Digital content detection method and device, storage medium and program product

Abstract

The invention provides a digital content detection method and device, a storage medium and a program product, and the method comprises the steps: obtaining to-be-processed target digital content and reference digital content in response to a detection task; the target digital content is input into a preset feature extraction model, so that the feature extraction model outputs a first data feature of the target digital content, the feature extraction model is obtained through training of a first sample data set, and the first sample data set comprises at least one sample digital content pair; labeling the sample digital content pairs with the same sample cue word and the same generation model as homologous labels; and according to the first data feature and the reference digital content, determining whether the target digital content is homologous with the reference digital content. According to the method, whether the target digital content and the reference digital content are homologous or not is effectively verified, the accuracy of digital content traceability and matching detection is improved, and reliable data support is provided for subsequent ownership protection of the digital content.

Description

Technical Field

[0001] The present application relates to the field of information processing technology, and in particular to a digital content detection method, device, storage medium and program product. Background Art

[0002] With the continuous advancement of artificial intelligence (AI) technology, artificial intelligence generated content (AIGC) has become a revolutionary force in digital content creation. In content question and answer, painting, photography, video production and other media fields, AIGC technology has demonstrated its powerful automation and innovation capabilities, providing creators with unprecedented tools and possibilities. For example, in the field of painting, through the AI ​​model, users only need to enter text descriptions to create images with a certain aesthetic and creativity. Similarly, on commercial platforms, users can more directly convert text into visual works, which has promoted the development of digital art.

[0003] However, with the widespread application of AIGC technology, especially in the field of art, the originality and intellectual property ownership of AI-generated content have gradually surfaced, sparking major disputes about creative property rights and copyright. Although AI can generate attractive visual works, the generation process of these works is usually lacking in transparency, and the ownership of the generated works remains unclear. Therefore, how to protect copyright ownership and ensure the authenticity of content in the context of the booming development of AIGC has become a key issue that needs to be resolved.

[0004] At present, some progress has been made in the detection of AI-generated content, and many solutions can detect whether digital content is generated by AI. However, most of these solutions only stay at the detection level and lack the ability to trace the source of the digital content generation method, making it difficult to effectively verify the originality and copyright information of digital content. Summary of the invention

[0005] The main purpose of the embodiments of the present application is to provide a digital content detection method, device, storage medium and program product, which can effectively verify whether the target digital content and the reference digital content are of the same origin, improve the accuracy of digital content traceability and matching detection, and provide reliable data support for subsequent digital content ownership protection.

[0006] In a first aspect, an embodiment of the present application provides a digital content detection method, comprising: in response to a detection task, obtaining target digital content to be processed and reference digital content; inputting the target digital content into a preset feature extraction model so that the feature extraction model outputs a first data feature of the target digital content, wherein the feature extraction model is obtained by training a first sample data set, the first sample data set includes at least one sample digital content pair, and the sample digital content pairs with the same sample prompt word and the same generation model are marked as homologous labels; based on the first data feature and the reference digital content, determining whether the target digital content and the reference digital content are homologous.

[0007] In one embodiment, before the target digital content is input into a preset feature extraction model so that the feature extraction model outputs the first data feature of the target digital content, it also includes: judging whether there is a target prompt word used to generate the target digital content in the input information of the detection task; if the target prompt word does not exist in the input information, executing the step of inputting the target digital content into a preset feature extraction model so that the feature extraction model outputs the first data feature of the target digital content.

[0008] In one embodiment, it also includes: if the target prompt word exists in the input information, determining whether the target prompt word is the same as the reference prompt word corresponding to the reference digital content; if the target prompt word is the same as the reference prompt word, inputting the target digital content and the reference digital content into a preset classification model respectively, so that the preset classification model outputs the first source information of the target digital content and the second source information of the reference digital content; and determining whether the target digital content and the reference digital content are of the same origin based on the first source information and the second source information.

[0009] In one embodiment, before respectively inputting the target digital content and the reference digital content into a preset classification model so that the preset classification model outputs the first source information of the target digital content and the second source information of the reference digital content, it also includes: obtaining a second sample data set, the second sample data set including a plurality of sample digital contents generated by different generation models, each of the sample digital contents being marked with information of the corresponding generation model; and according to the second sample data set, using a cross entropy loss function to train a preset neural network to obtain the preset classification model.

[0010] In one embodiment, the reference digital content is a comparison object specified by a user; determining whether the target digital content and the reference digital content are of the same origin based on the first data feature and the reference digital content includes: acquiring a second data feature of the reference digital content; calculating a feature similarity between the first data feature and the second data feature; if the feature similarity is greater than a first preset threshold, determining that the target digital content and the reference digital content are of the same origin; if the feature similarity is less than or equal to the first preset threshold, determining that the target digital content and the reference digital content are of different origins.

[0011] In one embodiment, there are multiple reference digital contents, and the second data feature and corresponding metadata of each reference digital content are stored in a preset database; determining whether the target digital content and the reference digital content are of the same origin based on the first data feature and the reference digital content includes: calculating the feature similarity between the first data feature and the second data feature of each reference digital content in the preset database respectively; determining whether the target digital content and the reference digital content are of the same origin based on the feature similarity and the metadata corresponding to each reference content.

[0012] In one embodiment, determining whether the target digital content and the reference digital content are of the same origin based on the feature similarity and the metadata corresponding to each of the reference digital contents includes: determining, from a plurality of the reference digital contents, a set of similar digital contents having a feature similarity with the target digital content greater than a second preset threshold value based on the feature similarity; comparing the metadata of the target digital content with the metadata of each digital content in the set of similar digital contents respectively; if there is specific digital content having the same metadata as the target digital content in the set of similar digital contents, determining that the target digital content and the specific digital content are of the same origin; and / or outputting the specific digital content; if there is no specific digital content having the same metadata as the target digital content in the set of similar digital contents, determining that the target digital content and the reference digital content are of different origins.

[0013] In one embodiment, before the target digital content is input into a preset feature extraction model so that the feature extraction model outputs the first data feature of the target digital content, it also includes: obtaining a plurality of different sample prompt words; obtaining a plurality of different content generation models; inputting the plurality of different sample prompt words into each of the content generation models one by one, so that each of the content generation models generates corresponding sample digital content according to the plurality of different sample prompt words, to obtain an original sample data set; pairing the sample digital contents in the original sample data set in pairs to form a plurality of sample digital content pairs, and marking the sample digital content pairs with the same sample prompt words and the same generation model as homologous labels to form the final first sample data set.

[0014] In one embodiment, before inputting the target digital content into a preset feature extraction model so that the feature extraction model outputs the first data feature of the target digital content, it also includes: according to the first sample data set, using a contrast loss function to fine-tune and train a preset feature extractor to obtain the feature extraction model.

[0015] In one embodiment, the first sample data set is a sample image set; before the target digital content is input into a preset feature extraction model so that the feature extraction model outputs the first data feature of the target digital content, it also includes: fine-tuning and training a preset feature extractor according to the sample image set, wherein the feature extractor includes an image feature extractor and a text feature extractor; during the training process, the image feature extractor and the text feature extractor are fine-tuned and trained using an inter-class contrast loss function and an image-text contrast loss function, respectively, to obtain the feature extraction model.

[0016] In a second aspect, an embodiment of the present application provides a method for detecting a product image, comprising: in response to a detection task, obtaining a target product image and a reference image to be processed, wherein the target product image includes product information; inputting the target product image into a preset feature extraction model so that the feature extraction model outputs a first image feature of the target product image, wherein the feature extraction model is trained by a first sample image set, and the first sample image set includes at least one sample image pair, and the sample image pairs with the same sample prompt word and the same generation model are annotated as homologous labels; and determining whether the target product image is homologous to the reference image based on the first image feature and the reference image.

[0017] In a third aspect, an embodiment of the present application provides a digital content detection device, including:

[0018] An acquisition module, used for acquiring target digital content and reference digital content to be processed in response to a detection task;

[0019] An extraction module, used for inputting the target digital content into a preset feature extraction model, so that the feature extraction model outputs a first data feature of the target digital content, wherein the feature extraction model is obtained by training a first sample data set, the first sample data set includes at least one sample digital content pair, and the sample digital content pairs with the same sample prompt word and the same generation model are marked as homologous labels;

[0020] The determination module is used to determine whether the target digital content and the reference digital content are of the same source according to the first data feature and the reference digital content.

[0021] In one embodiment, it also includes: a first judgment module, which is used to judge whether there is a target prompt word used to generate the target digital content in the input information of the detection task before inputting the target digital content into a preset feature extraction model so that the feature extraction model outputs the first data feature of the target digital content; if the target prompt word does not exist in the input information, the step of inputting the target digital content into a preset feature extraction model so that the feature extraction model outputs the first data feature of the target digital content is executed.

[0022] In one embodiment, it also includes: a second judgment module, which is used to judge whether the target prompt word is the same as the reference prompt word corresponding to the reference digital content if the target prompt word exists in the input information; a classification module, which is used to input the target digital content and the reference digital content into a preset classification model respectively if the target prompt word is the same as the reference prompt word, so that the preset classification model outputs the first source information of the target digital content and the second source information of the reference digital content; the classification module is also used to determine whether the target digital content and the reference digital content are of the same source based on the first source information and the second source information.

[0023] In one embodiment, it also includes: a first training module, which is used to obtain a second sample data set before respectively inputting the target digital content and the reference digital content into a preset classification model so that the preset classification model outputs the first source information of the target digital content and the second source information of the reference digital content, wherein the second sample data set includes multiple sample digital contents generated by different generation models, and each of the sample digital contents is marked with information of the corresponding generation model; based on the second sample data set, a preset neural network is trained using a cross entropy loss function to obtain the preset classification model.

[0024] In one embodiment, the reference digital content is a comparison object specified by a user; a determination module is used to obtain a second data feature of the reference digital content; calculate a feature similarity between the first data feature and the second data feature; if the feature similarity is greater than a first preset threshold, determine that the target digital content and the reference digital content are of the same source; if the feature similarity is less than or equal to the first preset threshold, determine that the target digital content and the reference digital content are of different sources.

[0025] In one embodiment, there are multiple reference digital contents, and the second data feature and corresponding metadata of each reference digital content are stored in a preset database; the determination module is used to calculate the feature similarity between the first data feature and the second data feature of each reference digital content in the preset database; based on the feature similarity and the metadata corresponding to each reference content, determine whether the target digital content and the reference digital content are of the same origin.

[0026] In one embodiment, the determination module is specifically used to determine, from the multiple reference digital contents, a set of similar digital contents whose feature similarity with the target digital content is greater than a second preset threshold value according to the feature similarity; respectively compare the metadata of the target digital content with the metadata of each digital content in the similar digital content set; if there is specific digital content with the same metadata as the target digital content in the similar digital content set, determine that the target digital content and the specific digital content are of the same origin; and / or output the specific digital content; if there is no specific digital content with the same metadata as the target digital content in the similar digital content set, determine that the target digital content and the reference digital content are not of the same origin.

[0027] In one embodiment, it also includes: a data construction module, which is used to obtain multiple different sample prompt words before inputting the target digital content into a preset feature extraction model so that the feature extraction model outputs the first data feature of the target digital content; obtain multiple different content generation models; input the multiple different sample prompt words into each of the content generation models one by one, so that each of the content generation models generates corresponding sample digital content according to the multiple different sample prompt words, thereby obtaining an original sample data set; pair the sample digital contents in the original sample data set in pairs to form multiple sample digital content pairs, and mark the sample digital content pairs with the same sample prompt words and the same generation model as homologous labels to form the final first sample data set.

[0028] In one embodiment, it also includes: a second training module, which is used to fine-tune the training of a preset feature extractor using a contrast loss function according to the first sample data set before inputting the target digital content into a preset feature extraction model so that the feature extraction model outputs the first data feature of the target digital content, so as to obtain the feature extraction model.

[0029] In one embodiment, the first sample data set is a sample image set; and further includes: a third training module, which is used to fine-tune and train a preset feature extractor according to the sample image set before inputting the target digital content into a preset feature extraction model so that the feature extraction model outputs the first data feature of the target digital content, wherein the feature extractor includes an image feature extractor and a text feature extractor; during the training process, the image feature extractor and the text feature extractor are fine-tuned and trained using an inter-class contrast loss function and an image-text contrast loss function, respectively, to obtain the feature extraction model.

[0030] In a fourth aspect, an embodiment of the present application provides an electronic device, including:

[0031] at least one processor; and

[0032] a memory communicatively coupled to the at least one processor;

[0033] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the electronic device to execute the method described in any one of the above aspects.

[0034] In a fifth aspect, an embodiment of the present application provides a cloud device, including:

[0035] at least one processor; and

[0036] a memory communicatively coupled to the at least one processor;

[0037] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the cloud device to execute the method described in any one of the above aspects.

[0038] In a sixth aspect, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions. When a processor executes the computer-executable instructions, the method described in any one of the above aspects is implemented.

[0039] In a seventh aspect, an embodiment of the present application provides a computer program product, including a computer program, which implements the method described in any of the above aspects when executed by a processor.

[0040] The digital content detection method, device, storage medium and program product provided in the embodiments of the present application automatically extract a representative first data feature from the target digital content by using a pre-trained feature extraction model, and the first data feature can characterize the source information of the target digital content. Since the feature extraction model is trained based on the first sample data set, in the first sample data set, the sample digital content pairs with the same sample prompt words and the same generation model are marked as homologous labels, and the prompt words and the generation model can both characterize the source information of the digital content, the trained feature extraction model can effectively capture the first data feature that accurately characterizes the source information in the target digital content, and can then effectively verify whether the target digital content and the reference digital content are homologous, thereby improving the accuracy of digital content traceability and matching detection, and providing reliable data support for the subsequent protection of digital content ownership. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The drawings herein are incorporated into the specification and constitute a part of the specification, illustrate embodiments consistent with the present application, and are used together with the specification to explain the principles of the present application. Obviously, the drawings described below are some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor.

[0042] Figure 1 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application;

[0043] Figure 2 A schematic diagram of an application scenario of a digital content detection system provided in an embodiment of the present application;

[0044] Figure 3 A schematic diagram of a flow chart of a digital content detection method according to an embodiment of the present application;

[0045] Figure 4 A schematic diagram of a process for constructing a data set provided in an embodiment of the present application;

[0046] Figure 5 A schematic diagram of a training process of a feature extraction model provided in an embodiment of the present application;

[0047] Figure 6 A schematic diagram of a training process of a feature extraction model provided in an embodiment of the present application;

[0048] Figure 7 A schematic diagram of a training process of a preset classification model provided in an embodiment of the present application;

[0049] Figure 8 A flowchart of a digital content detection method provided in an embodiment of the present application;

[0050] Fig. 9 A flowchart of a digital content detection method provided in an embodiment of the present application;

[0051] Fig.10 A flowchart of a method for detecting a commodity image provided in an embodiment of the present application;

[0052] Fig.11 A schematic diagram of the structure of a digital content detection device provided in an embodiment of the present application;

[0053] Fig.12 A schematic diagram of the structure of a cloud device provided in an embodiment of the present application.

[0054] The above drawings have shown clear embodiments of the present application, which will be described in more detail later. These drawings and text descriptions are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0055] Here, exemplary embodiments are described in detail, and examples thereof are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application.

[0056] The term "and / or" in this article is used to describe the association relationship of associated objects, specifically indicating that there may be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone.

[0057] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0058] In order to clearly describe the technical solution of the embodiment of the present application, the terms involved in the present application are first defined:

[0059] AI: Artificial Intelligence.

[0060] AIGC: Artificial Intelligence Generated Content, including but not limited to text, images, audio and video.

[0061] AI Editing: Use artificial intelligence technology to edit content, such as image style transfer, filter application, object recognition and removal, background replacement, etc.

[0062] SD: Stable Diffusion, a deep learning model for generating images.

[0063] GAN: Generative Adversarial Networks, Generative Adversarial Networks.

[0064] Prompt: The prompt word of the large model is an input used to guide the model to generate a specific type of output. It tells the model what type of content should be generated or in what style. In image generation, the prompt may be a text description of the desired image.

[0065] Diffusion Model: Diffusion Model is a model used to generate data, such as the SD model. It is based on the principle of random diffusion process, and gradually eliminates noise through optimization techniques such as gradient descent to generate highly realistic images. This model is particularly suitable for creating new visual content based on given text descriptions, and has been applied in many fields such as art creation and game design.

[0066] AI-generated images: AI-generated images refer to images automatically generated by artificial intelligence technologies such as diffusion models and generative adversarial networks (GANs). These images can be landscapes, objects, people, etc. They usually have a high sense of reality, so that ordinary people cannot distinguish their authenticity with the naked eye.

[0067] AI-Generated Image Provenance Detection: AI-Generated Image Provenance Detection aims to trace the source of AI-generated images by analyzing and comparing their features to determine which AI generation model or platform was used to generate them, so as to carry out applications such as copyright protection, responsibility tracing, and algorithm performance evaluation. It is of great significance to ensure the transparency and credibility of AI-generated content.

[0068] CLIP: Contrastive Language–Image Pretraining, is an innovative method proposed by OpenAI, which pre-trains from large-scale image-text pairs through contrastive learning. CLIP uses two independent encoders, one for images and one for text, to map them to the same embedding space, and maximizes the similarity of correctly paired images and texts through the contrastive loss function, while minimizing the similarity of unrelated pairs, thereby training a visual model with strong migration capabilities.

[0069] ViT: Vision Transformer, is a model that applies the Transformer architecture to computer vision tasks.

[0070] ConvNext: refers to a new type of convolutional neural network architecture.

[0071] BLIP: Bootstrapping Language-Image Pre-training, is a multimodal model that aims to improve the performance of the model on vision and language tasks by combining language and image data for pre-training. BLIP-2 is an improvement and extension of the original BLIP model, aiming to further improve the performance of multimodal learning. Similar to its predecessor, BLIP-2 combines language and image data for pre-training to enhance the performance of the model on vision and language tasks.

[0072] ResNet: Residual Network, residual network.

[0073] DenseNet: Densely Connected Convolutional Networks, densely connected convolutional networks.

[0074] MobileNet: A lightweight convolutional neural network architecture.

[0075] EfficientNet: An efficient neural network.

[0076] HRNet: High-Resolution Network, high-resolution network.

[0077] Swin Transformer: Shifted Window Transformer is a hierarchical visual Transformer model that achieves a balance between computational efficiency and performance by computing self-attention within a local window and shifting it between windows.

[0078] MobileViT: Mobile Vision Transformer, an efficient vision model designed specifically for mobile devices.

[0079] UGC: User-Generated Content, user-generated content.

[0080] The digital content detection method of the embodiment of the present application can be applied to any field involving watermark processing.

[0081] With the continuous advancement of artificial intelligence (AI) technology, artificial intelligence generated content (AIGC) has become a revolutionary force in digital content creation. In content question and answer, painting, photography, video production and other media fields, AIGC technology has demonstrated its powerful automation and innovation capabilities, providing creators with unprecedented tools and possibilities. For example, in the field of painting, through the AI ​​model, users only need to enter text descriptions to create images with a certain aesthetic and creativity. Similarly, on commercial platforms, users can more directly convert text into visual works, which has promoted the development of digital art.

[0082] However, with the widespread application of AIGC technology, especially in the field of art, the originality and intellectual property ownership of AI-generated content have gradually surfaced, sparking major disputes about creative property rights and copyright. Although AI can generate attractive visual works, the generation process of these works often lacks transparency, and the ownership of the generated works remains unclear. In addition, in order to generate high-quality images, AI models require a large amount of data for training. If the training data includes copyrighted works, the generated images may infringe the copyright of the original author, thus causing various disputes.

[0083] At the same time, with the rise of the open source community of AI generative models, more and more relevant practitioners have released models derived from the basic version of the diffusion model series. These various diffusion model versions bring higher generalization requirements to the retrieval matching of generated images in actual scenarios. Since generated images in actual scenarios are often transmitted through distorted channels such as social media platforms, the generated images may undergo post-processing such as scaling, cropping, and secondary compression during transmission. Therefore, in real-world detection scenarios, detectors must have the ability to resist these common editing processes.

[0084] Taking the e-commerce platform scenario as an example, merchants on the e-commerce platform can use AI big models to generate different product images to update the products. If others see that the product is selling well, they are likely to process the product image by scaling, cropping, secondary compression, etc., in an attempt to escape infringement, which seriously infringes on the information security of the original user.

[0085] In related technologies, some progress has been made in AI-generated content detection, which can detect whether digital content is generated by AI. However, most of these solutions only stay at the detection level and lack the traceability of the generation method of digital content, making it difficult to effectively verify the originality and copyright information of digital content.

[0086] In order to solve at least one of the above problems, the embodiment of the present application provides a digital content detection solution, which automatically extracts a representative first data feature from the target digital content by using a pre-trained feature extraction model, and the first data feature can characterize the source information of the target digital content. Since the feature extraction model is trained based on the first sample data set, in the first sample data set, the sample digital content pairs with the same sample prompt words and the same generation model are marked as homologous labels, so the trained feature extraction model can effectively capture the first data feature characterizing the source information in the target digital content, and then can effectively verify whether the target digital content and the reference digital content are homologous, thereby improving the accuracy of digital content traceability and matching detection, and providing reliable data support for the subsequent digital content ownership protection.

[0087] Some embodiments of the present application are described in detail below in conjunction with the accompanying drawings. In the case where there is no conflict between the embodiments, the following embodiments and the features in the embodiments can be combined with each other. In addition, the step sequence in the following method embodiments is only an example and not a strict limitation.

[0088] like Figure 1 As shown, this embodiment provides an electronic device 1, including: at least one processor 11 and a memory 12, Figure 1 A processor is taken as an example. The processor 11 and the memory 12 are connected via a bus 10. The memory 12 stores instructions that can be executed by the processor 11, and the instructions are executed by the processor 11, so that the electronic device 1 can execute all or part of the process of the method in the following embodiment, so as to effectively verify whether the target digital content and the reference digital content are of the same source, improve the accuracy of digital content traceability and matching detection, and provide reliable data support for the subsequent digital content ownership protection.

[0089] In one embodiment, the electronic device 1 may be a mobile phone, a tablet computer, a laptop computer, a desktop computer, or a large computing system composed of multiple computers.

[0090] Figure 2 Schematic diagram of a digital content detection system application scenario 200 provided in an embodiment of the present application. Figure 2 As shown, the system includes: a server 210 and a terminal 220, wherein:

[0091] Server 210 may be a data platform that provides digital content detection services, such as a digital information security service platform for e-commerce. In actual scenarios, an e-commerce platform may have multiple servers 210. Figure 2 In the figure, one server 210 is taken as an example.

[0092] The terminal 220 may be a computer, mobile phone, tablet or other device used by the user to log in to the digital information security service platform. There may be multiple terminals 220. Figure 2 Two terminals 220 are taken as an example for illustration.

[0093] The terminal 220 and the server 210 can transmit information via the Internet, so that the terminal 220 can access the data on the server 210. The terminal 220 and / or the server 210 can be implemented by the electronic device 1.

[0094] The digital content detection solution of the embodiment of the present application can be deployed on the server 210, can also be deployed on the terminal 220, or can be deployed partially on the server 210 and partially on the terminal 220. In actual scenarios, it can be selected based on actual needs, and this embodiment does not limit it.

[0095] When the digital content detection solution is fully or partially deployed on the server 210 , a calling interface may be opened to the terminal 220 to provide algorithm support to the terminal 220 .

[0096] The method provided in the embodiment of the present application can be implemented by executing the corresponding software code by the electronic device 1, and can be implemented by exchanging data with the server. The electronic device 1 can be a local terminal device. When the method is run on the server, the method can be implemented and executed based on the cloud interaction system, wherein the cloud interaction system includes a server and a client device.

[0097] In a possible implementation, the method provided in the embodiment of the present application provides a graphical user interface through a terminal device, wherein the terminal device can be the local terminal device mentioned above, or can be the client device in the cloud interaction system mentioned above.

[0098] Please see Figure 3 , which is a digital content detection method of an embodiment of the present application, the method can be Figure 1 The electronic device 1 shown is used to perform and can be applied to Figure 2In the digital content detection application scenario shown in , it is possible to effectively verify whether the target digital content and the reference digital content are of the same source, thereby improving the accuracy of digital content traceability and matching detection, and providing reliable data support for the subsequent digital content ownership protection. This embodiment takes the terminal 220 as an example of the execution end, and the method includes the following steps:

[0099] Step 301: In response to a detection task, obtain target digital content and reference digital content to be processed.

[0100] In this step, the detection task can be actively triggered by the user or automatically triggered by the system. The detection task can specify the target digital content to be processed. The target digital content refers to the digital content that needs to be traced and matched. The digital content includes but is not limited to images, texts, audio and video. Reference digital content refers to standard digital content used to assist in detection. The reference digital content includes but is not limited to images, texts, audio and video generated based on AI generation models. The reference digital content can be specified by the user. For example, if the user believes that his image has been infringed, the original image that may be infringed by the user can be used as the reference digital content, and the image that the user has provided as evidence of possible infringement can be determined as the target digital content to be tested. Optionally, the reference digital content can also be content specified by the system, such as a certified original image pre-stored in the system database.

[0101] Step 302: Input the target digital content into a preset feature extraction model so that the feature extraction model outputs a first data feature of the target digital content, wherein the feature extraction model is trained by a first sample data set, and the first sample data set includes at least one sample digital content pair, and the sample digital content pairs with the same sample prompt word and the same generation model are marked as homologous labels.

[0102] In this embodiment, the feature extraction model is trained based on a first sample data set. In the first sample data set, sample digital content pairs with the same sample prompt words and the same generation model are annotated as homologous labels. Both the prompt words and the generation model can represent the source information of the digital content. Therefore, the feature extraction model trained with the first sample data set can learn the feature representations related to the prompt words and the generation model in the digital content, so that the feature extraction model can effectively capture the first data features representing the source information in the target digital content during the inference stage, so that the first data features can accurately represent the source information of the target digital content.

[0103] In one embodiment, before step 302, a data set construction process is further included, including: obtaining a plurality of different sample prompt words. Obtaining a plurality of different content generation models. Inputting a plurality of different sample prompt words into each content generation model one by one, so that each content generation model generates corresponding sample digital content according to a plurality of different sample prompt words, respectively, to obtain an original sample data set. Pairing the sample digital contents in the original sample data set in pairs to form a plurality of sample digital content pairs, and marking the sample digital content pairs with the same sample prompt words and the same generation model as homologous labels to form a final first sample data set.

[0104] In this embodiment, by obtaining a plurality of different sample prompt words and content generation models, a variety of sample digital content is generated. This diversity helps to train the feature extraction model so that it can better adapt to original digital content of different types and styles and improve the generalization ability of the model. By pairing the generated original sample digital content in pairs and marking the sample digital content pairs with the same prompt words and generation model as homologous labels, a structured sample data set is constructed, which can provide more meaningful context information for subsequent feature extraction. By marking homologous labels, the model can more effectively identify and distinguish the data features of homologous and heterologous digital content, and improve the accuracy of feature extraction.

[0105] like Figure 4 As shown, it is a schematic diagram of a process of constructing a data set provided by the present application, taking image data as an example, which may include the following steps:

[0106] Step 1: Collect a large amount of text prompt words (for example, more than 100,000), and use a filtering algorithm to remove duplicates from the text prompt words to ensure that each sample text prompt word is unique and non-repetitive. Figure 4 The sample prompt in the example is "A knight in a space suit riding a Arabian horse walking on the moon."

[0107] Step 2: Build an automated AI image generation system that includes multiple open source AI generation models (such as SDv1.4 of the SD model, SDv2.1 of the SD model, SDXL, SD3, and FLUX (a generation model based on a diffusion model), etc.). Assume that multiple content generation models are numbered, including generation model 1, generation model 2, ..., generation model n, where n is an integer.

[0108] Step 3: Use the sample prompt words in step 1 to input them one by one into the automated AI generated image system in step 2 to generate batches of AI generated image pairs. Each input sample prompt word (with other input conditions remaining the same) generates at least 2 AI generated images in each AI generated model. For example Figure 4In the example, the sample prompt word "a knight in a space suit riding a sweat-blooded horse walking on the moon" is input into the generation model 1, the generation model 2, ..., the generation model n, and each generation model generates 2 images. Figure 4 The example shows an image pair generated by generative model 1 and an image pair generated by generative model 2. In this way, an original sample data set formed by multiple AI-generated images can be obtained.

[0109] Optionally, the first sample data set is used to train a feature extraction model. Before model training, the training data labels are first defined, and image pairs generated in the same AI generation model using the same text prompt words in the original data set (other input conditions are also the same) are regarded as the same category and marked with homologous labels, otherwise they are regarded as different categories (for example, two images are generated in the same AI generation model using different text prompt words or the same text prompt words are generated in different AI generation models and are regarded as different categories). Therefore, assuming that i different text prompt words and n AI generation models are used to construct a sample data set, the entire first sample data set can contain c=ixn different category labels, and each category can contain at least 2 images, where i is a positive integer and n is a positive integer.

[0110] Optionally, in order to adapt to different training tasks, the original sample data set can be processed accordingly. For example, in order to enhance the diversity of the sample data set, a random data enhancement scheme can be used to enhance the original sample data set. Taking the sample image set as an example, the specific enhancement scheme includes but is not limited to random cropping, maximum size of the long side, compression quality, horizontal flipping, Gaussian noise, Gaussian blur, random rotation of 90 degrees, edge completion, completion to a fixed size, brightness contrast transformation, grid dropout, grayscale, translation scale rotation, color jittering, random cropping and scaling, center cropping, random brightness adjustment, color jittering, color space conversion, random occlusion, salt and pepper noise, affine transformation, etc.

[0111] In one embodiment, before step 302, the method may further include: fine-tuning and training a preset feature extractor using a contrast loss function according to the first sample data set to obtain a feature extraction model.

[0112] In this embodiment, before inputting the target digital content into the preset feature extraction model, the preset feature extractor can be fine-tuned and trained using the contrast loss function according to the first sample data set, effectively improving the accuracy and robustness of the feature extraction model. The contrast loss function can enhance the model's ability to distinguish features of different categories, so that the feature extraction model can more accurately extract representative first data features when processing the target digital content. This not only improves the generalization ability of the feature extraction model, but also reduces the risk of overfitting, so that it can better adapt to the diverse digital content input in practical applications and improve the performance and reliability of the overall system.

[0113] Optionally, the preset feature extractor can be a feature extractor based on the CLIP pre-trained model. CLIP uses two independent encoders, one for images and one for text, to map them to the same embedding space, and maximizes the similarity of correctly paired images and texts through a contrast loss function, while minimizing the similarity of unrelated pairs, thereby training a visual model with strong migration capabilities. In actual scenarios, the original pre-trained CLIP image feature extractor can be used to extract the AI-generated image feature vector, but the feature vector can only distinguish whether the generation prompt words used by the two AI-generated images are the same, but cannot distinguish between two AI-generated images generated in two different generation models using the same generation prompt words, so it is impossible to accurately trace the AI-generated image. Therefore, in this embodiment, in the first sample data set, sample digital content pairs with the same sample prompt words and the same generation model are annotated as homologous labels, and the first sample data set is used as training data, and the CLIP image feature extractor is fine-tuned and trained using the contrast loss function, and then the fine-tuned CLIP image feature extractor is used as the feature extraction model, so that the feature extraction model can learn the feature representation related to the prompt words and the generation model in the digital content, so that the image feature vector extracted by the feature extraction model can not only distinguish whether the generation prompt words used in two AI-generated images are the same, but also distinguish between two AI-generated images generated in two different generation models using the same generation prompt words, thereby completing the accurate tracing of the AI-generated images.

[0114] Optionally, the backbone network of the CLIP pre-training model involved in the embodiments of the present application may adopt: ViT structure pre-training model and ConvNext structure pre-training model. In addition, the image feature extractor and text feature extractor in BLIP and BLIP2 (image-to-text model, the opposite task of CLIP) may also be used.

[0115] like Figure 5 FIG. 1 is a schematic diagram of a training process of a feature extraction model provided in an embodiment of the present application. Taking a sample image dataset as an example, it is assumed that the first sample dataset includes the following content with category labels:

[0116] Category 1: image pairs generated by model 1 for prompt word 1.

[0117] Category 2: Image pairs generated by model 2 for prompt word 1.

[0118] Category 3: Image pairs generated by model 1 for cue word 2.

[0119] …

[0120] Category c: image pairs generated by model n for cue word i.

[0121] The CLIP image feature extractor is fine-tuned and trained using the first sample data set. During the training process, the CLIP image feature extractor extracts a feature vector for each sample image, such as:

[0122] The feature vectors of the sample image pairs of category 1 are f1_1 and f1_2 respectively.

[0123] The feature vectors of the sample image pairs of category 2 are f2_1 and f2_2 respectively.

[0124] The feature vectors of the sample image pairs of category 3 are f3_1 and f3_2 respectively.

[0125] …

[0126] The feature vectors of the sample image pairs of category c are fc_1 and fc_2 respectively.

[0127] The contrast loss function is used to fine-tune the CLIP image feature extractor to obtain the final feature extraction model.

[0128] In an optional embodiment, the classification loss function used for feature extraction model training may also include: Cross-Entropy Loss and Focal Loss (a loss function for solving category imbalance), etc. The embodiment of the present application does not limit the type of loss function.

[0129] In one embodiment, the first sample data set is a sample image set. Before step 3602, the method may further include: fine-tuning and training a preset feature extractor according to the sample image set, wherein the feature extractor includes an image feature extractor and a text feature extractor. During the training process, the image feature extractor and the text feature extractor are fine-tuned and trained using an inter-class contrast loss function and an image-text contrast loss function, respectively, to obtain a feature extraction model.

[0130] In this embodiment, by combining image and text data for multimodal contrast learning, a more effective training method is provided for the traceability matching of AI-generated images. The feature extractor consists of an image feature extractor and a text feature extractor, and the inter-class contrast loss function and the image-text contrast loss function are used to fine-tune the feature extractor during the training process. The image feature extractor is enabled to better distinguish image features from different sources, thereby improving the accuracy of image feature extraction. At the same time, the image-text contrast loss function enhances the association ability between the text feature extractor and the image feature extractor, enabling the model to more effectively match and associate visual information with text information. This dual contrast loss training strategy not only improves the multimodal understanding ability of the model, but also enhances its generalization ability in different tasks, thereby providing a more reliable and accurate feature extraction effect in practical applications. By fine-tuning the feature extractor using a sample image set, the performance and adaptability of the feature extraction model are significantly improved.

[0131] Optionally, the feature extractor may be a CLIP image feature extractor and a CLIP text feature extractor.

[0132] like Figure 6 As shown, a training flow diagram of a feature extraction model provided by the present application is shown. Taking a sample image dataset as an example, the definition of sample labels can refer to the description of the above embodiment. Assume that the first sample dataset includes the following content with category labels:

[0133] Category 1: image pairs generated by model 1 for prompt word 1.

[0134] Category 2: Image pairs generated by model 2 for prompt word 1.

[0135] Category 3: Image pairs generated by model 1 for cue word 2.

[0136] …

[0137] Category c: image pairs generated by model n for cue word i.

[0138] The images in the above sample image set are applied to the CLIP image feature extractor, and the image features f of the sample images are extracted using the CLIP image feature extractor. The text of the sample prompt words corresponding to the above sample images is applied to the CLIP text feature extractor, and the text features v of the sample prompt words are extracted using the CLIP text feature extractor. The following feature vectors are obtained respectively:

[0139] The feature vectors of the first sample image in the sample image pair of category 1 are f1_1, corresponding to the feature vector v1_1 of the sample prompt word, and the feature vectors of the second sample image are f1_2, corresponding to the feature vector v1_2 of the sample prompt word.

[0140] The feature vectors of the first sample image in the sample image pair of category 2 are f2_1, corresponding to the feature vector v2_1 of the sample prompt word, and the feature vectors of the second sample image are f2_2, corresponding to the feature vector v2_2 of the sample prompt word.

[0141] The feature vectors of the first sample image in the sample image pair of category 3 are f3_1, corresponding to the feature vector v3_1 of the sample prompt word, and the feature vectors of the second sample image are f3_2, corresponding to the feature vector v3_2 of the sample prompt word.

[0142] …

[0143] The feature vectors of the first sample image in the sample image pair of category c are fc_1, corresponding to the feature vector vc_1 of the sample prompt word, and the feature vectors of the second sample image are fc_2, corresponding to the feature vector vc_2 of the sample prompt word.

[0144] The CLIP image feature extractor and text feature extractor are fine-tuned and trained using the inter-class contrast loss function and image-text contrast loss function respectively. The image-text contrast loss can ensure that the image features extracted from the target product image by the fine-tuned image feature extractor are aligned with the text features describing the content of the target product image, thereby achieving multimodal training and improving the performance of the final feature extraction model.

[0145] Optionally, the inter-class contrast loss function used for feature extraction model training can also be: contrast loss (Contrastive Loss), triplet loss (Triplet Loss), Circle Loss (a loss function through unified classification learning and sample pair learning), InfoNCE loss, N-pair Loss (information maximization unconditional estimation loss) and Soft-Nearest Neighbors Loss (soft neighbor loss), etc., which are not limited in this application.

[0146] Step 303: Determine whether the target digital content and the reference digital content are of the same source according to the first data feature and the reference digital content.

[0147] In this embodiment, the first data feature extracted in step 302 can accurately characterize the source information of the target digital content. Therefore, it is possible to effectively verify whether the target digital content and the reference digital content are of the same origin based on the first data feature, thereby improving the accuracy of digital content traceability and matching detection, and providing reliable data support for subsequent digital content ownership protection.

[0148] In one embodiment, before step 302, the method further includes: determining whether the input information of the detection task contains a target prompt word used to generate the target digital content. If the input information does not contain the target prompt word, the step of inputting the target digital content into a preset feature extraction model is executed so that the feature extraction model outputs a first data feature of the target digital content.

[0149] In the present embodiment, the user can specify input information in the detection task, such as the user uploads the specified target commodity image and reference image, and can also give the text prompt words used by the large model to generate the target commodity image, or give the text prompt words used by the large model to generate the reference image, wherein the text prompt words of the target commodity image are the target prompt words, and the text prompt words corresponding to the reference image are the reference prompt words. By first judging whether there is a target prompt word for generating the target digital content in the input information of the detection task before inputting the target digital content into the preset feature extraction model, when there is no target prompt word in the input information, it is indicated that the prompt word does not need to be processed, and the feature extraction step of step 302 can be directly executed, avoiding unnecessary prompt word processing process, thereby simplifying the data processing flow. It is ensured that the relevant processing of the prompt word is only carried out when necessary, reducing the waste of computing resources, and accelerating the speed of feature extraction. In addition, the method can also reduce the risk of feature extraction errors caused by the missing of the prompt word, improve the robustness and reliability of the model when processing diversified inputs, and then improve the performance of the overall system.

[0150] In one embodiment, the method further includes: if there is a target prompt word in the input information, determining whether the target prompt word is the same as the reference prompt word corresponding to the reference digital content. If the target prompt word is the same as the reference prompt word, inputting the target digital content and the reference digital content into a preset classification model respectively, so that the preset classification model outputs first source information of the target digital content and second source information of the reference digital content. Determining whether the target digital content and the reference digital content are of the same source is based on the first source information and the second source information.

[0151] In the present embodiment, when there is a target prompt word in the input information, it is further determined whether the target prompt word is the same as the reference prompt word corresponding to the reference digital content, so as to realize accurate identification and verification of the source of the digital content. When the target prompt word is the same as the reference prompt word, the target digital content and the reference digital content are further input into a preset classification model respectively to obtain the corresponding source information. The source information here includes but is not limited to the generation model information of the digital content, the model version information, the platform information to which the model belongs, and the like. This process can effectively identify the source information of the digital content, so as to determine whether the target digital content and the reference digital content are of the same source. Not only the ability to distinguish the source of the digital content is improved, but also the accuracy and reliability of the system when processing similar content are enhanced. By analyzing the source information using the classification model, the verification and tracing of the content can be better supported, and the possibility of misjudgment can be reduced. In addition, this method can also provide important support in application scenarios such as content audit, copyright protection and information tracing, and improve the intelligent level and application value of the overall system.

[0152] In one embodiment, before the target digital content and the reference digital content are respectively input into the preset classification model so that the preset classification model outputs the first source information of the target digital content and the second source information of the reference digital content, the method further includes: obtaining a second sample data set, the second sample data set including a plurality of sample digital contents generated by different generation models, each sample digital content being marked with information of the corresponding generation model. According to the second sample data set, a preset neural network is trained using a cross entropy loss function to obtain a preset classification model.

[0153] In this embodiment, before the target digital content and the reference digital content are input into the preset classification model, a training process of the preset classification model is also included. The second sample data set includes a plurality of sample digital contents generated by different generation models. Figure 4 The original data set constructed by the corresponding embodiment is used to generate the second sample data set. Taking the sample image set as an example, each sample image in the second sample data set is marked with the generation model information of its source. Using these marking information, the preset neural network is trained using the cross entropy loss function to obtain an optimized classification model. This training method enables the classification model to more accurately learn and distinguish the characteristics of different generation models, so that in practical applications, the source information of the target digital content and the reference digital content can be more effectively output, thereby improving the accuracy and reliability of whether the short digital content is homologous. In addition, the method also improves the generalization ability of the model, enabling it to adapt to a variety of generation models and content types, thereby enhancing the practicality and value of the overall system in application scenarios such as content review, copyright protection, and information traceability.

[0154] like Figure 7As shown, a schematic diagram of a training process of a preset classification model provided in an embodiment of the present application is shown, taking a sample image dataset as a training dataset as an example, which may include the following steps:

[0155] Step 1: Configure the training input images, define the real image set as category 0, define the image set generated by AI generation model 1 as category 1, define the image set generated by AI generation model 2 as category 2, and so on, define the image set generated by AI generation model n as category n. Among them, real images refer to images that are not generated by AI generation models, such as images captured by cameras or images drawn by users using drawing tools. Adding real images to the sample data set allows the classification model to learn what are AI-generated images and what are not AI-generated images.

[0156] Step 2: Use the cross entropy loss function to train the AI ​​generated image tracing classification model. The classification model can be a preset neural network structure. For example, the AI ​​generated model tracing classification model can adopt a series of deep models such as ResNet, DenseNet, MobileNet, EffcientNet, HRNet, ConvNext, ViT, Swin Transformer and MobileVit. By constraining the cross entropy loss between sample images (such as real images, images generated by model 1, images generated by model 2... images generated by model n) and corresponding category prediction results (such as real images corresponding to prediction 0, images generated by model 1 corresponding to prediction 1, images generated by model 2 corresponding to prediction 2,..., images generated by model n corresponding to prediction n), the trained classification model can accurately output the predicted category results of each input image, and the predicted category results can characterize the source information of the input image, such as which generation model it comes from, the version information of the model, etc.

[0157] In one embodiment, step 303 may specifically include: obtaining a second data feature of the reference digital content. Calculating a feature similarity between the first data feature and the second data feature. If the feature similarity is greater than a first preset threshold, determining that the target digital content and the reference digital content are of the same source. If the feature similarity is less than or equal to the first preset threshold, determining that the target digital content and the reference digital content are of different sources.

[0158] In this embodiment, the reference digital content is a comparison object specified by the user, such as a reference image specified by the user that is infringing. The second data feature can characterize the source information of the reference digital object, and the second data feature of the reference digital object can be extracted by the aforementioned feature extraction model. For example, when the user provides a reference image, the reference image can be input into the aforementioned feature extraction model to extract the second image feature of the reference image. Then the feature similarity between the first data feature and the second data feature is calculated. The calculation of feature similarity can effectively capture the potential association between the target digital content and the reference digital content. When the feature similarity is greater than the first preset threshold, the system can determine with high confidence that the target digital content and the reference digital content are homologous. Conversely, when the feature similarity is less than or equal to the first preset threshold, the system can accurately identify that the target digital content and the reference digital content are not homologous, thereby avoiding erroneous homology classification. This judgment method based on feature similarity not only improves the discrimination ability of the system, but also enhances its adaptability in a diverse and complex data environment, and provides solid technical support for applications such as content review, copyright protection and information traceability.

[0159] The first preset threshold can be selected according to actual needs, for example, by analyzing the similarity distribution in the training data set and observing the similarity values ​​of samples from the same source and different sources. Usually, a threshold that can effectively distinguish the two types of samples can be selected. The threshold can also be adjusted according to the business needs of the specific application scenario and the tolerance for misjudgment. For example, in some application scenarios, it may be more inclined to reduce false negatives (missed reports), while in other application scenarios, it may be more concerned with reducing false positives (false positives).

[0160] Optionally, feature similarity calculation methods include but are not limited to: Euclidean distance, cosine similarity, Manhattan distance, Hamming distance, etc.

[0161] like Figure 8 As shown, it is a flow chart of a digital content detection method provided by the present application. Taking image detection as an example, assuming that the user provides a reference image and a prompt word, the input information of the detection task is the image to be detected (i.e., the target product image) and its target prompt word, the reference image and its prompt word. The process may include the following steps:

[0162] Step 801: Determine whether the input information contains a prompt word. If yes, proceed to step 802; otherwise, proceed to step 805 to directly execute the source tracing matching branch.

[0163] Step 802: Determine whether the prompt word of the image to be tested is the same as the prompt word of the reference image. If so, proceed to step 803 to execute the source tracing classification branch. Assuming that the prompt word of the image to be tested is "a knight in a space suit riding a Ferghana horse walking on the moon", and the prompt word of the reference image is "a knight in a space suit riding a Ferghana horse walking on the moon", it can be seen that the prompt words are the same, and the process proceeds to step 803. Otherwise, the process proceeds to step 808.

[0164] Step 803: The AI ​​generation model tracing classification model (i.e., the preset classification model) obtained by training in the above-mentioned embodiment is used to respectively identify the tracing results of the image to be tested and the tracing results of the reference image.

[0165] Step 804: Determine whether the traceability result of the image to be tested is the same as the traceability result of the reference image. If so, proceed to step 807; otherwise, proceed to step 808.

[0166] Step 805: Use the CLIP image feature extractor fine-tuned by the above embodiment to extract the first image feature of the image to be tested and the second image feature of the reference image respectively, and calculate the feature distance (i.e., feature similarity) between the first image feature and the second image feature, and then enter step 806.

[0167] Step 806: Determine whether the feature distance is close, that is, determine whether the feature distance is greater than a first preset threshold. If so, proceed to step 807; otherwise, proceed to step 808.

[0168] Step 807: Determine whether the image to be tested and the reference image are homologous images.

[0169] Step 808: Determine whether the image to be tested and the reference image are non-homologous images.

[0170] The method of this embodiment can achieve the traceability of the target digital content to be tested and the reference digital content when the user provides the reference digital content, and can accurately match the specific generation source or model method of the digital content. The target digital content to be tested includes but is not limited to AI-generated images or texts, and the reference digital content includes but is not limited to reference images or texts.

[0171] In one embodiment, there are multiple reference digital contents, and the second data feature and corresponding metadata of each reference content are stored in a preset database. Determining whether the target digital content and the reference digital content are of the same origin based on the first data feature and the reference digital content includes: calculating the feature similarity between the first data feature and the second data feature of each reference digital content in the preset database, respectively. Determining whether the target digital content and the reference digital content are of the same origin based on the feature similarity and the metadata corresponding to each reference content.

[0172] In this embodiment, the reference digital content can be multiple and can be stored in a preset database, and the preset database can store the second data feature and corresponding metadata of each reference digital content. The metadata includes but is not limited to the source information of the reference digital content, such as which model the reference digital content is generated by, what the text prompt words used in the generation process are, the version information of the model, and the platform information to which the model belongs. By pre-storing the second data feature of the reference digital content in the preset database, the amount of data calculation for real-time reasoning is reduced. When implementing reasoning, the feature similarity between the first data feature of the target digital content and the second data feature of each reference digital content in the database is first calculated. The process of calculating the feature similarity one by one ensures a comprehensive analysis of all potential reference contents and avoids missing possible homologous contents. Subsequently, by combining the feature similarity and the metadata corresponding to each reference content, the system can more accurately determine whether the target digital content and the reference digital content are homologous. This strategy of combining metadata provides additional contextual information, which helps to improve the accuracy of judgment. In addition, the method optimizes the speed of data retrieval and calculation by utilizing the structured storage of the preset database, so that the system can run efficiently in a large-scale data environment. It not only enhances the capabilities of content review and copyright protection, but also provides strong support for applications such as information tracing and content recommendation.

[0173] In one embodiment, determining whether the target digital content and the reference digital content are of the same origin based on the feature similarity and the metadata corresponding to each reference content includes: determining a set of similar digital content whose feature similarity with the target digital content is greater than a second preset threshold from a plurality of reference digital contents based on the feature similarity. Comparing the metadata of the target digital content with the metadata of each digital content in the set of similar digital contents respectively. If there is a specific digital content with the same metadata as the target digital content in the set of similar digital contents, determining that the target digital content and the specific digital content are of the same origin. And / or, outputting the specific digital content. If there is no specific digital content with the same metadata as the target digital content in the set of similar digital contents, determining that the target digital content and the reference digital content are of different origins.

[0174] In this embodiment, by combining feature similarity and metadata for homology judgment, the method significantly improves the accuracy and reliability of homology identification between target digital content and reference digital content. Specifically, first, based on feature similarity, a set of similar digital content whose feature similarity with the target digital content is greater than a second preset threshold is screened out from multiple reference digital contents. This step ensures that only highly similar content is further analyzed, reducing the computational burden. Then, by comparing the metadata of the target digital content with the metadata of each digital content in the similar digital content set, the system can use the contextual information provided by the metadata to make a more refined homology judgment. If there is a specific digital content with the same metadata as the target digital content in the similar digital content set, it can be determined with high confidence that the target digital content is homologous to the specific digital content, and the specific digital content can be output. This strategy of combining metadata effectively reduces the possibility of misjudgment and ensures the accuracy of homology identification.

[0175] On the contrary, if there is no specific digital content with the same metadata as the target digital content in the similar digital content set, the system can accurately determine that the target digital content and the reference digital content are not from the same source. This not only improves the ability of content review and copyright protection, but also provides reliable technical support for applications such as information tracing and content recommendation, and enhances the adaptability and practicality of the system in complex data environments.

[0176] The second preset threshold value may be the same as the first preset threshold value, or may be different from the first preset threshold value, and may be selected according to actual needs. The description of the first preset threshold value may be referred to, and will not be repeated here.

[0177] like Fig. 9 FIG. 1 is a flow chart of a digital content detection method provided by an embodiment of the present application. Taking image detection as an example, assuming that the user does not provide a reference image, a reference image is pre-set in a preset database (i.e. Fig. 9The base database in the image database stores the relevant information of multiple reference images. For example, the reference digital content stored in the base database may include multiple AI-generated images with copyright protection certification. The feature vector of each AI-generated image and its bound json metadata (json_meta) are saved in the base database, where the json_meta information may include the name of the generated model (model_name) and the generated prompt word (prompt). For example, the json_meta information of base database image 1 includes the name of the generated model "SDv1.4", and the prompt word is "a knight in a space suit riding a Ferghana horse walking on the moon", the json_meta information of base database image 2 includes the name of the generated model "SDXL", and the prompt word is "a knight in a space suit riding a Ferghana horse walking on the moon", the json_meta information of base database image 3 includes the name of the generated model "SD3", and the prompt word is "Sun Wukong from Journey to the West flying over Wall Street", ..., the json_meta information of base database image n includes the name of the generated model "Midjouney", and the prompt word is "a wooden sailboat sailing in a rough sea".

[0178] The feature vector of each AI-generated image in the base library can be extracted by using the CLIP image feature extractor fine-tuned in any of the above embodiments and stored as the base library feature. Assume that the input information of the detection task is the image to be tested and its prompt word. The detection process for the image to be tested can include the following steps:

[0179] Step 901: Use the CLIP image feature extractor fine-tuned in any of the above embodiments to extract a feature vector of the image to be tested.

[0180] Step 902: Calculate the feature distance between the feature vector of the image to be tested and the feature vectors of all AI-generated images in the database.

[0181] Step 903: filtering the base library images according to the feature distance, and selecting a list of similar images whose feature distance is greater than a second preset threshold from the base library.

[0182] Step 904: Check one by one whether the json_meta information of the image to be tested is consistent with that of all images in the similar image list. If the model_name and prompt in the json_meta information of the base library image are the same as those of the image to be tested, go to step 905; otherwise, go to step 906.

[0183] Step 905: Determine the base library images in the similar image list whose json_meta information is the same as that of the image to be tested as homologous images of the image to be tested. The homologous images may include multiple ones, and a homologous image list may be output.

[0184] Step 906: Determine the base library images in the similar image list whose json_meta information is different from that of the image to be tested as non-homologous images of the image to be tested. The non-homologous images may include multiple ones, and a list of non-homologous images may be output.

[0185] The method of this embodiment can trace the AI-generated image or text to be tested, even if the user does not provide reference digital content, and can accurately match it to its specific generation source or model method.

[0186] The above-mentioned digital content detection method, for two AI-generated images and their corresponding text prompts, uses the AI-generated image source tracing classification algorithm and the multimodal AI-generated image source tracing matching algorithm to determine whether the two images are generated by the AI-generated model with the same source and the same prompt, and traces the generation method at the same time. The AI-generated image source tracing task is completed by jointly using the AI-generated model source tracing classification algorithm and the AI-generated model source tracing matching algorithm. In the AI-generated model source tracing classification algorithm, whether the two AI-generated images are of the same source is determined by predicting whether the classification categories are consistent. In the AI-generated image source tracing matching algorithm, whether the two AI-generated images are of the same source is determined by calculating the similarity of the feature vectors of the two AI-generated images. It can not only solve the copyright ownership problem of AI-generated content, but also greatly improve the management and copyright protection capabilities of generated images.

[0187] The technical solution of the embodiment of the present application has important applications in copyright protection, and its application scenarios include but are not limited to the following scenarios:

[0188] 1. Platform governance of original image AIGC detection: In actual scenarios, there are post-processing of copyrighted model images and clothing images through AIGC technology synthesis, attempting to bypass copyright protection detection. This kind of plagiarism is difficult to identify with the naked eye, but the detection technology of this application solution can effectively identify AI-synthesized images, thereby preventing copyright infringement and protecting the legitimate rights and interests of creators.

[0189] 2. Platform AIGC image content identification obligations: According to relevant Internet information service management regulations, platforms must clearly identify content including AIGC images. Since images posted by users may come from a variety of sources, platforms can identify AI synthetic images based on the technical solution of this application and mark the corresponding content to ensure information transparency and users' right to know.

[0190] 3. Traffic diversion and offline information security prevention: Through the AI-generated image tracing detection technology of the embodiment of the present application, it is possible to trace the source and determine which model platform the AI-generated image came from, so as to identify information security risks in advance and prevent potential information losses. At the same time, it can also inform the platform that provides AI-generated images to strengthen the restriction and interception of unsafe generated images.

[0191] 4. Advertising and brand protection: In actual scenarios, the advertising industry can use AI-generated images to promote brands. However, there may be a phenomenon of using AI to edit advertising images of famous brands, causing damage to the brand image. Through the multimodal AI-generated image traceability matching method provided by this application, false advertising images can be identified and processed early to protect brand reputation.

[0192] 5. Anti-counterfeiting of e-commerce platforms: Online shopping platforms need to ensure the authenticity of product images. The AI-generated image traceability detection technology provided in the embodiments of this application can identify and trace the source of product images, protect consumer rights, and enhance the trust of the platform.

[0193] 6. Traceability matching of maliciously reposted UGC content: User-generated content (UGC) may be maliciously reposted on social media and content platforms, resulting in damage to the rights of the original author. The AI-generated image traceability detection technology provided by this application can accurately identify and trace the source of reposted UGC content, ensure the rights of the original author, prevent the spread of bad content, and improve information security.

[0194] For details of each step of the above method, please refer to the relevant description of the above embodiment, which will not be repeated here.

[0195] Please see Fig.10 , which is a method for detecting a commodity image according to an embodiment of the present application, the method can be Figure 1 The electronic device 1 shown is used to perform and can be applied to Figure 2 In the digital content detection application scenario shown in , it is used to generate a product image with a watermark. In this embodiment, the terminal 220 is used as an execution end, and the product image scene in the e-commerce scene is used as an example. The method includes the following steps:

[0196] Step 1001: In response to a detection task, a target product image to be processed and a reference image are acquired, where the target product image includes product information.

[0197] In this step, the product information includes but is not limited to product-related knowledge information, appearance information, product promotion information, product-related order information, product-related service information, and the like.

[0198] Step 1002: Input the target product image into a preset feature extraction model so that the feature extraction model outputs a first image feature of the target product image, wherein the feature extraction model is trained by a first sample image set, and the first sample image set includes at least one sample image pair, and the sample image pairs with the same sample prompt word and the same generation model are marked as homologous labels.

[0199] Step 1003: Determine whether the target product image and the reference image are of the same source based on the first image feature and the reference image.

[0200] The above-mentioned commodity image detection method automatically extracts a representative first image feature from the target commodity image by using a pre-trained feature extraction model, and the first image feature can characterize the source information of the target commodity image. Since the feature extraction model is trained based on the first sample image set, in the first sample image set, sample image pairs with the same sample prompt word and the same generation model are annotated as homologous labels, and both the prompt word and the generation model can characterize the source information of the image, the trained feature extraction model can effectively capture the first image feature that accurately characterizes the source information in the target commodity image, and then can effectively verify whether the target commodity image and the reference image are homologous, thereby improving the accuracy of image traceability and matching detection, and providing reliable data support for the subsequent protection of the ownership of commodity images.

[0201] For details of each step of the above method, please refer to the relevant description of the above embodiment, which will not be repeated here.

[0202] Please see Fig.11 , which is a digital content detection device 1100 of an embodiment of the present application, the device can be applied to Figure 1 The electronic device 1 shown can be applied to Figure 2 In the digital content detection application scenario shown in , it is possible to effectively verify whether the target digital content and the reference digital content are of the same origin, thereby improving the accuracy of digital content traceability and matching detection, and providing reliable data support for the subsequent digital content ownership protection. The device includes: an acquisition module 1101, an extraction module 1102 and a determination module 1103, and the functional principles of each module are as follows:

[0203] The acquisition module 1101 is used to acquire the target digital content and reference digital content to be processed in response to the detection task.

[0204] The extraction module 1102 is used to input the target digital content into a preset feature extraction model so that the feature extraction model outputs the first data feature of the target digital content, wherein the feature extraction model is obtained by training a first sample data set, and the first sample data set includes at least one sample digital content pair, and the sample digital content pairs with the same sample prompt word and the same generation model are marked as homologous labels.

[0205] The determination module 1103 is used to determine whether the target digital content and the reference digital content are of the same source according to the first data feature and the reference digital content.

[0206] In one embodiment, the method further includes: a first judgment module, which is used to judge whether there is a target prompt word used to generate the target digital content in the input information of the detection task before the target digital content is input into the preset feature extraction model so that the feature extraction model outputs the first data feature of the target digital content. If there is no target prompt word in the input information, the step of inputting the target digital content into the preset feature extraction model so that the feature extraction model outputs the first data feature of the target digital content is executed.

[0207] In one embodiment, it further includes: a second judgment module, which is used to judge whether the target prompt word is the same as the reference prompt word corresponding to the reference digital content if the target prompt word exists in the input information. A classification module, which is used to input the target digital content and the reference digital content into a preset classification model respectively if the target prompt word is the same as the reference prompt word, so that the preset classification model outputs the first source information of the target digital content and the second source information of the reference digital content. The classification module is also used to determine whether the target digital content and the reference digital content are of the same source according to the first source information and the second source information.

[0208] In one embodiment, the method further includes: a first training module, which is used to obtain a second sample data set before inputting the target digital content and the reference digital content into the preset classification model respectively so that the preset classification model outputs the first source information of the target digital content and the second source information of the reference digital content, wherein the second sample data set includes a plurality of sample digital contents generated by different generation models, and each sample digital content is marked with information of the corresponding generation model. According to the second sample data set, a preset neural network is trained using a cross entropy loss function to obtain a preset classification model.

[0209] In one embodiment, the reference digital content is a comparison object specified by a user. The determination module 1103 is used to obtain a second data feature of the reference digital content. The feature similarity between the first data feature and the second data feature is calculated. If the feature similarity is greater than a first preset threshold, it is determined that the target digital content and the reference digital content are of the same source. If the feature similarity is less than or equal to the first preset threshold, it is determined that the target digital content and the reference digital content are of different sources.

[0210] In one embodiment, there are multiple reference digital contents, and the second data feature and corresponding metadata of each reference digital content are stored in a preset database. The determination module 1103 is used to calculate the feature similarity between the first data feature and the second data feature of each reference digital content in the preset database. According to the feature similarity and the metadata corresponding to each reference content, it is determined whether the target digital content and the reference digital content are of the same source.

[0211] In one embodiment, the determination module 1103 is specifically used to determine a set of similar digital contents whose feature similarity with the target digital content is greater than a second preset threshold value from a plurality of reference digital contents according to the feature similarity. The metadata of the target digital content is respectively compared with the metadata of each digital content in the similar digital content set. If there is a specific digital content with the same metadata as the target digital content in the similar digital content set, it is determined that the target digital content and the specific digital content are of the same origin. And / or, the specific digital content is output. If there is no specific digital content with the same metadata as the target digital content in the similar digital content set, it is determined that the target digital content and the reference digital content are not of the same origin.

[0212] In one embodiment, it also includes: a data construction module, which is used to obtain multiple different sample prompt words before inputting the target digital content into a preset feature extraction model so that the feature extraction model outputs the first data feature of the target digital content. Obtain multiple different content generation models. Input multiple different sample prompt words into each content generation model one by one, so that each content generation model generates corresponding sample digital content according to multiple different sample prompt words, and obtain an original sample data set. Pair the sample digital contents in the original sample data set in pairs to form multiple sample digital content pairs, and mark the sample digital content pairs with the same sample prompt words and the same generation model as homologous labels to form a final first sample data set.

[0213] In one embodiment, it also includes: a second training module, which is used to fine-tune the training of a preset feature extractor using a contrast loss function according to a first sample data set before inputting the target digital content into a preset feature extraction model so that the feature extraction model outputs the first data feature of the target digital content to obtain a feature extraction model.

[0214] In one embodiment, the first sample data set is a sample image set. It also includes: a third training module, which is used to fine-tune and train a preset feature extractor according to the sample image set before inputting the target digital content into the preset feature extraction model so that the feature extraction model outputs the first data feature of the target digital content, wherein the feature extractor includes an image feature extractor and a text feature extractor. During the training process, the image feature extractor and the text feature extractor are fine-tuned and trained using the inter-class contrast loss function and the image-text contrast loss function respectively to obtain a feature extraction model.

[0215] For a detailed description of the digital content detection device 1100 , please refer to the description of the relevant method steps in the above embodiment. The implementation principle and technical effect are similar and will not be described in detail in this embodiment.

[0216] Fig.12 The following is a schematic diagram of a cloud device 120 provided in an exemplary embodiment of the present application. The cloud device 120 can be used to run the method provided in any of the above embodiments. Fig.12 As shown, the cloud device 120 may include: a memory 1204 and at least one processor 1205, Fig.12 A processor is taken as an example.

[0217] The memory 1204 is used to store computer programs and can be configured to store various other data to support operations on the cloud device 120. The memory 1204 can be an object storage service (OSS).

[0218] Memory 1204 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0219] The processor 1205 is coupled to the memory 1204 and is used to execute the computer program in the memory 1204 to implement the solution provided by any of the above method embodiments. The specific functions and technical effects that can be achieved are not repeated here.

[0220] Furthermore, if Fig.12 The cloud device also includes: a firewall 1201, a load balancer 1202, a communication component 1206, a power supply component 1203 and other components. Fig.12 Only some components are shown schematically, which does not mean that the cloud device only includes Fig.12 Components shown.

[0221] In one embodiment, the above Fig.12 The communication component 1206 in is configured to facilitate wired or wireless communication between the device where the communication component 1206 is located and other devices. The device where the communication component 1206 is located can access a wireless network based on a communication standard, such as WiFi, 2G, 3G, 4G, LTE (Long Term Evolution, Long Term Evolution, referred to as LTE), 5G and other mobile communication networks, or a combination thereof. In an exemplary embodiment, the communication component 1206 receives a broadcast signal or broadcast-related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 1206 also includes a near field communication (Near Field Communication, referred to as NFC) module to facilitate short-range communication. For example, the NFC module can be based on Radio Frequency Identification (Radio Frequency Identification, referred to as RFID) technology, Infrared Data Association (Infrared Data Association, referred to as IrDA) technology, Ultra Wide Band (UWB) technology, Bluetooth (bluetooth, referred to as BT) technology and other technologies to achieve.

[0222] In one embodiment, the above Fig.12 The power supply component 1203 provides power to various components of the device where the power supply component 1203 is located. The power supply component 1203 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the device where the power supply component is located.

[0223] An embodiment of the present application further provides a computer-readable storage medium, in which computer-executable instructions are stored. When a processor executes the computer-executable instructions, the method of any of the aforementioned embodiments is implemented.

[0224] An embodiment of the present application also provides a computer program product, including a computer program, which implements the method of any of the aforementioned embodiments when executed by a processor.

[0225] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of modules is only a logical function division, and there may be other division methods in actual implementation, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0226] The above-mentioned integrated module implemented in the form of a software function module can be stored in a computer-readable storage medium. The above-mentioned software function module is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to perform some steps of the methods of various embodiments of the present application.

[0227] It should be understood that the above-mentioned processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in the application can be directly embodied as being executed by a hardware processor, or can be executed by a combination of hardware and software modules in the processor. The memory may include a high-speed RAM (Random Access Memory) memory, and may also include non-volatile storage NVM (Nonvolatile memory, NVM for short), such as at least one disk storage, and can also be a USB flash drive, a mobile hard disk, a read-only memory, a disk or an optical disk, etc.

[0228] The above storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The storage medium can be any available medium that can be accessed by a general or special computer.

[0229] An exemplary storage medium is coupled to a processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can be located in an application specific integrated circuit (ASIC). Of course, the processor and the storage medium can also exist as discrete components in an electronic device or a main control device.

[0230] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, clothing or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, clothing or device. In the absence of more restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, clothing or device including the element.

[0231] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0232] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a disk, or an optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods of each embodiment of the present application.

[0233] In the technical solution of this application, the collection, storage, use, processing, transmission, provision and disclosure of user data and other information involved are in compliance with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0234] The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A digital content detection method, characterized in that: include: In response to the detection task, obtaining target digital content and reference digital content to be processed; Inputting the target digital content into a preset feature extraction model so that the feature extraction model outputs a first data feature of the target digital content, wherein the feature extraction model is obtained by training a first sample data set, the first sample data set includes at least one sample digital content pair, and the sample digital content pairs with the same sample prompt word and the same generation model are marked as homologous labels; According to the first data feature and the reference digital content, it is determined whether the target digital content and the reference digital content have the same origin.

2. The method according to claim 1, characterized in that Before inputting the target digital content into a preset feature extraction model so that the feature extraction model outputs the first data feature of the target digital content, the method further includes: Determining whether there is a target prompt word used to generate the target digital content in the input information of the detection task; If the target prompt word does not exist in the input information, the step of inputting the target digital content into a preset feature extraction model is performed so that the feature extraction model outputs a first data feature of the target digital content.

3. The method according to claim 2, characterized in that Also includes: If the target prompt word exists in the input information, determining whether the target prompt word is the same as the reference prompt word corresponding to the reference digital content; If the target prompt word is the same as the reference prompt word, the target digital content and the reference digital content are respectively input into a preset classification model, so that the preset classification model outputs first source information of the target digital content and second source information of the reference digital content; It is determined whether the target digital content and the reference digital content have the same source according to the first source information and the second source information.

4. The method according to claim 3, characterized in that Before respectively inputting the target digital content and the reference digital content into a preset classification model so that the preset classification model outputs the first source information of the target digital content and the second source information of the reference digital content, the method further includes: Acquire a second sample data set, wherein the second sample data set includes a plurality of sample digital contents generated by different generation models, each of the sample digital contents is marked with information of a corresponding generation model; According to the second sample data set, a preset neural network is trained using a cross entropy loss function to obtain the preset classification model.

5. The method according to claim 1, characterized in that The reference digital content is a comparison object specified by a user; and determining whether the target digital content and the reference digital content are of the same origin according to the first data feature and the reference digital content includes: Acquiring a second data feature of the reference digital content; Calculating feature similarity between the first data feature and the second data feature; If the feature similarity is greater than a first preset threshold, determining that the target digital content and the reference digital content are of the same source; If the feature similarity is less than or equal to the first preset threshold, it is determined that the target digital content and the reference digital content are from different sources.

6. The method according to claim 1, characterized in that There are a plurality of reference digital contents, and the second data feature and corresponding metadata of each of the reference digital contents are stored in a preset database; determining whether the target digital content and the reference digital content are of the same origin according to the first data feature and the reference digital contents comprises: respectively calculating the feature similarity between the first data feature and the second data feature of each of the reference digital contents in the preset database; According to the feature similarity and the metadata corresponding to each of the reference contents, it is determined whether the target digital content and the reference digital content are of the same origin.

7. The method according to claim 6, characterized in that The step of determining whether the target digital content and the reference digital content are of the same origin according to the feature similarity and the metadata corresponding to each of the reference contents includes: Determine, from the plurality of reference digital contents according to the feature similarity, a set of similar digital contents whose feature similarity to the target digital content is greater than a second preset threshold; respectively comparing the metadata of the target digital content with the metadata of each digital content in the similar digital content set; If there is a specific digital content having the same metadata as the target digital content in the similar digital content set, determining that the target digital content and the specific digital content are of the same origin; and / or outputting the specific digital content; If there is no specific digital content having the same metadata as the target digital content in the similar digital content set, it is determined that the target digital content and the reference digital content have different sources.

8. The method according to claim 1, characterized in that: Before inputting the target digital content into a preset feature extraction model so that the feature extraction model outputs the first data feature of the target digital content, the method further includes: Get multiple different sample prompt words; Obtain multiple different content generation models; Inputting the plurality of different sample prompt words into each of the content generation models one by one, so that each of the content generation models generates corresponding sample digital content according to the plurality of different sample prompt words, to obtain an original sample data set; The sample digital contents in the original sample data set are paired in pairs to form a plurality of sample digital content pairs, and the sample digital content pairs with the same sample prompt words and the same generation model are marked as homologous labels to form the final first sample data set.

9. The method according to claim 1 or 8, characterized in that: Before inputting the target digital content into a preset feature extraction model so that the feature extraction model outputs the first data feature of the target digital content, the method further includes: According to the first sample data set, a contrast loss function is used to fine-tune and train a preset feature extractor to obtain the feature extraction model.

10. The method according to claim 1 or 8, characterized in that: The first sample data set is a sample image set; before the target digital content is input into a preset feature extraction model so that the feature extraction model outputs the first data feature of the target digital content, the method further includes: Fine-tune and train a preset feature extractor according to the sample image set, wherein the feature extractor includes an image feature extractor and a text feature extractor; During the training process, the image feature extractor and the text feature extractor are fine-tuned and trained using an inter-class contrast loss function and an image-text contrast loss function, respectively, to obtain the feature extraction model.

11. A method for detecting a commodity image, characterized in that: include: In response to the detection task, acquiring a target commodity image to be processed and a reference image, wherein the target commodity image includes commodity information; Inputting the target product image into a preset feature extraction model so that the feature extraction model outputs a first image feature of the target product image, wherein the feature extraction model is trained by a first sample image set, the first sample image set includes at least one sample image pair, and the sample image pairs having the same sample prompt word and the same generation model are annotated as homologous labels; According to the first image feature and the reference image, it is determined whether the target product image and the reference image are of the same origin.

12. An electronic device, characterized in that: include: at least one processor; as well as a memory communicatively coupled to the at least one processor; The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the electronic device to perform the method described in any one of claims 1-11.

13. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and when the processor executes the computer-executable instructions, the method according to any one of claims 1 to 11 is implemented.

14. A computer program product, characterized in that The method comprises a computer program, which implements the method according to any one of claims 1 to 11 when being executed by a processor.

Images