is input into the depth image watermarking pool module to add watermarks, and the output is an image watermark dataset containing depth image watermarks.

[0019] The deepfake detection model is trained in two stages using an image watermark dataset containing depth image watermarks, resulting in a fully trained deepfake detection model.

[0020] Preferably, the depth image watermark pool module adds a watermark to the image according to a preset probability p, specifically using the following formula:

[0021]

[0022] Where x′ represents the image containing the depth image watermark, x represents the original image data, and w e This represents the watermark embedding tensor, and rand(·) represents the random generation function.

[0023] As a preferred approach, the deepfake detection model is trained in two stages using an image watermark dataset to obtain the trained deepfake detection model, specifically as follows:

[0024] The classification module was trained using the original image data and optimized using the cross-entropy loss function to obtain the parameters of the classification module.

[0025] The parameters of the classification module are frozen, and the feature discarding module is adversarially trained using an image watermark dataset containing depth image watermarks, thereby obtaining a trained deepfake detection model.

[0026] The beneficial effects of this invention are:

[0027] 1. The feature discarding module of this invention adopts an encoder-decoder structure to eliminate redundant features from the original image, thereby reducing the interference of redundant features on the decision-making process of the classification model. A skip connection is added between the encoder and the decoder, that is, the output of the first encoding unit and the output of the transposed convolution unit are input to the decoding unit through the element-wise addition unit, which can retain the shallow features in the input image and prevent the loss of key information related to forgery.

[0028] 2. This invention integrates an efficient multi-scale attention mechanism into the Xception network. This mechanism captures multi-scale contextual features by aggregating information across different spatial resolutions, thereby enhancing the network's ability to model hierarchical visual patterns in the input phase (entry flow) and optimize global-local feature interactions in the output phase (exit flow). This improves the multi-scale feature extraction capability of the classification module, fuses channel and spatial information, and strengthens the attention mechanism.

[0029] In a second aspect, a deepfake detection system resistant to depth image watermark interference includes a processor for executing the deepfake detection method resistant to depth image watermark interference as described in the first aspect.

[0030] Thirdly, a computer-readable storage medium stores computer instructions, wherein, in response to a computer reading the computer instructions in the storage medium, the computer executes the depth forgery detection method against depth image watermark interference as described in the first aspect. Attached Figure Description

[0031] Figure 1 The diagram shows a flowchart of a depth forgery detection method that resists depth image watermark interference, as provided in Embodiment 1 of the present invention.

[0032] Figure 2 The diagram shown is a structural diagram of the feature discarding module provided in Embodiment 1 of the present invention.

[0033] Figure 3 The diagram shown is a structural diagram of the classification module provided in Embodiment 1 of the present invention. Detailed Implementation

[0034] Exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the drawings are merely exemplary and are intended to illustrate the principles and spirit of the invention, and are not intended to limit the scope of the invention.

[0035] Example 1:

[0036] like Figure 1 As shown, a depth forgery detection method resistant to depth image watermark interference includes the following steps:

[0037] S1. Construct a deepfake detection model;

[0038] S2. Input the image data to be detected into the deepfake detection model to perform deepfake detection, and output the detection result.

[0039] In this embodiment, let ε and D represent the encoder and decoder of the deep watermarking model, respectively, x represent the input image, w represent the embedded information, and F represent the image similarity loss function. In the HiDDeN model, the image similarity loss function F usually refers to the l2 loss used during training and optimization. Other models use different loss functions, such as mean squared error (MSE), structural similarity index (SSIM), or peak signal-to-noise ratio (PSNR) loss. To enhance the realism of the watermarked image, the discriminator evaluates the encoder's output, and its loss function is usually the cross-entropy loss function. B represents the bit error rate (BER) for watermark extraction. Although the specific details of different algorithms vary, the global loss function of such models can be expressed by the following formula:

[0040]

[0041] Among them, E (x,w)Let Y represent the expected value of the results of the operations with respect to the encoder ε, decoder D, and bit error rate under the joint distribution of the input image x and the embedded information w. Y represents the image with embedded watermark information. In the image with embedded deep watermark, the watermark information is inserted into the feature layer of the image, and then reconstructed into a watermarked image x through a deconvolution process. w Therefore, the embedded watermark can be approximated as w. e =x w -x. Deep image watermarking typically doesn't produce a noticeable visual difference between the original and watermarked images. However, even minor modifications can affect the representation of features in a convolutional neural network. Therefore, traditional deep forgery detection models may extract misleading features during representation learning, leading to prediction errors and degraded detection performance.

[0042] In this embodiment, the deepfake detection model includes a feature discarding module and a classification module;

[0043] The feature discarding module is used to remove redundant features from the input image to obtain an image after removing redundant features;

[0044] The classification module receives the image after removing redundant features, and performs deepfake detection on the image after removing redundant features through multi-scale feature extraction and attention mechanism, and outputs the detection result.

[0045] In this embodiment, as Figure 2 As shown, the feature discarding module adopts an encoder-decoder structure, following the U-Net architecture. Skip connections are added between the encoder and decoder to preserve shallow features in the input image. Although fully convolutional networks and autoencoders are widely used, deep forgery detection tasks require the preservation of shallow features that often contain forgery traces. Traditional convolutional networks often discard these shallow features during processing; therefore, skip connections help preserve them and prevent the loss of crucial information related to forgery. The feature discarding module includes a first encoding unit, a max-pooling unit, a second encoding unit, a transposed convolutional unit, an element-wise addition unit, a decoding unit, a final convolutional unit, and a sigmoid activation unit connected in sequence.

[0046] Another output of the first encoding unit is connected to another input of the element-wise addition unit. This skip connection allows shallow features to be preserved during feature suppression. The feature discarding module maintains the spatial dimension of the input image while processing it, ensuring that the size of the output image is consistent with the input.

[0047] In this embodiment, the first coding unit and the second coding unit have the same structure; the first coding unit is composed of a first convolutional layer, a first ReLU layer, a second convolutional layer and a second ReLU layer connected in sequence.

[0048] In this embodiment, as Figure 3 As shown, the classification module is the EMA-Xception network, which incorporates an efficient multi-scale attention (EMA) mechanism into the Xception network. The Xception network exhibits strong performance in visual tasks such as image classification and semantic segmentation. It originates from the Inception architecture and employs depthwise separable convolutions, cross-channel activation, and global average pooling for feature aggregation. By replacing the standard Inception convolutions with depthwise separable convolutions, Xception introduces a novel modular structure that combines higher efficiency and computational power. This design significantly reduces the number of parameters and model complexity, while improving accuracy and generalization in visual recognition tasks. However, the feature discarding module compresses the image through convolution operations and then reconstructs it using deconvolution, a process that removes watermark features. Since convolutional compression is irreversible, some key features used to determine the authenticity of the image are also lost. Therefore, it is necessary to improve the multi-scale feature extraction capability of the Xception backbone network, fuse channel and spatial information, and enhance the attention mechanism. Furthermore, since the task of this invention is to detect data with unknown distributions, the model's generalization ability is crucial. To meet this need, this invention integrates an efficient multi-scale attention (EMA) mechanism into the Xception network, thereby improving its applicability to this task.

[0049] The EMA-Xception network comprises a first entry flow unit, a second entry flow unit, a first efficient multi-scale attention unit, an eight-layer middle flow unit, a Conv-BN-ReLU unit, a second efficient multi-scale attention unit, and a Conv-BN unit, connected in sequence.

[0050] The Conv-BN-ReLU unit comprises a third convolutional (Conv) layer, a first batch normalization (BN) layer, and a third ReLU layer connected in sequence;

[0051] The Conv-BN unit, or Exit Flow unit, consists of a fourth convolutional layer and a second batch normalization layer connected in sequence.

[0052] In this embodiment, the construction of the deepfake detection model specifically includes the following steps:

[0053] Obtain the raw image data;

[0054] A depth image watermarking pool module is constructed. The original image data is input into the depth image watermarking pool module to add watermarks, and the output is an image watermark dataset containing depth image watermarks.

[0055] The deepfake detection model is trained in two stages using an image watermark dataset containing depth image watermarks, resulting in a fully trained deepfake detection model.

[0056] In this embodiment, the current deepfake detection dataset only contains original or forged images and does not consider the case where the images contain watermarks. However, since the model of this invention needs to be evaluated on watermarked data, environmental simulation becomes crucial. To achieve this, this embodiment selects two models: the general deep image watermarking model MBRS and the active forensics method FaceSigns for deepfake detection. The watermark embedded by FaceSigns remains intact even after processing by the deepfake generation algorithm. After training these two models, they are integrated into a unified deep image watermarking pool to obtain the deep image watermarking pool module. The deep image watermarking pool module adds watermarks to the input images in a probabilistic manner. During the watermark embedding process, the specific model used for embedding can be manually selected. In addition, the deep image watermarking pool module allows embedding different types of watermarks on different samples in the dataset to improve the fidelity of the simulation. The deep image watermarking pool module adds watermarks to the images according to a preset probability p, where the preset probability p is set to a value of 0-1 during module operation. The preset probability p determines that p×100% of the images in the dataset will be watermarked, and the specific formula is as follows:

[0057]

[0058] Where x′ represents the image containing the depth image watermark, x represents the original image data, and w e This represents the watermark embedding tensor, and rand(·) represents the random generation function.

[0059] In this embodiment, the deepfake detection model is trained in two stages using an image watermark dataset to obtain the trained deepfake detection model, specifically as follows:

[0060] The classification module was trained using the original image data and optimized using the cross-entropy loss function to obtain the parameters of the classification module.

[0061] The parameters of the classification module are frozen, and the feature discarding module is adversarially trained using an image watermark dataset containing depth image watermarks, thereby obtaining a trained deepfake detection model.

[0062] Example 2:

[0063] Based on Example 1, this embodiment of the invention uses the widely recognized datasets UADFV, CelebDF-v1, and CelebDF-v2 in the field of deepfake detection to verify and illustrate the deepfake detection method against depth image watermark interference proposed in this invention.

[0064] Although the datasets UADFV, CelebDF-v1, and CelebDF-v2 are publicly available, they still require preprocessing, including frame extraction, face cropping, and alignment. Therefore, this embodiment of the invention selected a preprocessed dataset from DeepfakeBench—this dataset randomly selects 32 frames from each video, aligns faces through facial key point detection, crops them to 256×256 resolution, and provides structured labels. Using the DeepfakeBench dataset as a baseline, this embodiment of the invention created four watermark datasets using a depth image watermarking pool: 50% or 100% image embedding MBRS watermarking, and 50% or 100% image embedding FaceSigns watermarking.

[0065] Baseline Method: Xception is a general image classification network derived from Inception, which improves efficiency and accuracy through stacked depthwise separable convolutions. It utilizes SRM filters to extract frequency and spatial features and fuses them through a cross-attention mechanism to enhance the model's generalization ability in forgery detection. UCF image information is decoupled into three parts: forgery-independent features, method-specific forgery cues, and general forgery patterns. Images are reconstructed based on these forgery-related features, and contrastive regularization is applied to distinguish different types of forgeries. CORE generates augmented views of the input image, extracts features using a shared encoder, enforces similarity representation through cosine distance, and classifies each view using supervised labels.

[0066] Implementation details of the method proposed in this invention: The Middle Flow of the Xception architecture integrates depthwise separable convolutions and residual connections, with these combined units arranged in a multi-layered stacked manner. This design reduces the number of model parameters and computational burden while maintaining efficient feature extraction capabilities. Based on the Xception network, this embodiment of the invention introduces the EMA attention mechanism only in the Entry Flow and Exit Flow, while stacking eight convolutional-residual units in the Middle Flow. The model training process is divided into two stages: the first stage trains the classifier on the original dataset images, using the cross-entropy loss function for optimization. The second stage loads the trained weights into the classifier and freezes its parameters, performing adversarial training on the feature dropout module. Each batch contains two classes of images: real images from the dataset, and images processed by the frozen HiDDeN encoder and feature dropout module. The classifier optimizes the feature dropout module by comparing these images. The entire training process uses only cross-entropy loss to maintain computational simplicity and employs an early stopping strategy to prevent overfitting.

[0067] In this embodiment, the impact of deep image watermarking on detection models is evaluated using three backbone networks commonly used in deepfake detection (Xception, EfficientNet, and ResNet). All models were trained on the Celeb-V2 dataset. To investigate the factors leading to performance degradation, Gradient Weighted Class Activation Mapping (Grad-CAM) was used to analyze images correctly classified by the Xception model on the original dataset but misclassified after the FaceSigns watermark was introduced. A heatmap generated by the last convolutional layer of Xception was used to visualize the model's region of interest. The results show that the model's focus differs significantly between the original and watermarked images, ultimately leading to classification errors.

[0068] In this embodiment, comparative experiments are conducted on three mainstream datasets to comprehensively evaluate the effectiveness of the proposed method for image detection with depth watermarks. All models are trained using an early stopping strategy to prevent overfitting from affecting the results.

[0069] Experimental Results on the UADFV Dataset: Experiments were conducted on the UADFV dataset, and the performance of our proposed method was compared with the baseline models described in the baseline section. As shown in Table 1, on the original dataset and the dataset with embedded MBRS watermarks, the performance of our proposed model is slightly inferior to the baseline models; however, on the dataset with FaceSigns watermarks, our method outperforms all baseline models. Furthermore, the stable performance of the model on different datasets highlights its robustness to depth image watermarking interference.

[0070] Table 1. Performance comparison on the UADFV dataset (ACC / AUC)

[0071]

[0072]

[0073] Results on the Celeb-DF-V1 dataset: To evaluate the performance of the proposed method on the Celeb-DF-V1 dataset, experiments were conducted in this embodiment of the invention, and the detection results are shown in Table 2. The results show that the proposed method outperforms other methods in almost all scenarios, clearly demonstrating its superior performance on Celeb-DF-V1.

[0074] Table 2 Performance comparison on the Celeb-DF-V1 dataset (ACC / AUC)

[0075]

[0076] Results on the Celeb-DF-V2 dataset: Among the datasets used, Celeb-DF-V2 has the largest data volume, containing 176,177,207,951 images. As shown in Table 3, the overall performance of the method proposed in this invention is superior to the comparison methods. These results demonstrate that even on large-scale datasets, the method proposed in this invention maintains strong detection performance.

[0077] Table 3 Performance comparison on the Celeb-DF-V2 dataset (ACC / AUC)

[0078]

[0079]

[0080] The results show that adding a depth image watermark to the original image leads to a decrease in detector performance. Existing depth forgery detection models exhibit significant performance degradation when processing images with unknown depth watermarks. To alleviate this problem, this invention proposes a depth forgery detection method resistant to depth image watermark interference. This method can effectively combat watermark interference and does not require watermark data for training.

[0081] Example 3:

[0082] Based on Embodiment 1, this embodiment of the invention provides a depth forgery detection system resistant to depth image watermark interference, which is configured to execute a depth forgery detection method resistant to depth image watermark interference in Embodiment 1.

[0083] In this embodiment, the system may be an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor. The processor executes the program to implement some or all of the steps of the depth forgery detection method against depth image watermark interference as described in Embodiment 1.

[0084] In this embodiment, the electronic device may include: a processor, a memory, a bus, and a communication interface. The processor, the communication interface, and the memory are connected via the bus. The memory stores a computer program that can run on the processor. When the processor runs the computer program, it executes some or all of the steps of the depth forgery detection method against depth image watermark interference provided in Embodiment 1 of this application.

[0085] The system in this embodiment of the invention may also be a computer-readable storage medium storing a computer program that, when executed, implements some or all of the steps of the depth forgery detection method against depth image watermark interference as described in Embodiment 1.

[0086] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0087] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0088] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0089] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with embodiments of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0090] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0091] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for detecting depth forgery that resists interference from depth image watermarks, characterized in that, Includes the following steps: Build a deepfake detection model; The image data to be detected is input into the deepfake detection model for deepfake detection, and the detection result is output.

2. The depth forgery detection method against depth image watermark interference according to claim 1, characterized in that, The deepfake detection model includes a feature discarding module and a classification module; The feature discarding module is used to remove redundant features from the input image to obtain an image after removing redundant features; The classification module receives the image after removing redundant features, and performs deepfake detection on the image after removing redundant features through multi-scale feature extraction and attention mechanism, and outputs the detection result.

3. The depth forgery detection method against depth image watermark interference according to claim 2, characterized in that, The feature discarding module includes a first encoding unit, a max pooling unit, a second encoding unit, a transposed convolution unit, an element-wise addition unit, a decoding unit, a final convolution unit, and a Sigmoid activation unit connected in sequence. The other output of the first encoding unit is connected to the other input of the element-by-element addition unit.

4. The depth forgery detection method against depth image watermark interference according to claim 3, characterized in that, The first coding unit and the second coding unit have the same structure; the first coding unit consists of a first convolutional layer, a first ReLU layer, a second convolutional layer and a second ReLU layer connected in sequence.

5. The depth forgery detection method against depth image watermark interference according to claim 2, characterized in that, The classification module is an EMA-Xception network obtained by introducing an efficient multi-scale attention mechanism into the Xception network; the EMA-Xception network includes a first entry flow unit, a second entry flow unit, a first efficient multi-scale attention unit, an eight-layer intermediate flow unit, a Conv-BN-ReLU unit, a second efficient multi-scale attention unit, and a Conv-BN unit connected in sequence. The Conv-BN-ReLU unit comprises a third convolutional layer, a first batch normalization layer, and a third ReLU layer connected in sequence. The Conv-BN unit consists of a fourth convolutional layer and a second batch normalization layer connected in sequence.

6. The depth forgery detection method against depth image watermark interference according to claim 2, characterized in that, The construction of the deepfake detection model specifically includes the following steps: Obtain the raw image data; A depth image watermarking pool module is constructed. The original image data is input into the depth image watermarking pool module to add watermarks, and the output is an image watermark dataset containing depth image watermarks. The deepfake detection model is trained in two stages using an image watermark dataset containing depth image watermarks, resulting in a fully trained deepfake detection model.

7. The depth forgery detection method against depth image watermark interference according to claim 6, characterized in that, The depth image watermark pool module adds a watermark to the image according to a preset probability p, and the specific formula is as follows: Where x′ represents the image containing the depth image watermark, x represents the original image data, and w e This represents the watermark embedding tensor, and rand(·) represents the random generation function.

8. The depth forgery detection method against depth image watermark interference according to claim 6, characterized in that, The deepfake detection model is trained in two stages using an image watermark dataset to obtain the trained deepfake detection model, specifically: The classification module was trained using the original image data and optimized using the cross-entropy loss function to obtain the parameters of the classification module. The parameters of the classification module are frozen, and the feature discarding module is adversarially trained using an image watermark dataset containing depth image watermarks, thereby obtaining a trained deepfake detection model.

9. A depth forgery detection system resistant to depth image watermark interference, characterized in that, Includes a processor for performing the depth forgery detection method against depth image watermark interference as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The storage medium stores computer instructions. In response to the computer reading the computer instructions in the storage medium, the computer executes the depth forgery detection method against depth image watermark interference as described in any one of claims 1-8.

Images