Generalized fake media detection method for real large-flow scene

Abstract

The invention discloses a generalized fake media detection method oriented to a real large-flow scene. The method comprises the following steps: step 1, distributing data requested by a network to a plurality of servers through a weighted polling algorithm; step 2, the access traffic of each server is cached in an access queue, data of a queue head is taken out each time and subjected to standardization processing to obtain a detection packet, and the standardization processing comprises framing, clipping, segmentation and packaging; 3, the detection packets are cached in the detection queue, and the detection packet at the queue head is taken out every time to carry out forgery detection, according to the method, a multi-model joint decision algorithm is introduced, and by integrating multiple different detection models, the coping capacity for multiple forgery technologies is improved, and the generalization performance and robustness of the algorithm are enhanced. And the detection efficiency of the system is improved through a modular process and streaming forgery detection, and efficient feedback can be realized even on the basis of multiple models.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a generalized forged media detection method for realistic high-traffic scenarios. Background Art

[0002] In the current Internet environment, the speed of propagation of forged images and videos increases rapidly with the large traffic of network platforms, especially in high-concurrency scenarios such as social media and news websites. The spread range and speed of false information often exceed expectations, and the public is easily affected by these forged contents. Therefore, providing online forged image / video detection to the public is not only an inevitable choice to deal with the rapid spread of forged content, but also an important means to maintain public trust and ensure the security of the network environment. However, online forgery detection needs to deal with a large number of access requests, and traditional detection systems are prone to efficiency bottlenecks and resource overload problems when processing large-scale data. In order to ensure that the detection system can still maintain efficient and stable performance when a large amount of data streams are pouring in, the present invention introduces a load balancing mechanism to reasonably allocate computing resources and avoid excessive pressure on a single point, thereby ensuring that the system can quickly process massive data.

[0003] In the context of such high traffic volume, the limitations of a single detection model become more pronounced, especially when counterfeiting methods become increasingly diverse and complex, the difficulty of detection increases significantly.

[0004] Existing generative model detection methods can be divided into three categories: biometric-based, statistical-based, and deep learning-based methods.

[0005] Biometric-based methods use the slight differences in physiological features between generated images and real images for detection. This type of method focuses on capturing biological features that naturally exist in real images, such as the regularity and consistency of pupil shape, the consistency of eye highlight reflections, or judging the authenticity of images by detecting the natural appearance of facial texture and detail features. These biometric features are often interfered with by counterfeiting technology during the generation process, resulting in unnatural appearance, which becomes the basis for detection.

[0006] Forged image detection methods based on statistical features identify forgery traces by analyzing the statistical properties of images. For example, the grayscale co-occurrence matrix and cross-color channel co-occurrence matrix of images are often used for texture analysis and as effective features for forgery detection. There is also frequency domain feature extraction based on discrete cosine transform (DCT), which identifies forgery patterns in images by converting images to the frequency domain. By combining machine learning algorithms such as support vector machines (SVM), these statistical features can effectively identify forged images.

[0007] The deep learning-based method automatically extracts deep features from images by training neural networks. It does not rely on traditional hand-designed features, is more adaptable, and can effectively deal with complex generation patterns. The core advantage of this type of method is that it can learn complex features of images through large-scale data and has stronger discrimination capabilities. Network models such as VGGNet, ResNet, and Xception are often used to detect forged images or videos.

[0008] However, current counterfeit detection technology faces many challenges when faced with large amounts of traffic data. With the rapid development of generation technology, detection systems need to process large amounts of data in a short period of time, which places high demands on the real-time and efficiency of algorithms. In high-traffic environments, traditional detection methods may not be able to respond in a timely manner, resulting in potential counterfeit content not being identified in a timely manner. In addition, the system may experience performance degradation under high load, which in turn affects the overall detection accuracy.

[0009] In addition, the limitations of single-model detection methods cannot be ignored. Although single models perform well in certain specific tasks, their generalization ability and robustness are often insufficient, making it difficult to adapt to diverse generated images and videos. When faced with new generation technologies, a single model may not be able to fully capture the characteristics of forged images, resulting in missed detections or false detections. In addition, the vulnerability of a single model in dealing with data noise and adversarial attacks also limits its effectiveness in practical applications. Therefore, single-model detection methods need to be combined with other technical means to improve the accuracy and reliability of their detection. Summary of the invention

[0010] In view of the shortcomings of the existing technology, the present invention proposes a generalized forged media detection method for realistic high-traffic scenarios, introduces a multi-model joint decision-making algorithm, and improves the ability to cope with a variety of forgery technologies by integrating multiple different detection models, enhancing the generalization performance and robustness of the algorithm. The detection efficiency of the system is improved through modular processes and streaming forgery detection, and efficient feedback can be achieved even based on multiple models.

[0011] In order to solve the above technical problems, the technical solution of the present invention is:

[0012] A generalized forged media detection method for realistic high-traffic scenarios includes the following steps:

[0013] Step 1: Distribute the data requested by the network to multiple servers through a weighted round-robin algorithm;

[0014] Step 2: Cache the access traffic of each server in the access queue, and take out the data at the head of the queue each time to perform standardization processing to obtain a detection packet, wherein the standardization processing includes framing, clipping and segmentation, and encapsulation;

[0015] Step 3: Cache the detection packets in the detection queue, and take out the detection packet at the head of the queue for forgery detection each time.

[0016] The method of forgery detection is:

[0017] First, we build a multimodal forgery detection framework consisting of three expert models: Xception (model A), ResNet-50 (model B), and ConvNeXt (model C);

[0018] Then, a dataset with real images and corresponding generated forged images is constructed, and the dataset is used to train and verify the multimodal forgery detection network model;

[0019] The detection package to be detected is extracted, and the detection package is decapsulated to obtain several detection clusters. Each detection cluster is decapsulated to obtain a detection unit, and the detection unit outputs the detection result through a multimodal forgery detection network model.

[0020] Preferably, the data requested by the network includes images and videos.

[0021] Preferably, in step 1, the method for allocating network requests is as follows:

[0022] Quantify the performance indicators of each server and select a The image is used as the reference. The inference time of all servers for the image is obtained by performing inference in the servers. And memory usage ratio , the performance quantitative index of each server is obtained by the following formula:

[0023] ;

[0024] in, ;

[0025] The performance required for network request data is quantified based on the size of the image or video frame and the number of key frames. The quantization expression of an image or video frame is as follows:

[0026] ;

[0027] in is the number of key frames, for images , for videos ;

[0028] Weighted polling based on performance quantification of server and network request data.

[0029] Preferably, the weighted polling method is: first initialize Value , which are equal to , select servers in order , process the requests in sequence, and the performance requirement for each processing is The image or video corresponding to minus ,like If the number of servers is insufficient, the next server will be selected. If insufficient, Reset to initial value.

[0030] As a preferred method, the video segmentation method is: for a video, use the video codec tool FFmpeg to extract all key frames in the video ( frame), get the frame set ,in is the cardinality of the frame set, i.e. the number of frames, and ; For images, there is no need to perform frame processing, and the frame set only contains the image, that is, ;

[0031] The cropping and segmentation method is: for each frame in the frame set, performing face cropping and image segmentation operations respectively;

[0032] The encapsulation method is as follows: after cropping and segmentation, the image blocks required by each model are encapsulated into corresponding detection units, and an identifier of a designated detection model is added to realize directional sending of the detection units in a routing module in the detection phase.

[0033] Preferably, in step 3, three data sets including real images and corresponding forged type generated images are constructed respectively, wherein the real images are labeled as 0 and the forged images are labeled as 1, 80% of the images are used as training sets, 10% of the images are used as validation sets, and 10% of the images are used as test sets.

[0034] Preferably, the training method of the multimodal forgery detection framework is:

[0035] The three training sets are used as input to train Xception (model A), ResNet-50 (model B) and ConvNeXt (model C). Each model predicts the input image and finally outputs a confidence score. , represents the probability that this image is a forged image, and the cross entropy function used in training is:

[0036] ;

[0037] in The label representing the image, i.e. the real image , fake image , the training goal of this loss is to make As small as possible until convergence;

[0038] The weights of the confidence of each model output are trained. First, the weights corresponding to the three models are defined as , the images in the three validation sets are combined, and each image can be predicted by three trained models to obtain three confidence levels Finally, several triplets with real or fake labels are obtained as the data set for weight training. First, the triple confidence is aggregated, and then the three weights are continuously trained and optimized through the loss function. , the expression is as follows:

[0039] ;

[0040] ;

[0041] in Represents the label of the image corresponding to the confidence triplet, that is, the real image , fake image , the training goal of this loss is to ensure Under the premise of As small as possible until convergence. Represents the confidence after aggregation.

[0042] Preferably, in step 3, the three decapsulated detection units use a routing module to send them to the corresponding model for detection according to the identifier on the detection unit.

[0043] Preferably, in step 3, the confidences of the three detection units in the detection cluster output by the corresponding model are aggregated through the trained weights to obtain the confidence of each detection cluster. , set the threshold ,when When , the image is predicted to be a real image, When , the image is predicted to be a forged image, and the predicted label is recorded as .

[0044] Preferably, the threshold The method of determining is: all Sort from small to large and select them as thresholds in turn, and get the predicted label according to the authenticity judgment rules mentioned above. , and by comparing and Get the prediction accuracy of the threshold, and finally Select the threshold value that corresponds to the maximum prediction accuracy.

[0045] Preferably, in step 3, for video detection, a detection package usually contains multiple detection clusters, corresponding to multiple cluster confidences, so by combining sensitive parameters Together to judge the authenticity of , when the proportion of forged detection clusters to all detection clusters is greater than or equal to When the video is judged as fake, sensitive parameters The smaller the video is, the easier it is to identify it as fake. Videos with detection clusters, where we have obtained greater than or equal If a fake detection cluster is found, the detection will not continue and the detection result will be directly returned as fake.

[0046] In the present invention, a load balancing algorithm for both image and video formats is designed. The performance requirements of data in different formats are uniformly quantified, and weighted polling is performed according to the performance of the server to achieve load balancing.

[0047] A standardized processing flow was designed to encapsulate images and videos into a unified format, which simplified the data processing process and ensured compatibility and consistency between different input data types.

[0048] A multi-model joint decision-making mechanism was designed to improve the system's generalization ability and detection accuracy by integrating the judgment results of multiple models for images and videos generated by various forgery methods.

[0049] Through load balancing, key frame extraction, image segmentation in the standardization process, and streaming authenticity judgment in the multi-model joint decision-making process, the present invention significantly improves the reasoning efficiency and realizes efficient detection feedback.

[0050] The present invention has the following characteristics and beneficial effects:

[0051] Using the above technical solutions, this method is first based on a load balancing strategy, which is designed for large-volume forged image and video detection. By dynamically allocating computing resources and tasks, we can effectively process large amounts of data streams, ensuring that the system maintains efficient detection performance under high load conditions. On this basis, a multi-model joint decision-making framework is adopted to combine the outputs of multiple detection models to improve the recognition accuracy and robustness of forged content. Different models have their own advantages in dealing with diverse generation technologies, allowing the system to more comprehensively capture the characteristics of forged content. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0053] Figure 1 The figure is a schematic diagram of a method flow of an embodiment of the present invention.

[0054] Figure 2 Schematic diagram of data standardization process in an embodiment of the present invention.

[0055] Figure 3 Schematic diagram of the multi-model joint detection process in an embodiment of the present invention. DETAILED DESCRIPTION

[0056] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0057] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and the like are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, features defined as "first", "second", and the like may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0058] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood by specific circumstances.

[0059] Example 1

[0060] The present invention provides a generalized forged media detection method for realistic high-traffic scenarios, such as Figure 1 As shown in the figure, the overall framework consists of three parts: load balancing, data standardization, and multi-model joint decision-making.

[0061] Load balancing: In order to disperse traffic pressure, improve service performance and ensure high availability of the system, the present invention distributes network requests to multiple servers through a weighted round-robin algorithm. The servers are assigned weights according to their processing capabilities, and the number of requests received by each server is proportional to its weight.

[0062] Specifically, the load balancing implementation method is as follows:

[0063] Quantify the performance indicators of each server, and use the server's memory and inference time as the weight. The image is used as the reference. The inference time of all servers for the image is obtained by performing inference in the servers. And memory usage ratio , the performance quantitative index of each server is obtained by the following formula:

[0064] ;

[0065] in, ;

[0066] Request performance quantification: Due to the different performance requirements for images and videos, the performance required for network request data is quantified based on the size of the image or video frame and the number of key frames. The quantization expression of an image or video frame is as follows:

[0067] ;

[0068] in is the number of key frames, for images , for videos ;

[0069] Weighted round-robin: Based on the quantified results of the first two steps, weighted round-robin is performed to achieve load balancing. First, initialize Value , which are equal to , select servers in order , process the requests in sequence, and the performance requirement for each processing is The image or video corresponding to minus ,like If the number of servers is insufficient, the next server will be selected. If insufficient, Reset to initial value.

[0070] Data standardization: The access traffic of each server is cached in the access queue. The image or video at the head of the queue is taken out for standardization each time and encapsulated into detection packages with the same format. Figure 2 As shown, an image or video needs to go through three steps of framing, cropping and segmentation, and packaging to obtain the corresponding detection package.

[0071] It can be understood that in this embodiment, only the head of the queue is taken out, which means that the queue is a buffer, and each detection request is put into the queue and placed at the end of the queue. According to the idea of ​​first come first served, one data (request) is taken out for detection each time, and the next one is taken out after the detection is completed, and processed in sequence.

[0072] Specifically, the data standardization implementation method is as follows:

[0073] The method of video framing is as follows: for a video, use the video codec tool FFmpeg to extract all key frames in the video ( frame), get the frame set ,in is the cardinality of the frame set, i.e. the number of frames, and ; For images, there is no need to perform frame processing, and the frame set only contains the image, that is, ;

[0074] The cropping and segmentation method is: for each frame in the frame set, face cropping and image segmentation operations are performed respectively.

[0075] It should be noted that this step is to provide all detection models with image blocks of required content and size. Xception (model A) analyzes the face area of ​​the image, because DeepFake only tampered with the face area of ​​the image, and the other areas are real; ResNet-50 (model B) and ConvNeXt (model C) do not need to crop the face area, because the Diffusion Model and GAN forgery methods detected by these two models are all full-image synthesis, and there is no real area. However, if the image size is too large, it will cause the GPU to be overloaded and the inference speed to slow down, so the image needs to be divided into several image blocks for the model to detect in parallel.

[0076] The encapsulation method is as follows: after cropping and segmentation, the image blocks required by each model are encapsulated into corresponding detection units, and an identifier of the specified detection model is added to realize the directional sending of the detection unit in the routing module of the detection stage. The detection unit corresponding to Xception (model A) contains only one cropped face image block; the detection units corresponding to ResNet-50 (model B) and ConvNeXt (model C) contain segmented image blocks. The detection units of the three detection models are then encapsulated to obtain detection clusters, and each detection cluster corresponds to a frame in the frame set. All detection clusters of a video or image are encapsulated into detection packages, and a detection package is the result of the standardization of a video or image.

[0077] Multi-model joint decision-making: After the images or videos are standardized, detection packages with a unified format are obtained and cached in the detection queue. Each time the detection package at the head of the queue is taken out for forgery detection, the detection results of the image or video corresponding to each detection package are finally obtained.

[0078] like Figure 3 As shown, the specific method is as follows:

[0079] Decapsulation and routing: The detection packet corresponding to each image or video is decapsulated to obtain several detection clusters, and each detection cluster is decapsulated to obtain three detection units. The routing module determines which model it should be sent to based on the identifier on the detection unit.

[0080] Model detection: Build a multimodal forgery detection framework consisting of Xception (model A), ResNet-50 (model B) and ConvNeXt (model C), which respectively detect DeepFake, Diffusion Model, and GAN forged images or videos.

[0081] Each model needs to be trained before it can be used. The three models are trained using different data sets. Each data set contains real images and generated images of the corresponding forged type. The real images are labeled as 0, and the forged images are labeled as 1. 80% of the images are used as training sets, 10% of the images are used as validation sets, and 10% of the images are used as test sets. When the model is trained on the training set, each model predicts the input image and finally outputs a confidence score. , represents the probability that this image is a fake image. The training uses the cross entropy function:

[0082]

[0083] in The label representing the image, i.e. the real image , fake image , the training goal of this loss is to make As small as possible until convergence and evaluate using the test set data.

[0084] During the use process, the image blocks in each detection unit are passed into the corresponding detection model as a batch, and the confidence of each detection unit can be obtained. Since the detection unit corresponding to Xception (model A) contains only one face image, only one confidence is obtained. , while the detection unit corresponding to ResNet-50 (model B) and ConvNeXt (model C) may contain multiple image blocks, so it is necessary to take the average of the confidence of these image blocks to obtain the confidence of the detection unit .

[0085] Joint decision-making: Before putting it into use, in addition to training the model, it is also necessary to train the confidence weight of each detection unit. Models A, B, and C correspond to weights respectively. First, the images in the three validation sets are combined. Each image can be predicted by three trained models to obtain three confidence levels. Finally, we get several triplets with real or fake labels as the data set for weight training. Then, we continuously train and optimize the three weights through the loss function. :

[0086] ;

[0087] ;

[0088] in Represents the label of the image corresponding to the confidence triplet, that is, the real image , fake image , the training goal of this loss is to ensure Under the premise of As small as possible until convergence. Represents the confidence after aggregation

[0089] During the commissioning process, the confidence of the three detection units in the detection cluster will be obtained after the model is tested. The weights obtained through training , the confidence of each detection cluster is obtained by the same calculation method as the confidence aggregation formula .

[0090] Authenticity streaming judgment: Before putting it into use, you need to pre-select the threshold for authenticity judgment ,when When , the image is judged to be a real image. When , the image is judged as a forged image. In the joint decision step, we have obtained several confidence triplets as a data set for weight training. Here, this embodiment also uses this data set to obtain the threshold For each confidence triplet , can be obtained through the trained weights and the confidence aggregation formula to obtain the aggregated confidence , and each Corresponding to a true label . All Sort from small to large and select them as thresholds in turn, and get the predicted label according to the authenticity judgment rules mentioned above. , and by comparing and Get the prediction accuracy of the threshold, and finally Select the threshold value that corresponds to the maximum prediction accuracy.

[0091] It should be noted that during the commissioning process, the joint decision will obtain the confidence of the detection cluster , by comparison and The authenticity of the detection cluster is obtained. For an image, a detection package contains only one detection cluster, and the authenticity of the detection cluster is the authenticity of the image.

[0092] Example 2

[0093] The difference between this embodiment and embodiment 1 is that, during the detection process, for video detection, a detection package usually contains multiple detection clusters, corresponding to multiple cluster confidences, so by combining sensitive parameters Together to judge the authenticity of , which is preset before use and is usually set to 0.5. When the proportion of forged detection clusters to all detection clusters is greater than or equal to When the video is judged as fake, sensitive parameters The smaller the video is, the easier it is to identify it as fake. Videos with detection clusters, where we have obtained greater than or equal If a fake detection cluster is found, the detection will not continue and the detection result will be directly returned as fake.

[0094] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments including components are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.

Claims

1. A generalized forged media detection method for realistic high-traffic scenarios, characterized in that: The steps include: Step 1: Distribute the data requested by the network to multiple servers through a weighted round-robin algorithm; Step 2: Cache the access traffic of each server in the access queue, and take out the data at the head of the queue each time to perform standardization processing to obtain a detection packet, wherein the standardization processing includes framing, clipping and segmentation, and encapsulation; Step 3: Cache the detection packets in the detection queue, and take out the detection packet at the head of the queue for forgery detection each time. The method of forgery detection is: Firstly, a multimodal forgery detection framework consisting of three expert models, Xception, ResNet-50 and ConvNeXt, is constructed; Then, a dataset with real images and corresponding generated forged images is constructed, and the dataset is used to train and verify the multimodal forgery detection network model; Extract the detection package to be detected, decapsulate the detection package to obtain Each detection cluster is decapsulated to obtain a detection unit, and the detection unit outputs the detection result through a multimodal forgery detection network model.

2. A generalized forged media detection method for realistic high-traffic scenarios according to claim 1, characterized in that: The data requested by the network includes images and videos.

3. According to claim 2, a generalized forged media detection method for realistic high-traffic scenarios is characterized in that: In step 1, the method for allocating network requests is as follows: Quantify the performance indicators of each server and select a The image is used as the reference. The inference time of all servers for the image is obtained by performing inference in the servers. And memory usage ratio , the performance quantitative index of each server is obtained by the following formula: ； in, ; The performance required for network request data is quantified based on the size of the image or video frame and the number of key frames. The quantization expression of an image or video frame is as follows: ； in is the number of key frames, for images , for videos ; Weighted polling based on performance quantification of server and network request data.

4. According to claim 3, a generalized forged media detection method for realistic high-traffic scenarios is characterized in that: The weighted polling method is as follows: first initialize Value , which are equal to , select servers in order , process the requests in sequence, and the performance requirement for each processing is The image or video corresponding to minus ,like If the number of servers is insufficient, the next server will be selected. If insufficient, Reset to initial value.

5. According to claim 3, a generalized forged media detection method for realistic high-traffic scenarios is characterized in that: The method of video segmentation is as follows: for a video, use the video codec tool FFmpeg to extract all key frames in the video to obtain a key frame set. ,in is the cardinality of the frame set, i.e. the number of frames, and ; For images, there is no need to perform frame processing, and the frame set only contains the image, that is, ; The cropping and segmentation method is: for each frame in the frame set, performing face cropping and image segmentation operations respectively; The encapsulation method is as follows: after cropping and segmentation, the image blocks required by each model are encapsulated into corresponding detection units, and an identifier of a specified detection model is added to realize directional sending of the detection units in a routing module in the detection phase.

6. According to claim 5, a generalized forged media detection method for realistic high-traffic scenarios is characterized in that: In step 3, three data sets containing real images and corresponding forged type generated images are constructed respectively, where the real images are labeled as 0 and the forged images are labeled as 1, 80% of the images are used as training sets, 10% of the images are used as validation sets, and 10% of the images are used as test sets.

7. A generalized forged media detection method for realistic high-traffic scenarios according to claim 6, characterized in that: The training method of the multimodal forgery detection network model is: The three training sets are used as input to train Xception, ResNet-50 and ConvNeXt respectively. Each model predicts the input image and finally outputs a confidence score. , represents the probability that this image is a forged image, and the cross entropy function used in training is: ； in The label representing the image, i.e. the real image , fake image , the training goal of this loss is to make As small as possible until convergence; The weights of the confidence of each model output are trained. First, the weights corresponding to the three models are defined as , the images in the three validation sets are combined, and each image can be predicted by three trained models to obtain three confidence levels Finally, several triplets with real or fake labels are obtained as the data set for weight training. First, the triple confidence is aggregated, and then the three weights are continuously trained and optimized through the loss function. , the expression is as follows: ； ； in Represents the label of the image corresponding to the confidence triplet, that is, the real image , fake image , the training goal of this loss is to ensure Under the premise of As small as possible until convergence. Represents the confidence after aggregation.

8. According to claim 5, a generalized forged media detection method for realistic high-traffic scenarios is characterized in that: In step 3, the three decapsulated detection units are sent to the corresponding models for detection by a routing module according to the identifiers on the detection units.

9. The generalized forged media detection method for realistic high-traffic scenarios according to claim 7, characterized in that: In step 3, the confidences of the three detection units in the detection cluster output by the corresponding model are aggregated through the trained weights to obtain the confidence of each detection cluster. , set the threshold ,when When , the image is predicted to be a real image, When , the image is predicted to be a forged image, and the predicted label is recorded as .

10. The generalized forged media detection method for realistic high-traffic scenarios according to claim 9, characterized in that: The threshold The method of determining is: all Sort from small to large and select them as thresholds in turn, according to the predicted label , and by comparing and Get the prediction accuracy of the threshold, and finally Select the threshold value that corresponds to the maximum prediction accuracy.

11. The generalized forged media detection method for realistic high-traffic scenarios according to claim 9, characterized in that: In step 3, for video detection, a detection package usually contains multiple detection clusters, corresponding to multiple cluster confidences, so by combining sensitive parameters Together to judge the authenticity, , when the proportion of forged detection clusters to all detection clusters is greater than or equal to , the video will be judged as forged. The smaller the sensitive parameter, the easier it is to judge the video as forged. Therefore, for a video containing Videos with detection clusters, where we have obtained greater than or equal If a fake detection cluster is found, the detection will not continue and the detection result will be directly returned as fake.

Images