A data processing method, device, apparatus, and storage medium

Abstract

The application provides a data processing method, device and equipment and a storage medium. The method comprises the following steps: receiving sensitive data of at least one mode and projecting the sensitive data to a unified semantic space to generate corresponding sensitive features; identifying display data of a screen and projecting the display data to obtain display features; matching the display features and the sensitive features of each mode to obtain corresponding mode risk values; if any mode risk value exceeds a local threshold value, shielding the display data of the mode; if there are two or more modes and each mode risk value is lower than the local threshold value, calculating a semantic consistency score of the multi-mode display features, determining a synergistic gating coefficient, and combining the mode risk values to obtain a joint risk score; and when the joint risk score is higher than a global threshold value, visually shielding all mode display data. The application realizes individualized and semantic-level sensitive content identification and protection for different mode contents, and significantly improves the fine degree of identification and the protection pertinence.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a data processing method, apparatus, device, and storage medium. Background Technology

[0002] Current mainstream internet content filtering methods mostly adopt a unified global control strategy for the platform, which is difficult to meet the personalized protection needs of different users. At the same time, they lack the ability to understand cross-modal semantics of multimodal content such as text, images and videos, making it difficult to accurately identify sensitive content that may cause psychological stimulation or induce negative emotions in specific users, resulting in insufficient precision in identification and targeted protection. Summary of the Invention

[0003] This application provides a data processing method, apparatus, device, and storage medium to at least solve the above-mentioned technical problems existing in the prior art.

[0004] A first aspect of this application provides a data processing method, the method comprising: Receive sensitive data for at least one modality input, project the sensitive data of each modality onto a unified semantic space, and obtain the sensitive features corresponding to each modality; The display data of each modality on the screen is identified, and the display data of each modality is projected onto the unified semantic space to obtain the display features of each modality. The display features of each modality are matched with all the sensitive features corresponding to each modality to obtain the matching results; Determine the risk value for each modality based on the matching results; In response to a risk value for any modality exceeding the corresponding local threshold, visual masking is performed on the display data for the corresponding modality. In response to the existence of at least two modalities and the risk values ​​of all modalities being less than the corresponding local thresholds, the semantic consistency scores of the display features of at least two modalities in the unified semantic space are calculated. The cooperative gating coefficients are determined based on the semantic consistency scores. The joint risk score is determined based on the cooperative gating coefficients and the risk values ​​of each modality. In response to the joint risk score being greater than the global threshold, visual masking processing is performed on the display data of all modalities.

[0005] In one possible implementation, projecting the sensitive data of each modality onto a unified semantic space includes: Determine the feature extraction model and the mapping model corresponding to each modality; By using the feature extraction models corresponding to each modality, the sensitive data of each modality are extracted to obtain the initial sensitive features of each modality; By using the mapping model corresponding to each modality, the initial sensitive features of each modality are mapped to the unified semantic space to obtain the sensitive features corresponding to each modality.

[0006] In one possible implementation, the mapping model is trained through the following operations: The mapping layer weights corresponding to the modalities in the target cross-modal model are used to initialize the training mapping model for each modality. Multiple sensitive sample pairs are obtained, wherein the sensitive sample pairs include anchor sensitive data, positive sample sensitive data that matches the anchor sensitive data, and negative sample sensitive data that does not match the anchor sensitive data; The sensitive sample pairs are respectively input into the initialized mapping models of each modality to be trained, so as to map them to the unified semantic space; With the goal of minimizing the distance between anchor-sensitive data and positive sample-sensitive data in the unified semantic space, and maximizing the distance between anchor-sensitive data and negative sample-sensitive data in the unified semantic space, the parameters of the mapping model to be trained for each modality are updated to obtain the trained mapping model.

[0007] In one possible implementation, the at least one modality includes a text modality and / or a visual modality; The display data for each modality on the identification screen includes: The screen display content of multiple consecutive frames is obtained, and dirty rectangle detection is performed on the screen display content of multiple consecutive frames to obtain the dirty rectangle region. In response to the detection of text content at the screen position corresponding to the dirty rectangular area, the text content is extracted and used as the display data of the text modality; In response to the detection of static image content at the screen location corresponding to the dirty rectangular area, the static image content is captured and used as display data for the visual modality; or, in response to the detection that the screen location corresponding to the dirty rectangular area is a video playback area, a video frame sequence is captured from the video playback area at a target acquisition frequency and used as display data for the visual modality; wherein, the target acquisition frequency is determined based on the risk value of the text modality and the degree of screen change in the video playback area.

[0008] In one possible implementation, projecting the display data of each modality onto the unified semantic space to obtain the display features of each modality includes: In response to the fact that the displayed data is text content, initial text features are extracted from the text content by a text feature encoder and projected onto the unified semantic space to obtain text features as the display features of the text modality; In response to the display data being static image content, initial image features are extracted from the static image content using an image feature encoder and projected onto the unified semantic space to obtain image features as display features of the visual modality; or, in response to the display data being a video frame sequence, initial video features are extracted from each video frame in the video frame sequence using an image feature encoder and projected onto the unified semantic space to obtain video features corresponding to each video frame, and one video feature is selected from all video features as the display feature of the visual modality.

[0009] In one possible implementation, selecting one video feature from all video features as the display feature of the visual modality includes: The video features corresponding to each video frame in the video frame sequence are matched with all sensitive features corresponding to the visual modality to obtain the risk value of each video frame. The video features corresponding to the video frame with the highest risk value are selected as the display features of the visual modality.

[0010] In one possible implementation, matching the display features of each modality with all the sensitive features corresponding to each modality to obtain the matching result includes: Calculate the similarity between the displayed features of each modality and all the sensitive features corresponding to each modality, and use the maximum similarity as the risk value of each modality.

[0011] In one possible implementation, determining the collaborative gating coefficient based on the semantic consistency score, and determining the joint risk score based on the collaborative gating coefficient and each modality risk value, includes: The semantic consistency score is compared with zero, and the maximum value is taken as the collaborative gating coefficient; Based on the risk value of each modality, determine the attention weight of each modality; Based on the attention weights of each modality, the risk values ​​of each modality are weighted and fused to obtain a weighted fusion term; Multiplying the weighted fusion term by the collaborative gating coefficient yields the joint reward; The collaborative reward is determined based on the minimum value among the collaborative gating coefficient, the preset reward coefficient, and the risk values ​​of each modality. The joint risk score is determined based on the joint reward and the synergistic reward.

[0012] In one possible implementation, the method further includes: In response to detecting that the user has marked the display data for which visual masking is performed as a false alarm, the corresponding local or global threshold is increased; In response to detecting that a user has marked display data for which visual masking has not been performed as a missed report, the corresponding local or global threshold is reduced.

[0013] In one possible implementation, the method further includes: The parameters of the mapping model are updated in response to any of the following: The sum of the number of display data that the user marked as false alarms after performing visual masking and the number of display data that the user marked as missed alarms after not performing visual masking reaches a preset number; The time since the last mapping model update exceeds the preset time; Within a preset sliding window, the proportion of users marking display data that has undergone visual masking as false alarms or display data that has not undergone visual masking as missed alarms exceeds a preset proportion.

[0014] In one possible implementation, the method further includes: In response to the detection that available memory is below the memory threshold, the visual modality is disabled, and only the text modality is enabled; In response to screen capture permission being denied; or, if the number of consecutive failures to acquire display data for any mode exceeds a preset number, the corresponding mode will be disabled; If the single-frame inference delay of the image feature encoder continues to exceed a preset duration and reaches a preset number of frames consecutively, the visual modality is disabled. In response to the detection of a shutdown command for any modality, the corresponding modality is disabled.

[0015] A second aspect of this application provides a data processing apparatus, the apparatus comprising: The first projection module is used to receive sensitive data for at least one modality input, project the sensitive data of each modality onto a unified semantic space, and obtain the sensitive features corresponding to each modality. The second projection module is used to identify the display data of each modality on the screen, and project the display data of each modality onto the unified semantic space to obtain the display features of each modality. The matching module is used to match the display features of each modality with all the sensitive features corresponding to each modality to obtain the matching results; The determination module is used to determine the risk value of each modality based on the matching results; The first shielding module is used to perform visual shielding processing on the display data of the corresponding modality in response to the risk value of any modality being greater than the corresponding local threshold. The second masking module is used to calculate the semantic consistency score of the display features of at least two modalities in the unified semantic space in response to the existence of at least two modalities and the risk value of all modalities being less than the corresponding local threshold, determine the cooperative gating coefficient based on the semantic consistency score, determine the joint risk score based on the cooperative gating coefficient and the risk value of each modality, and perform visual masking processing on the display data of all modalities in response to the joint risk score being greater than the global threshold.

[0016] A third aspect of this application provides an electronic device comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method described in this application.

[0017] A fourth aspect of this application provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform the methods described in this application.

[0018] The data processing method, apparatus, device, and storage medium of this application receive sensitive data for at least one modality input, project the sensitive data of each modality onto a unified semantic space to obtain sensitive features corresponding to each modality; identify display data of each modality on the screen, project the display data of each modality onto the unified semantic space to obtain display features of each modality; match the display features of each modality with all sensitive features corresponding to each modality to obtain matching results; determine the risk value of each modality based on the matching results; in response to the risk value of any modality being greater than the corresponding local threshold, perform visual masking processing on the display data of the corresponding modality; in response to the existence of at least two modalities and the risk values ​​of all modalities being less than the corresponding local thresholds, calculate the semantic consistency score of the display features of at least two modalities in the unified semantic space, determine a cooperative gating coefficient based on the semantic consistency score, determine a joint risk score based on the cooperative gating coefficient and the risk values ​​of each modality, and in response to the joint risk score being greater than a global threshold, perform visual masking processing on the display data of all modalities. It enables personalized, semantically sensitive content identification and protection for different modalities, significantly improving the precision of identification and the targeted nature of protection.

[0019] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0020] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of this application are illustrated in the drawings by way of example and not limitation, in which: In the accompanying drawings, the same or corresponding reference numerals indicate the same or corresponding parts.

[0021] Figure 1 A schematic diagram illustrating the implementation flow of the data processing method provided in an embodiment of this application is shown; Figure 2 This illustration shows a schematic diagram of the implementation flow of the mapping model training operation of the data processing method provided in the embodiments of this application; Figure 3 This illustration shows a schematic diagram of the implementation flow of the display data recognition operation of the data processing method provided in the embodiments of this application; Figure 4 A schematic diagram illustrating the implementation flow of the feature extraction operation of the data processing method provided in this application embodiment is shown; Figure 5 This illustration shows a schematic diagram of the implementation flow of the joint risk value determination operation of the data processing method provided in the embodiments of this application; Figure 6 A schematic diagram of the composition structure of the data processing apparatus provided in an embodiment of this application is shown; Figure 7 A schematic diagram of the composition structure of an electronic device according to an embodiment of this application is shown. Detailed Implementation

[0022] To make the objectives, features, and advantages of this application more apparent and understandable, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] First, the application scenario of this application will be explained. Current mainstream internet content filtering methods have fundamental flaws: the unified global control strategy adopted by the platform side is irreconcilably contradictory to the highly personalized psychological safety needs of users. Firstly, users cannot configure fine-grained settings and large-scale content blocking based on their own psychological sensitivities (such as specific phobias, traumatic trigger words, or insulting language in the context of cyberbullying). Whether it's parental controls in operating systems or the "content blurring" function of social media platforms, the strategies are uniformly formulated by the platform. Secondly, existing platform filtering mainly relies on keyword matching, lacking the ability to understand the semantics of text implying malice, and also unable to effectively identify and block sensitive content in images and videos, resulting in crude identification capabilities. Furthermore, theoretically, uploading user data to the cloud for powerful model analysis could achieve optimal filtering, but this would lead to the risk of privacy leaks of sensitive user content and behavioral data, creating a dilemma of "exposure is necessary for protection." In high-stress industries such as healthcare, education, and social media, many practitioners suffer severe psychological trauma due to persistent malicious comments, image attacks, or secondary dissemination. Existing "content compliance" reviews that only target the general public are unable to identify and intercept covert attacks that target the psychological sensitivities of specific individuals.

[0024] The aforementioned shortcomings are specifically reflected in the two current mainstream internet content filtering schemes: 1) Keyword filtering and youth mode. The technical principle is based on string matching using a pre-set sensitive word database, or restricting access to specific websites or application categories at the system level. The disadvantages of this approach are: lack of semantic understanding, making it unable to identify malicious comments without keywords (such as sarcastic or mocking remarks); lack of image recognition capabilities, rendering it completely ineffective against images and videos that cause psychological discomfort; and rigid rules that cannot adapt to individual psychological differences.

[0025] 2) Platform-side content blurring and warnings. The technical principle is to use a cloud-based AI model to uniformly identify images and videos published on the platform, blurring or warning about content that may cause discomfort. The disadvantages of this solution are: centralized strategy, users cannot customize sensitive content types (for example, the platform will not specifically filter insect images for users who are "afraid of insects"); limited coverage, limited to a single platform and unable to cover the entire operating system (such as local files, other applications, malicious images in bullet comments, and comment sections); and all content must be uploaded to the platform server for analysis, posing privacy risks.

[0026] Therefore, to solve the above-mentioned technical problems, this application provides a data processing method, apparatus, device, and storage medium. This data processing method can be executed by a data processing system (e.g., corresponding functional software or plugins) deployed locally on an electronic device. The electronic device can be a terminal device with conventional data processing and screen capture capabilities, such as a laptop, tablet, smartphone, smart monitor, or desktop computer.

[0027] Figure 1 A schematic diagram illustrating the implementation flow of the data processing method provided in an embodiment of this application is shown.

[0028] refer to Figure 1 This application provides a data processing method, which includes: Operation 101: Receive sensitive data for at least one modality input, project the sensitive data of each modality onto a unified semantic space, and obtain the sensitive features corresponding to each modality.

[0029] The data processing method of this application embodiment can be executed by a data processing system deployed in an electronic device, such as a psychological sensitivity protection system or a content security plugin. This data processing system provides a user-defined sensitivity setting interface. When a user first starts the system or actively triggers configuration, the system responds to this event by prompting the user to input their personal sensitive data. The system supports multiple modalities, including but not limited to text, visual, and audio.

[0030] For example, in text modality, the system can receive sensitive data such as keywords, phrases, or complete sentences that trigger psychological responses in user input; in visual modality, the system can receive sensitive data such as static images or videos that may cause discomfort and uploaded or captured by the user; and in audio modality, the system can receive sensitive data such as specific sound clips or music recorded or selected by the user.

[0031] After receiving sensitive data input by the user, in order to perform subsequent matching of data from different modalities within the same framework, the sensitive data of each modality needs to be projected onto a unified semantic space. The unified semantic space is a vector space with a preset fixed dimension (e.g., 128 dimensions). The projection operation transforms the original data of each modality into vectors in this space; these vectors are called the sensitive features corresponding to that modality. The specific implementation of the projection can be any mathematical transformation capable of transforming original data of different dimensions and properties into the same dimensional space. Each modality can correspond to one or more sensitive features, and all sensitive features are stored together on the local device, forming a user-specific sensitive feature library.

[0032] For example, the data processing system can support receiving user input in both text and visual modalities. Text modal input may include, but is not limited to, text descriptions or keywords (such as "insects," "dense patterns," "insulting comments") and emotion / semantic tags (such as "fear," "trauma," "feeling of being attacked"). Visual modal input may include triggering image examples uploaded or annotated by the user. After processing, the system outputs a user sensitivity library: First, text descriptions and semantic tags are input into the lightweight language model TinyBERT (768 dimensions), outputting an initial set of text-sensitive features; images are input into the lightweight visual network MobileNetV3 (960 dimensions), outputting an initial set of visual-sensitive features; then, a cross-modal semantic alignment mechanism is used to map the two types of features to a unified semantic space (e.g., dimension d=128), obtaining text-sensitive features for the text modality and visual-sensitive features for the visual modality.

[0033] Operation 102: Identify the display data of each modality on the screen, project the display data of each modality onto a unified semantic space, and obtain the display features of each modality.

[0034] The screen capture interface of the operating system can be used to acquire the current display content of an electronic device's screen in real time or periodically. Then, display data belonging to different modalities, such as text, static images, or dynamic videos on the screen, can be identified from the displayed content. After identifying the display data for each modality, the display data for each modality is projected into a unified semantic space to obtain the display features corresponding to each modality. These display features are in the same vector space as the previously stored sensitive features and have the same dimensions, thus laying the foundation for subsequent matching operations.

[0035] Operation 103: Match the display features of each modality with all the sensitive features corresponding to each modality to obtain the matching results.

[0036] Within a unified semantic space, the display features of each modality are matched against one or more pre-stored sensitive features for that modality. The matching operation is based on similarity calculations between features, such as dot product or cosine similarity. For each modality, if multiple sensitive features exist, the highest similarity is taken as the matching result for that modality. The matching result reflects the semantic proximity between the currently displayed content and the user's preset psychological sensitivities.

[0037] Operation 104: Determine the risk value of each modality based on the matching results.

[0038] Based on the matching results, a corresponding risk value can be determined for each modality. The risk value is positively correlated with the matching result; the higher the degree of matching, the more likely the displayed content is to cause psychological stimulation to the user, and the higher the risk value. Different modalities can independently define the conversion relationship from the matching result to the risk value, for example, by directly using the matching result as the risk value, or by mapping it through a preset non-linear function. The risk value can be cosine similarity, also known as risk confidence.

[0039] In one embodiment of this application, when a mode has multiple risk values, the risk value with the largest value among the multiple risk values ​​can be taken as the final risk value of the mode.

[0040] Operation 105: In response to the risk value of any modality being greater than the corresponding local threshold, perform visual masking processing on the display data of the corresponding modality.

[0041] When the risk value of any modality exceeds the corresponding local threshold, visual masking is applied only to the display data of that modality (e.g., blurring or folding text areas, blurring or obscuring image areas), while the display data of other modalities remains displayed normally. This approach is used to address content where a single modality clearly poses a psychological risk, such as purely insulting text or images of isolated violence. The local threshold can be set based on actual needs or experience.

[0042] Operation 106: In response to the existence of at least two modalities and the risk values ​​of all modalities being less than the corresponding local thresholds, calculate the semantic consistency scores of the display features of at least two modalities in a unified semantic space, determine the cooperative gating coefficients based on the semantic consistency scores, determine the joint risk score based on the cooperative gating coefficients and the risk values ​​of each modality, and in response to the joint risk score being greater than the global threshold, perform visual masking processing on the display data of all modalities.

[0043] When there are at least two modalities and the risk values ​​of all modalities are less than their respective local thresholds, the system does not allow passage directly. Instead, it further calculates the semantic consistency score of the display features of at least two modalities in a unified semantic space. For example, it calculates the similarity of at least two modalities as the semantic consistency score.

[0044] For example, considering two modalities (a text modality and a visual modality), the semantic consistency score c between the two can be calculated using the following formula:

[0045] in, For text features, As a visual feature, when c→+1, it indicates that the text and visual semantics are highly consistent, with both modalities pointing to the same or similar attack themes, emotional orientations, or psychological suggestions; when c≈0, it indicates that the text and visuals are basically unrelated semantically; when c<0, it indicates that the text and visuals are semantically conflicting or interfering with each other, such as insulting text paired with completely unrelated images, or violent images paired with neutral descriptive text. It should be noted that c is used to characterize whether a semantic synergy is formed between cross-modalities. It does not directly represent the magnitude of risk itself, but is used to determine whether the combination of text and images may produce an additional risk-enhancing effect.

[0046] Furthermore, to prevent semantic conflict scenarios from falsely boosting joint judgments, this application embodiment also introduces a collaborative gating coefficient, which is determined based on the semantic consistency score. A higher score results in a larger coefficient, indicating stronger semantic synergy among multimodal content and a more significant risk superposition effect; a lower score results in a smaller coefficient, effectively suppressing the amplification of misjudgments caused by semantic conflicts or irrelevant combinations. Finally, a joint risk score is determined based on the collaborative gating coefficient and the risk value of each modality. If the joint risk score is greater than a preset global threshold, overall visual masking processing is performed on the display data of all modalities (e.g., intercepting, blurring, or covering the entire screen area or combined content). This can be used to identify scenarios where a single modality has low risk, but the combination of text and images creates semantic implications, implicit attacks, or psychological triggers (e.g., neutral text paired with a neutral image, but the combination produces a clear ironic or insulting effect). The global threshold can be set according to actual needs or experience.

[0047] In one embodiment of this application, visual masking processing includes, but is not limited to, occlusion, blurring, blurring, folding, replacement, or interception.

[0048] To illustrate the visual masking process in this application embodiment more specifically, this application provides a two-layer, three-channel parallel decision architecture, taking a modality including a first modality and a second modality as an example. Under this architecture, the visual masking process is as follows: The first layer, independent single-modal decision-making: Two parallel channels are set up. Channel one is for the first mode: when the risk value of the first mode... Reaching or exceeding a local threshold When the default value is 0.6, local protection measures such as masking and collapsing are applied to the display data corresponding to the first modality. Channel 2 targets the second modality: when the risk value of the second modality... Reaching or exceeding a local threshold When the default value is 0.6, local protection such as blurring and occlusion is applied to the display data corresponding to the second modality. The protection actions of the first-layer channels are independent of each other and can take effect simultaneously.

[0049] The second layer involves cross-modal joint determination: a third channel is set up, which is activated only when the risk values ​​of both modes have not reached their respective local thresholds. The joint risk value is determined based on the risk values ​​of the first and second modes. Reaching or exceeding the global threshold When the value is 0.55 (e.g., the default value), comprehensive visual masking processing is performed on the entire content of the current screen, such as interception, blurring, or folding.

[0050] Thus, this embodiment of the application projects sensitive data input by the user and screen display data into the same unified semantic space for matching. Based on the matching results, it determines the risk value of each modality and performs targeted protection for single high-risk modalities. At the same time, it further calculates semantic consistency and introduces a collaborative gating coefficient for multimodal low-risk combinations, comprehensively judges joint risks, and performs overall protection. This achieves personalized, semantic-level sensitive content identification and protection for different modalities while protecting user privacy on the local device. It balances accurate interception of single modalities with identification of hidden risks in multimodalities, significantly improving the precision of identification, the targeted nature of protection, and the adaptability to complex scenarios. It effectively solves the technical defects of traditional global control strategies that cannot adapt to personalized psychological security needs and that inaccurate identification is caused by insufficient cross-modal semantic understanding.

[0051] In one embodiment of this application, in the above operation 101, the sensitive data of each modality are projected onto a unified semantic space, including: determining the feature extraction model and the mapping model corresponding to each modality; extracting features from the sensitive data of each modality using the feature extraction model corresponding to each modality to obtain the initial sensitive features of each modality; and mapping the initial sensitive features of each modality to the unified semantic space using the mapping model corresponding to each modality to obtain the sensitive features corresponding to each modality.

[0052] Different modalities of data have different original forms. For example, text needs to be processed by a language model, and images need to be processed by a vision model. Therefore, it is necessary to pre-configure or train a dedicated feature extraction model for each modality. At the same time, in order to transform the features of different modalities to the same vector space, it is also necessary to configure an independent mapping model for each modality.

[0053] After determining the different feature extraction models corresponding to different modalities, features are extracted from the sensitive data of each modality using the feature extraction model corresponding to each modality to obtain the initial sensitive features for each modality. For example, for user-inputted text sensitive data, it is fed into a text feature extraction model (such as a lightweight language model, TinyBERT), and a high-dimensional text semantic vector is output as the initial sensitive feature; for user-inputted image sensitive data, it is fed into an image feature extraction model (such as a lightweight vision model, MobileNetV3), and a high-dimensional image feature vector is output as the initial sensitive feature.

[0054] The initial sensitive features retain the original semantic information of each modality, but their dimensions differ, making direct comparison impossible. Therefore, a mapping model corresponding to each modality is needed to map the initial sensitive features of each modality to a unified semantic space, thus obtaining the corresponding sensitive features for each modality. For example, inputting text semantic vectors into a text mapping model (e.g., a fully connected layer) transforms them into vectors of fixed low dimensions (e.g., 128 dimensions), yielding sensitive features. Similarly, inputting image feature vectors into an image mapping model also transforms them into vectors of the same dimension, obtaining sensitive features. In this way, the sensitive features of different modalities reside in the same vector space with unified dimensions, allowing for direct similarity calculation or distance measurement, thereby laying the foundation for subsequent matching operations.

[0055] Figure 2 This illustration shows a schematic diagram of the implementation flow of the mapping model training operation of the data processing method provided in the embodiments of this application.

[0056] refer to Figure 2 In one embodiment of this application, the mapping model is obtained through the following training: Operation 201 uses the mapping layer weights of the corresponding modes in the target cross-modal model to initialize the training mapping model for each mode; Operation 202: Obtain multiple sensitive sample pairs. The sensitive sample pairs include anchor sensitive data, positive sample sensitive data that matches the anchor sensitive data, and negative sample sensitive data that does not match the anchor sensitive data. Operation 203 involves inputting the sensitive sample pairs into the initialized mapping models of each modality to be trained, so as to map them to a unified semantic space. Operation 204 aims to minimize the distance between anchor-sensitive data and positive sample-sensitive data in the unified semantic space and maximize the distance between anchor-sensitive data and negative sample-sensitive data in the unified semantic space. It updates the parameters of the mapping model to be trained for each modality and obtains the trained mapping model.

[0057] In this application, independent mapping models are trained for different modalities. Specifically, each mapping model is a linear mapping head, employing a single-layer fully connected (FC) structure. To achieve multimodal feature alignment and fusion, enabling semantic matching calculations for features from different modalities, the mapping models corresponding to different modalities are collaboratively trained within the same unified semantic space. The training process for each modality's mapping model is as follows: 1) First, in order to reduce the difficulty of model training convergence by relying on the general semantic prior knowledge of mature cross-modal models and improve the initial mapping effect, this embodiment uses the mapping layer weights of the corresponding modality inside the target cross-modal model to complete the parameter initialization of the mapping model to be trained in each modality; for example, the pre-trained weights disclosed by CLIP (Contrastive Language-Image Pre-training) can be used as the weight initialization basis to complete the initial configuration of the mapping model to be trained in each modality.

[0058] 2) Perform few-shot fine-tuning on the initialized mapping model to be trained using K sets of anchor-positive sample pairs (K≥1) (i.e., K sets of sensitive sample pairs) provided by the user. The fine-tuning process is as follows: Obtain multiple sets of sensitive sample pairs provided by the user, each set including anchor, positive sample, and negative sample; input these sample pairs into the corresponding mapping models and map them to a unified semantic space; then update the parameters of the mapping model with the goal of minimizing the distance between the anchor sensitive data and the positive sample sensitive data in the unified semantic space and maximizing the distance between the anchor sensitive data and the negative sample sensitive data in the unified semantic space, to obtain the final mapping model.

[0059] In one embodiment of this application, in order to ensure a balance between computational efficiency and semantic distinguishability, the dimension of the unified semantic space can be configured to a preset value.

[0060] Taking text modality (original feature dimension 768) and visual modality (original feature dimension 960) as examples, if independent linear mapping heads are trained for each modality, projecting their respective original features onto a unified semantic space of dimension d, the total number of parameters for the mapping heads is approximately 768×d + 960×d. When d is 128, the total number of parameters is approximately 220,000, resulting in minimal computational cost, allowing for real-time inference on terminal devices. Furthermore, 128 dimensions are sufficient to distinguish different semantics in cosine similarity calculation. In addition, the computational complexity of subsequent matching is proportional to the dimension d of the unified semantic space and the number of user-sensitive features |S|, i.e., O(d·|S|). Since the number of user-sensitive features typically does not exceed 50, when d=128, the time consumed in a single matching operation can be controlled within 1 millisecond. Therefore, this application preferably sets the dimension of the unified semantic space to 128 dimensions.

[0061] It should be noted that the embodiments of this application do not specifically limit the dimensions of the unified semantic space, and can be flexibly adjusted according to the actual deployment scenario, accuracy requirements, and device performance. The selection of 128 dimensions is only a preferred embodiment of this application.

[0062] In this application, the mapping model output dimension of each modality is consistent with the unified semantic space. The objective is to minimize the distance between anchor-sensitive data and positive sample-sensitive data in the unified semantic space, and to maximize the distance between anchor-sensitive data and negative sample-sensitive data. This can be viewed as minimizing the contrastive loss function. The embodiments of this application may use the standard triplet loss function:

[0063] A (anchor point) is the mapping vector of anchor point sensitive data in the unified semantic space, P is the mapping vector of positive sample sensitive data (matching the anchor point) in the unified semantic space, N is the mapping vector of negative sample sensitive data (not matching the anchor point) in the unified semantic space, D() is the Euclidean distance, and m is the safety interval parameter with an initial value of 0.3, which can be adaptively updated based on user feedback.

[0064] In this embodiment, anchor point A and positive sample P and negative sample N originate from different modalities, while positive sample P and negative sample N originate from the same modality. For example, in a typical but not limited application, anchor point A adopts a text modality (such as descriptive text input by the user), while positive sample P and negative sample N adopt a visual modality (such as images that semantically match or do not match the anchor point, respectively). Accordingly, the text modality and the visual modality each have independent mapping models: the text mapping model is used to map anchor point A to a unified semantic space, and the image mapping model is used to map positive sample P and negative sample N to the same unified semantic space. During training, the parameters of the text mapping model and the image mapping model are updated simultaneously through a triplet loss function, thereby reducing the distance between anchor point A and positive sample P and increasing the distance between anchor point A and negative sample N. In this way, the two mapping models from different modalities achieve collaborative training and semantic alignment in a unified semantic space.

[0065] In one embodiment of this application, the training of the mapping model can be performed once when the user first completes the configuration, and then incrementally updated as user feedback samples accumulate. The entire training is completed locally, ensuring that user data is not transmitted externally.

[0066] Verification shows that, through the above training process, the cosine similarity between insect text descriptions and insect images in a unified semantic space is significantly improved and can stably reach the target threshold (greater than 0.7). Therefore, this embodiment of the application, using the weights of a pre-trained cross-modal model as the initialization basis, and utilizing a small number of sensitive sample pairs provided by the user, collaboratively trains mapping models for different modalities such as text and images. This optimizes the mapping model to map semantically related cross-modal content to features in similar positions within a unified semantic space, providing a reliable foundation for subsequent real-time semantic matching and visual masking of sensitive content.

[0067] Figure 3 This illustration shows a schematic diagram of the implementation flow of the display data recognition operation of the data processing method provided in the embodiments of this application.

[0068] refer to Figure 3 In one embodiment of this application, at least one modality includes a text modality and / or a visual modality; the above-described operation 102, identifying the display data of each modality on the screen, includes: Operation 301: Obtain the screen display content of multiple consecutive frames, and perform dirty rectangle detection on the screen display content of multiple consecutive frames to obtain the dirty rectangle region.

[0069] This application embodiment periodically or event-drivenly acquires the current screen display content through the operating system's desktop compositing interface (e.g., Windows' DXGI Desktop Duplication API or macOS's ScreenCaptureKit). To reduce resource consumption, the system only analyzes areas that actually change, specifically through dirty rectangle detection. The detection steps are as follows: A screen image is captured at a preset fixed period (e.g., 200 milliseconds), and pixel-by-pixel differencing is performed with the previous frame to generate a difference image. The difference image is binarized, and a threshold for pixel grayscale change is set (e.g., 15), marking all pixels whose changes exceed the threshold. Connectivity analysis is performed on the marked changed pixel regions, extracting the bounding rectangle of each connected region as candidate dirty rectangle regions. The area of ​​each dirty rectangle region is calculated, and only regions with an area exceeding a certain proportion (e.g., 2%) of the total screen area are retained, filtering out minor changes such as cursor blinking and icon highlighting. The final retained regions are the dirty rectangle regions analyzed in this study, and the system only processes these regions subsequently.

[0070] Operation 302: In response to the detection of text content at the screen position corresponding to the dirty rectangular area, the text content is extracted and used as display data for the text modality.

[0071] After acquiring the dirty rectangular region, the content type of the screen area it covers is determined. If text content exists at that location (e.g., dialog box, comment section, or bullet screen area), text extraction is initiated. Extraction methods include, but are not limited to: performing local optical character recognition (OCR) on the screen area; or directly obtaining the text stream of that area through the operating system's accessibility API, such as comment content, title, bullet screen, or body text. The extracted raw text data serves as the display data for the text modality, used for subsequent feature extraction and projection.

[0072] Operation 303: In response to detecting that there is static image content at the screen position corresponding to the dirty rectangular area, the static image content is captured as the display data of the visual modality; or, in response to detecting that the screen position corresponding to the dirty rectangular area is a video playback area, a video frame sequence is captured from the video playback area at the target acquisition frequency as the display data of the visual modality.

[0073] The target acquisition frequency is determined based on the risk value of the text modality and the degree of change in the video playback area.

[0074] For visual modalities, a further distinction is made between static images and dynamic videos. If the screen position corresponding to the dirty rectangular area is a static image (such as a picture on a webpage or a preview image in a local photo album), then the image data of that area is directly extracted as the display data for the visual modal.

[0075] If the screen location corresponding to the dirty rectangle area is a video playback area (e.g., an online video window, a local player), then a continuous sequence of video frames needs to be captured from that area. An adaptive dynamic frame extraction strategy is used during acquisition, specifically including: Normal mode, which uses a lower base frequency as the target acquisition frequency (e.g., 1 frame per second) to extract keyframes by default. Linked trigger mode, when the risk value of bullet comments or subtitles in the currently displayed content is detected to be increased through text modality, the acquisition frequency is immediately increased to a higher frequency as the target acquisition frequency (e.g., 5 frames per second), entering a high-frequency monitoring state; if the risk value remains below a preset threshold for more than a set duration (e.g., 3 seconds), it automatically reverts to the base frequency. Sudden change trigger mode, based on the dirty rectangle detection results, if the video playback area undergoes a drastic change (e.g., the area of ​​change in the difference map exceeds 20% of the playback area, commonly seen in transitions and rapid flashing), the current sudden change frame is immediately forcibly captured.

[0076] The acquired video frames are organized by time windows (e.g., one second per window), and all frames within each window are used as visual modal display data for that time period. In subsequent processing, one or more frames can be selected from the window as representatives of that window according to preset rules, such as selecting the first frame, middle frame, or last frame within the window, but this application does not limit this.

[0077] Figure 4 The diagram illustrates the implementation flow of the feature extraction operation of the data processing method provided in this application embodiment.

[0078] refer to Figure 4 In one embodiment of this application, in the above operation 102, the display data of each modality is projected onto a unified semantic space to obtain the display features of each modality, including: Operation 401, in response to the display data being text content, extracts initial text features from the text content through a text feature encoder and projects them into a unified semantic space to obtain text features as display features of the text modality.

[0079] When the identified display data is text content (such as comments, bullet comments, titles, etc. extracted from the screen), the text content is first fed into a text feature encoder. The text feature encoder can, for example, employ the lightweight language model TinyBERT (trained with INT8 quantization), which can extract semantic features from the text, recognizing explicit keywords and subtle offensive expressions such as sarcasm and irony. TinyBERT outputs an initial text feature vector with 768 dimensions. Subsequently, the system projects this initial text feature into a unified semantic space (e.g., 128 dimensions) using a linear mapping head; the resulting vector is the display feature of the text modality.

[0080] Operation 402: In response to the display data being static image content, initial image features are extracted from the static image content using an image feature encoder and projected onto a unified semantic space to obtain image features as display features of the visual modality; or, in response to the display data being a video frame sequence, initial video features are extracted from each video frame in the video frame sequence using an image feature encoder and projected onto a unified semantic space to obtain video features corresponding to each video frame, and one video feature is selected from all video features as the display feature of the visual modality.

[0081] In one embodiment of this application, selecting one video feature from all video features as the display feature of the visual modality includes: matching the video features corresponding to each video frame in the video frame sequence with all sensitive features corresponding to the visual modality to obtain the risk value of each video frame; and selecting the video feature corresponding to the video frame with the highest risk value as the display feature of the visual modality.

[0082] If the displayed data is a static image (such as a picture or screenshot on a webpage), the image data is fed into an image feature encoder. For example, the image feature encoder can use the lightweight visual network MobileNetV3-Small, which outputs initial image features after global average pooling, with a dimension of 960. Then, the initial image features are projected onto the same unified semantic space (with the same dimension as the text features, e.g., 128 dimensions) through a corresponding linear mapping head to obtain the display features of the image modality.

[0083] If the displayed data is a sequence of continuous video frames (organized by time window) collected from the video playback area, perform the same operation as for static images on each frame: input each frame into an image feature encoder (such as MobileNetV3-Small), extract the initial frame-level features (960 dimensions), and then project them onto a unified semantic space (128 dimensions) through the same linear mapping head to obtain the video features corresponding to each frame.

[0084] Since a time window typically contains multiple frames, and the subsequent matching stage requires outputting the display features of a visual modality on a per-window basis, it is necessary to select a representative feature from the video features of all frames. In one embodiment of this application, the selection rule is as follows: the video features corresponding to each video frame in the video frame sequence are matched with the sensitive features of the visual modality in the user's sensitive database (for example, calculating cosine similarity and taking the maximum similarity as the risk value of that frame); then, the video features corresponding to the video frame with the highest risk value are selected as the visual modality display features of that time window. In this way, the image with the strongest psychological stimulation to the user within the window can be captured, improving the sensitivity of the warning.

[0085] Figure 5 This illustration shows a schematic diagram of the implementation process of the joint risk value determination operation of the data processing method provided in the embodiments of this application.

[0086] refer to Figure 5 In one embodiment of this application, the collaborative gating coefficient is determined based on the semantic consistency score, and the joint risk score is determined based on the collaborative gating coefficient and the risk values ​​of each modality, including: Operation 501 compares the semantic consistency score with zero and takes the maximum value as the collaborative gating coefficient.

[0087] To prevent semantic conflict scenarios from causing erroneous boosting of joint decision-making, embodiments of this application also incorporate a cooperative gating coefficient. Defined as Specifically: if the semantic consistency score ,but This indicates that there is no positive semantic collaboration among the current multiple modalities, and the joint channel does not enhance the overall risk; if ,but Furthermore, the gating coefficient increases with semantic consistency, indicating a stronger degree of validity in multimodal semantic combinations. This gating coefficient controls whether the second-layer joint decision takes effect and the magnitude of the enhancement, thereby ensuring that "irrelevant text / images" or "semantically conflicting text / images" are not incorrectly identified as combined attack scenarios.

[0088] Operation 502 determines the attention weight of each modality based on the risk value of each modality.

[0089] To characterize the relative contribution of each modality to the joint risk, the risk values ​​of each modality are transformed according to the Softmax function to obtain the attention weights of each modality.

[0090] Taking a modality comprising a first mode and a second mode as an example, the risk values ​​of the two modes can be transformed using the following softmax function:

[0091] in, This is the temperature adjustment coefficient (default value is 1), used to control the sensitivity of weight allocation. The calculated... and These are the attention weights for the first and second modes, respectively, and their sum is 1. The first mode risk value, This is the risk value for the second modality. This calculation allows the joint decision to naturally favor the higher-risk modality: for example, when the first modality has a higher risk, Larger modalities have a higher proportion of the first modality in the joint risk score; conversely, smaller modalities have a higher proportion of the first modality.

[0092] Operation 503: Based on the attention weights of each modality, the risk values ​​of each modality are weighted and fused to obtain a weighted fusion term.

[0093] The weighted fusion term is obtained by summing the products of the attention weights of each modality and their risk values. For example, assuming the modalities include a first modality and a second modality, the weighted fusion term would be: This value represents a weighted fusion of the risk levels of each modality, assuming semantic consistency.

[0094] Operation 504 multiplies the weighted fusion term by the collaborative gating coefficient to obtain the joint reward.

[0095] Multiplying the weighted fusion term by the cooperative gating coefficient yields the joint reward. For example, taking a two-modal model, the joint reward is... .

[0096] Operation 505 determines the collaborative reward based on the minimum value among the collaborative gating coefficient, the preset reward coefficient, and the risk values ​​of each modality.

[0097] To effectively identify combinatorial attacks when the semantics are valid and the risks of both modalities are at a moderate level, a cooperative reward term is further introduced. The cooperative gating coefficient... Multiply by the preset reward coefficient (For example, 0.2 by default), then multiply by the minimum risk value among all modalities to obtain the collaborative reward value. For example, taking a two-modality model, the collaborative reward is... This reward can provide an additional boost to the joint risk value when the individual risks of each modality are not high, but the semantics are consistent when combined, thereby identifying exemplary scenarios of "neutral graph + neutral text but the combination implies an attack".

[0098] Operation 506: Determine the joint risk score based on joint rewards and synergistic rewards.

[0099] The joint reward and the synergistic reward are added together to obtain the final joint risk value. For example, taking a two-modal scenario, the joint risk value is... :

[0100] The joint risk value has the following characteristics: when there is a semantic conflict ( ,but ), The combined channel automatically fails; when only a single mode is high-risk, The risk is relatively small, with limited collaborative reward items, mainly relying on the first-layer single-modal channel to perform local processing; when the individual risks of the two modalities are not high, but the semantics are valid after combination and At higher levels, the combined effect of collaborative gating items and collaborative reward items can make... Exceeding the joint threshold (For example, 0.55), thus triggering overall protection. Compared with a preset global threshold, if the risk is greater than the threshold and none of the single-modal risks exceed their respective local thresholds, for example... Then, perform overall visual masking processing (such as interception, overall blurring, overall occlusion, or overall folding) on ​​all modal display data within the current screen area.

[0101] In one embodiment of this application, the threshold is adaptively updated in response to user feedback on visual blocking processing. Specifically: when a user marks displayed data that has undergone visual blocking processing as a "false alarm" (i.e., normal content was blocked), the sample is recorded as a negative sample, and the local or global threshold of the corresponding modality is increased to reduce the probability of false blocking of similar content in the future. When a user marks displayed data that has not undergone visual blocking processing as a "missed alarm" (i.e., harmful content that should be blocked was not identified), the sample is recorded as a positive sample, and the local or global threshold of the corresponding modality is decreased to improve the sensitivity of identifying similar content in the future.

[0102] The threshold updates mentioned above involve only adjustments to scalar parameters, requiring minimal computation and can be completed immediately through real-time user feedback, without the need for background retraining. Through this lightweight adaptive mechanism, the system can continuously optimize its judgment criteria based on user interaction, better aligning with the dynamic changes in individual psychological sensitivity.

[0103] In one embodiment of this application, the parameters of the mapping model can also be updated in response to any of the following triggering conditions: the sum of the number of display data marked as false positives by the user and the number of display data not marked as missed positives by the user reaches a preset number (e.g., 20 by default); the time since the last mapping model update exceeds a preset duration (e.g., 7 days by default); within a preset sliding window (e.g., the last 50 predictions), the proportion of display data marked as false positives by the user or the proportion of display data not marked as missed positives by the user exceeds a preset proportion (e.g., 15%). To avoid affecting the user's normal operation, the above parameter update operation is performed asynchronously in the background when the device is charging and the system is idle (e.g., CPU utilization is below 30%).

[0104] When updating the parameters of the mapping model, training samples can be constructed as follows: Anchor point: Uses a set of existing sensitive features from the user's sensitive feature library. These sensitive features serve as a reference baseline and remain fixed in this update, not participating in parameter optimization. Positive sample: Display features corresponding to display data labeled as "missed" by the user. Negative sample: Display features corresponding to display data labeled as "false positive" by the user.

[0105] To prevent "catastrophic forgetting" (i.e., the degradation of the model's original recognition ability due to training with new samples), this application also introduces an old sample replay mechanism. Specifically, during each training session, a certain proportion (e.g., 30%) of old samples are randomly sampled from the historical sample pool and combined with the newly added feedback samples to form the training set. In this way, the model can retain its memory of old sensitive features while adapting to new feedback. The process of updating the parameters of the mapping model using training samples is consistent with the aforementioned training process of the mapping model, and will not be described again here.

[0106] It is worth noting that the update process in this application embodiment only updates the parameters of the mapping model (i.e., the linear mapping head) corresponding to each modality, without updating the weights of the backbone network (e.g., the text feature encoder TinyBERT, the image feature encoder MobileNetV3), thereby keeping the computational load within an acceptable range.

[0107] In addition, to prevent overfitting, an upper limit is set on the number of training epochs for the mapping model (e.g., 10 epochs by default), and an early stopping strategy is implemented: training is terminated early when the validation loss no longer decreases for three consecutive epochs. After training, the updated mapping model parameters replace the original parameters and are used for subsequent display feature projection and matching.

[0108] In one embodiment of this application, the entire mapping model parameter update process is executed locally, and all training data (including anchor points, positive samples, negative samples, and historical sample pools) are not transmitted to the network, thus fully protecting user privacy. Through the above mechanism, a closed loop of "self-learning-self-correction-self-optimization" is achieved, which can continuously adapt to the dynamic changes in the user's psychological sensitivities as the user uses the device, thereby improving the accuracy and personalization of the recognition.

[0109] In one embodiment of this application, the application also supports a single-modal degradation mode. The system automatically switches from multimodal mode to single-modal degradation mode when any of the following conditions are met: visual modality is disabled and only text modality is enabled in response to available memory falling below a memory threshold; screen capture permission is denied; or, the corresponding modality is disabled when the number of consecutive failed data acquisitions for any modality exceeds a preset number; visual modality is disabled in response to the single-frame inference latency of the image feature encoder continuously exceeding a preset duration and continuously reaching a preset number of frames; or the corresponding modality is disabled in response to the detection of a shutdown command for any modality.

[0110] Specifically, the embodiments of this application also support the following operations: When the available memory of an electronic device is below a preset threshold (e.g., 512MB), the visual modality is disabled, and only the text modality is enabled; If screen capture permission is denied, or if the number of consecutive optical character recognition (OCR) failures exceeds a preset number (e.g., 3 times), the corresponding mode that failed will be disabled. When the single-frame inference latency of an image feature encoder (e.g., MobileNetV3) exceeds a preset time (e.g., 100 milliseconds) and reaches a preset number of frames (e.g., 5 frames) consecutively, the visual modality is temporarily disabled. Users can manually disable a specific mode through the system settings interface.

[0111] In the single-modal degradation mode, the risk assessment formula can be simplified to: , Indicates the currently available modality (e.g., text modality or visual modality). This represents the risk value for that modality. The original local and global thresholds for each modality remain unchanged, and protection strategies (including local and overall protection) are still triggered normally; only risk assessment is based on the available modalities. This mechanism ensures that even when device resources are limited or some functions are malfunctioning, the system can still provide basic early warning and protection capabilities against psychologically sensitive content.

[0112] In one embodiment of this application, all data involved in the entire data processing process, including all features, thresholds, samples, etc., are stored in a local encrypted sandbox (AES-256 encryption), isolated at the file system level, and not exposed to other applications. Model inference (TinyBERT, MobileNetV3) and training are completed entirely locally, without relying on any network interface or remote API. Screen capture content is only temporarily processed in memory on a frame-by-frame basis, and released immediately after analysis, without being written to disk. The text content extracted by OCR is destroyed immediately after feature encoding, retaining only the features. Users can clear all local feature data and feedback samples with a single click in the settings interface at any time, restoring the factory settings.

[0113] Figure 6 A schematic diagram of the composition structure of the data processing apparatus provided in the embodiments of this application is shown.

[0114] refer to Figure 6 This application also provides a data processing apparatus, which includes: The first projection module 601 is used to receive sensitive data for at least one modality input, project the sensitive data of each modality onto a unified semantic space, and obtain the sensitive features corresponding to each modality. The second projection module 602 is used to identify the display data of each modality on the screen, project the display data of each modality onto a unified semantic space, and obtain the display features of each modality. The matching module 603 is used to match the display features of each modality with all the sensitive features corresponding to each modality to obtain the matching results; Module 604 is used to determine the risk value of each mode based on the matching results; The first shielding module 605 is used to perform visual shielding processing on the display data of the corresponding modality in response to the risk value of any modality being greater than the corresponding local threshold. The second masking module 606 is used to calculate the semantic consistency score of the display features of at least two modalities in a unified semantic space in response to the existence of at least two modalities and the risk values ​​of all modalities being less than the corresponding local thresholds, determine the cooperative gating coefficient based on the semantic consistency score, determine the joint risk score based on the cooperative gating coefficient and the risk values ​​of each modality, and perform visual masking processing on the display data of all modalities in response to the joint risk score being greater than the global threshold.

[0115] It should be noted that the description of the apparatus in this application embodiment is similar to the description of the method embodiment above, and has similar beneficial effects as the method embodiment; therefore, it will not be repeated. For any technical details not covered in the data processing apparatus provided in this application embodiment, please refer to... Figures 1 to 5 The meaning is understood in accordance with the description of any of the accompanying drawings.

[0116] According to embodiments of this application, this application also provides an electronic device and a readable storage medium.

[0117] Figure 7 A schematic block diagram of an example electronic device 700 that can be used to implement embodiments of this application is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the application described and / or claimed herein.

[0118] like Figure 7 As shown, the electronic device 700 includes a computing unit 701, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 702 or a computer program loaded from a storage unit 708 into a random access memory (RAM) 703. The RAM 703 may also store various programs and data required for the operation of the electronic device 700. The computing unit 701, ROM 702, and RAM 703 are interconnected via a bus 704. An input / output (I / O) interface 705 is also connected to the bus 704.

[0119] Multiple components in electronic device 700 are connected to I / O interface 705, including: input unit 706, such as keyboard, mouse, etc.; output unit 707, such as various types of displays, speakers, etc.; storage unit 708, such as disk, optical disk, etc.; and communication unit 709, such as network card, modem, wireless transceiver, etc. Communication unit 709 allows electronic device 700 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0120] The computing unit 701 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 701 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 701 performs the various methods and processes described above, such as data processing methods. For example, in some embodiments, the data processing method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 708. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 700 via ROM 702 and / or communication unit 709. When the computer program is loaded into RAM 703 and executed by the computing unit 701, one or more steps of the data processing method described above may be performed. Alternatively, in other embodiments, the computing unit 701 may be configured to perform data processing methods by any other suitable means (e.g., by means of firmware).

[0121] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transferring data and instructions to the storage system, the at least one input device, and the at least one output device.

[0122] The program code used to implement the methods of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0123] In the context of this application, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0124] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0125] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0126] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0127] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this application can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this application can be achieved, and this is not limited herein.

[0128] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified.

[0129] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A data processing method, characterized in that, The method includes: Receive sensitive data for at least one modality input, project the sensitive data of each modality onto a unified semantic space, and obtain the sensitive features corresponding to each modality; The display data of each modality on the screen is identified, and the display data of each modality is projected onto the unified semantic space to obtain the display features of each modality. The display features of each modality are matched with all the sensitive features corresponding to each modality to obtain the matching results; Determine the risk value for each modality based on the matching results; In response to a risk value for any modality exceeding the corresponding local threshold, visual masking is performed on the display data for the corresponding modality. In response to the existence of at least two modalities and the risk values ​​of all modalities being less than the corresponding local thresholds, the semantic consistency scores of the display features of at least two modalities in the unified semantic space are calculated. The collaborative gating coefficients are determined based on the semantic consistency scores. The joint risk score is determined based on the collaborative gating coefficients and the risk values ​​of each modality. In response to the joint risk score being greater than the global threshold, visual masking processing is performed on the display data of all modalities.

2. The method according to claim 1, characterized in that, The step of projecting the sensitive data of each modality onto a unified semantic space includes: Determine the feature extraction model and the mapping model corresponding to each modality; By using the feature extraction models corresponding to each modality, the sensitive data of each modality are extracted to obtain the initial sensitive features of each modality; By using the mapping model corresponding to each modality, the initial sensitive features of each modality are mapped to the unified semantic space to obtain the sensitive features corresponding to each modality.

3. The method according to claim 2, characterized in that, The mapping model is trained through the following operations: The mapping layer weights corresponding to the modalities in the target cross-modal model are used to initialize the training mapping model for each modality. Multiple sensitive sample pairs are obtained, wherein the sensitive sample pairs include anchor sensitive data, positive sample sensitive data that matches the anchor sensitive data, and negative sample sensitive data that does not match the anchor sensitive data; The sensitive sample pairs are respectively input into the initialized mapping models of each modality to be trained, so as to map them to the unified semantic space; With the goal of minimizing the distance between anchor-sensitive data and positive sample-sensitive data in the unified semantic space, and maximizing the distance between anchor-sensitive data and negative sample-sensitive data in the unified semantic space, the parameters of the mapping model to be trained for each modality are updated to obtain the trained mapping model.

4. The method according to claim 1, characterized in that, The at least one modality includes a text modality and / or a visual modality; The display data for each modality on the identification screen includes: The screen display content of multiple consecutive frames is obtained, and dirty rectangle detection is performed on the screen display content of multiple consecutive frames to obtain the dirty rectangle region. In response to the detection of text content at the screen position corresponding to the dirty rectangular area, the text content is extracted and used as the display data of the text modality; In response to the detection of static image content at the screen location corresponding to the dirty rectangular area, the static image content is captured and used as display data for the visual modality; or, in response to the detection that the screen location corresponding to the dirty rectangular area is a video playback area, a video frame sequence is captured from the video playback area at a target acquisition frequency and used as display data for the visual modality; wherein, the target acquisition frequency is determined based on the risk value of the text modality and the degree of screen change in the video playback area.

5. The method according to claim 4, characterized in that, The step of projecting the display data of each modality onto the unified semantic space to obtain the display features of each modality includes: In response to the fact that the displayed data is text content, initial text features are extracted from the text content by a text feature encoder and projected onto the unified semantic space to obtain text features as the display features of the text modality; In response to the display data being static image content, initial image features are extracted from the static image content using an image feature encoder and projected onto the unified semantic space to obtain image features as display features of the visual modality; or, in response to the display data being a video frame sequence, initial video features are extracted from each video frame in the video frame sequence using an image feature encoder and projected onto the unified semantic space to obtain video features corresponding to each video frame, and one video feature is selected from all video features as the display feature of the visual modality.

6. The method according to claim 5, characterized in that, The step of selecting one video feature from all video features as the display feature of the visual modality includes: The video features corresponding to each video frame in the video frame sequence are matched with all sensitive features corresponding to the visual modality to obtain the risk value of each video frame. The video features corresponding to the video frame with the highest risk value are selected as the display features of the visual modality.

7. The method according to claim 1, characterized in that, The process of matching the display features of each modality with all the sensitive features corresponding to each modality to obtain the matching results includes: Calculate the similarity between the displayed features of each modality and all the sensitive features corresponding to each modality, and use the maximum similarity as the risk value of each modality.

8. The method according to claim 1, characterized in that, The step of determining the collaborative gating coefficient based on the semantic consistency score, and determining the joint risk score based on the collaborative gating coefficient and the risk values ​​of each modality, includes: The semantic consistency score is compared with zero, and the maximum value is taken as the collaborative gating coefficient; Based on the risk value of each modality, determine the attention weight of each modality; Based on the attention weights of each modality, the risk values ​​of each modality are weighted and fused to obtain a weighted fusion term; Multiplying the weighted fusion term by the collaborative gating coefficient yields the joint reward; The collaborative reward is determined based on the minimum value among the collaborative gating coefficient, the preset reward coefficient, and the risk values ​​of each modality. The joint risk score is determined based on the joint reward and the synergistic reward.

9. The method according to claim 1, characterized in that, The method further includes: In response to detecting that the user has marked the display data for which visual masking is performed as a false alarm, the corresponding local or global threshold is increased; In response to detecting that a user has marked display data for which visual masking has not been performed as a missed report, the corresponding local or global threshold is reduced.

10. The method according to claim 2, characterized in that, The method further includes: The parameters of the mapping model are updated in response to any of the following: The sum of the number of display data that the user marked as false alarms after performing visual masking and the number of display data that the user marked as missed alarms after not performing visual masking reaches a preset number; The time since the last mapping model update exceeds the preset time; Within a preset sliding window, the proportion of users marking display data that has undergone visual masking as false alarms or display data that has not undergone visual masking as missed alarms exceeds a preset proportion.

11. The method according to claim 1, characterized in that, The method further includes: In response to the detection that available memory is below the memory threshold, the visual modality is disabled, and only the text modality is enabled; In response to screen capture permission being denied; or, if the number of consecutive failures to acquire display data for any mode exceeds a preset number, the corresponding mode will be disabled; If the single-frame inference delay of the image feature encoder continues to exceed a preset duration and reaches a preset number of frames consecutively, the visual modality is disabled. In response to the detection of a shutdown command for any modality, the corresponding modality is disabled.

12. A data processing apparatus, characterized in that, The device includes: The first projection module is used to receive sensitive data for at least one modality input, project the sensitive data of each modality onto a unified semantic space, and obtain the sensitive features corresponding to each modality. The second projection module is used to identify the display data of each modality on the screen, and project the display data of each modality onto the unified semantic space to obtain the display features of each modality. The matching module is used to match the display features of each modality with all the sensitive features corresponding to each modality to obtain the matching results; The determination module is used to determine the risk value of each modality based on the matching results; The first shielding module is used to perform visual shielding processing on the display data of the corresponding modality in response to the risk value of any modality being greater than the corresponding local threshold. The second masking module is used to calculate the semantic consistency score of the display features of at least two modalities in the unified semantic space in response to the existence of at least two modalities and the risk value of all modalities being less than the corresponding local threshold, determine the cooperative gating coefficient based on the semantic consistency score, determine the joint risk score based on the cooperative gating coefficient and the risk value of each modality, and perform visual masking processing on the display data of all modalities in response to the joint risk score being greater than the global threshold.

13. An electronic device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-11.

14. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-11.

Images