A home video privacy degree evaluation method fusing VLM and micro-rough set optimization

Abstract

The present application relates to the technical field of computer vision and pattern recognition, in particular to a home video privacy evaluation method fusing VLM and differentiable rough set optimization, comprising: performing multi-layer compressed sensing processing on the input video; introducing a visual language model (VLM) privacy evaluation mechanism based on human cognitive alignment; generating privacy level labels conforming to expert evaluation scales through expert scoring samples and feedback type prompts optimization; constructing a space-time dual-flow network to extract spatial visual representation and time sequence dynamic representation in the privacy protection video respectively, realizing semantic alignment of heterogeneous dual-flow features through contrastive learning, promoting space and time sequence information fusion by using cross attention mechanism, and obtaining unified video privacy feature representation; constructing a privacy level prototype with ordered relationship constraints, introducing a differentiable fuzzy rough set theory to model the uncertainty between different privacy levels through fuzzy membership relationship and boundary area, and realizing continuous reasoning and quantitative scoring of the degree of video privacy protection.

Description

Technical Field

[0001] This invention relates to the field of computer vision and pattern recognition technology, specifically to a method for evaluating the privacy of home videos that integrates VLM and differentiable rough set optimization. Background Technology

[0002] With the widespread adoption of video surveillance, the technology has been integrated into all aspects of production and daily life. Especially in recent years, its cutting-edge applications in areas such as home-based elderly care have demonstrated a clear trend of expansion from public environments to private settings. While video surveillance systems deployed in home environments can provide functions such as elderly care, abnormal behavior detection, and security alerts, they also inevitably and continuously collect visual information related to personal habits, family structures, and private behaviors, thus posing a serious risk of privacy breaches.

[0003] To address this issue, existing research has proposed various privacy protection strategies. Among them, compressed sensing, as a typical video privacy protection method, maps high-dimensional video data to a low-dimensional observation space through a random projection mechanism. This causes nonlinear mixing and structural perturbation of the original visual information in the observation domain, thereby reducing the direct identifiability of sensitive content. Furthermore, by adjusting parameters such as the number of compression layers, the degree of visual information loss can be controlled, achieving different levels of video privacy protection.

[0004] However, existing research on visual privacy protection mainly focuses on the design and optimization of protection algorithms, lacking an effective quantitative evaluation mechanism for privacy protection effectiveness. Existing methods typically use metrics such as reconstruction error, image quality changes, or information retention rate to measure processing results, but these metrics primarily reflect the degree of visual information degradation and are insufficient to accurately describe the impact of privacy protection operations on sensitive semantic information and identifiability. In practical applications, different privacy protection strategies and parameter configurations can lead to varying degrees of changes in video spatial structure, semantic content, and temporal dynamic patterns, but currently, there is a lack of unified evaluation standards to objectively measure and compare the privacy protection capabilities of different protection schemes. Furthermore, video data exhibits significant spatiotemporal correlation characteristics; privacy protection operations not only alter the spatial visual information in a single frame but also affect the dynamic consistency and behavioral semantic expression between consecutive frames. Therefore, relying solely on traditional image quality metrics or spatial information loss measures is insufficient to comprehensively reflect the risk of video privacy leakage. Therefore, it is necessary to construct a video privacy protection evaluation method that integrates video spatial and temporal features and can achieve a unified quantification of privacy protection effectiveness. Summary of the Invention

[0005] The purpose of this invention is to address the lack of an effective quantitative evaluation mechanism in existing video privacy protection methods, which makes it difficult to accurately measure the spatial information and temporal dynamic changes of videos under different privacy protection levels. This invention proposes a home video privacy evaluation method that integrates Visual Language Modeling (VLM) and differentiable rough set optimization. This method utilizes a visual language model to guide privacy semantic cognition, learns video privacy representations through spatiotemporal dual-stream feature extraction and cross-modal fusion mechanisms, then constructs an ordered privacy level prototype space and combines it with differentiable fuzzy rough sets for boundary-aware level inference, achieving a continuous quantitative evaluation of the degree of video privacy protection.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] A method for evaluating home video privacy that integrates VLM and differentiable rough set optimization includes the following steps:

[0008] Step 1: Input the privacy-protected home video data and process the original home video at different intensities to generate video samples with different levels of privacy protection;

[0009] Step 2: Using expert subjective evaluation samples as a cognitive benchmark, the VLM is iteratively calibrated through privacy evaluation prompts and feedback optimization mechanisms to construct a VLM model aligned and calibrated with human cognition; then, video samples with different levels of privacy protection are fed into the VLM model aligned and calibrated with human cognition to generate corresponding privacy protection level labels.

[0010] Step 3: Input the privacy-protected video into the spatiotemporal dual-stream network to extract features;

[0011] Step 4: Perform cross-modal alignment on the extracted features and use an attention mechanism to perform spatiotemporal information interaction and fusion to obtain spatiotemporal fusion features;

[0012] Step 5: Represent the spatiotemporal fusion features of different privacy protection levels as learnable hierarchical prototypes with ordered structures, and constrain the arrangement relationship between adjacent hierarchical prototypes to construct a privacy level prototype space with hierarchical constraints, so that the privacy level representation conforms to the continuous change law from low to high.

[0013] Step 6: Based on the ordered privacy level prototype space, use the upper and lower approximate adaptive characterization of differentiable fuzzy rough sets to depict the transition boundary between different privacy protection levels.

[0014] Step 7: Based on the size of the boundary domain, perform boundary-aware privacy level inference for video privacy protection level, output the probability distribution of each privacy level, and finally map it into a continuous video privacy protection level score.

[0015] Preferably, step 2 specifically includes:

[0016] S201. Construct a set of expert subjective evaluations, selecting them from the training set. A cognitive calibration dataset is composed of representative videos, processed with different levels of privacy protection. ; then invited Each video sample was scored for its level of privacy protection by a team of experts with relevant experience. The final average opinion score is obtained by employing an extreme value removal averaging strategy. ;

[0017] in, Indicates the first The human subjective privacy score corresponding to each video is used as a real reference benchmark in the subsequent VLM cognitive calibration process to constrain the consistency between model output and human privacy perception.

[0018] S202. Using VLM to probabilistically model video privacy levels:

[0019] Let the set of privacy levels be These represent different levels of privacy protection for input videos. VLM based on the current prompt word Output the probability distribution corresponding to each privacy level. ,in Subsequently, the discrete 0-4 ratings were converted to a continuous 0-10 privacy score using an expectation mapping method.

[0020] ;

[0021] S203. Design a cue word calibration mechanism based on scoring error feedback:

[0022] For the In the evaluation process, the VLM prediction error is defined as: ,in This indicates the expert MOS score. Indicates the first VLM prediction scores were then used; subsequently, high-biased samples were selected based on the magnitude of the error. ,in The error threshold is used to convert high-bias samples and their corresponding expert evaluation differences into structured feedback information. ,in This represents the feedback generation function; finally, the next round of prompts is updated based on the generated feedback information. ,in This indicates a context fusion operation;

[0023] S204. Use correlation indicators to measure the consistency between model predictions and expert MOS to evaluate whether VLM has achieved cognitive alignment:

[0024] Spearman's rank correlation coefficient was used to evaluate the consistency of predicted rankings. The linear consistency of the predicted scores was evaluated using the Pearson correlation coefficient. Then when and And in multiple iterations At that time, it is considered that VLM has reached a stable state of human cognitive alignment;

[0025] S205. Input video samples with different levels of privacy protection into the calibrated VLM model to generate corresponding privacy protection level labels; wherein, the privacy protection level labels include five levels from 0 to 4, which correspond to five privacy protection levels: none, weak, medium, good, and excellent, respectively, and are used to characterize the overall privacy protection level of the video.

[0026] Preferably, step 3 specifically includes:

[0027] S301, The spatiotemporal dual-stream network includes:

[0028] Spatial branching is used to model the structural information and identifiability changes of a single frame image, and to characterize the impact of privacy protection operations on visual appearance.

[0029] The temporal branch is used to model the motion dependencies and behavioral evolution characteristics between consecutive frames, reflecting the degree to which privacy protection undermines the consistency of temporal semantics.

[0030] S302. Perform feature modeling on the input multi-layer compressed sensing processed video:

[0031] Parallel receive batch size is Collection of privacy-protected video clips and the corresponding set of intermediate video frames extracted from it. As input, a spatiotemporal encoder is used to capture feature vectors containing information on motion continuity and temporal distortion from video clips. The appearance encoder is used to extract appearance feature vectors containing spatial resolution and pixel occlusion from a single frame image. .

[0032] Preferably, step 4 specifically includes:

[0033] S401. Introduce a contrastive learning mechanism to analyze the spatiotemporal features from the same protected video segment. Appearance features Positive sample pairs are formed, and features from other videos within the same batch are formed negative sample pairs. By optimizing the InfoNCE loss function, two-dimensional features under the same privacy state are forced to be mapped to the same hypersphere. The symmetric alignment loss function is defined as follows:

[0034] ;

[0035] ;

[0036] in, This represents the set of all target modal feature vectors within the same batch. This indicates temperature hyperparameters. It represents the mathematical expectation of the distribution;

[0037] S402. A Transformer decoding layer with a cross-attention mechanism is introduced to perform deep interaction and fusion of features. In this module, the aligned 3D spatiotemporal features... Serving as query vectors, they guide the model to actively focus on and extract 2D appearance features as keys and values. The key spatial distortion information in the image; its linear mapping and attention calculation formulas are as follows:

[0038] ;

[0039] ;

[0040] ;

[0041] ;

[0042] in, express , express ; Representation layer normalization; This represents the learnable weight matrix. Indicates the scaling dimension of the key vector;

[0043] The complementary interaction features extracted by the cross-attention module are then linked with the original spatiotemporal features via residual connections. The mixture is then added together and fused before being fed into a multilayer perceptron for processing.

[0044] ;

[0045] ;

[0046] Ultimately, the output fusion features Dimensionally reduced by MLP Compact feature space of dimensionality .

[0047] Preferably, step 5 specifically includes:

[0048] S501, Extracting dual-stream fusion features Next, we first construct a learnable, ordered privacy level prototype space and introduce... Each privacy level prototype vector: Each prototype vector corresponds to a semantic center state with a privacy level, and the privacy levels increase progressively from low to high.

[0049] S502. Introduce chained ordering constraints to explicitly constrain the relative distance relationships between prototypes. The privacy level prototype space is defined as follows:

[0050] .

[0051] Preferably, step 6 specifically includes:

[0052] S601. After constructing the ordered prototype space, in order to characterize the sample features To establish a continuous association between samples and privacy levels, a fuzzy membership modeling method based on Gaussian kernel function is introduced, defining the relationship between samples and the first privacy level. The similarity between the prototypes at each level is: ;in, Indicates temperature parameter;

[0053] The similarity function extends the hard determination of a sample belonging to a certain category to a representation of multiple privacy levels with different degrees of membership, thus forming a fuzzy membership vector: ;

[0054] S602. Introducing the concept of lower approximation in fuzzy rough sets, then for the true privacy level... The following approximation is defined as: ;

[0055] This metric measures whether a sample is simultaneously far from all non-target levels.

[0056] Preferably, step 7 specifically includes:

[0057] S701. To explicitly characterize the fuzzy boundary structure between privacy levels, a boundary domain function is introduced:

[0058] ;

[0059] in, Used to control boundary sensitivity; this function maps the lower approximation to... Intervals, thus forming continuous boundary measures; at this time when When the value is close to 0, it indicates that the sample is in the deterministic region; when it is close to 1, it indicates that the sample is in the fuzzy transition region of the privacy level.

[0060] Based on the aforementioned boundary domain function, a boundary-aware fuzzy inference mechanism is constructed. During training, the overall optimization objective is defined as:

[0061] ;

[0062] in, For cross-entropy loss, Label smoothing loss;

[0063] Then, a boundary regularization term is introduced: , used to control the overall boundary area size;

[0064] S702. Map the privacy level probability distribution output by the model to a continuous privacy protection score. Specifically, the model first outputs probability distribution vectors for different privacy levels. Subsequently, based on the ordered semantic relationships of privacy levels, a linear weighted mapping strategy is introduced to map the discrete probability space to the continuous scoring space.

[0065] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0066] 1. A privacy assessment mechanism is constructed based on Video Rendering Model (VLM). This mechanism utilizes expert scoring priors and feedback mechanisms to cognitively calibrate the VLM, enabling the model to accurately learn and transfer the subjective evaluation patterns of human experts regarding video privacy protection, achieving semantic alignment of privacy assessments. This solves the problem of high costs associated with traditional manual subjective evaluations and provides an automated privacy assessment standard for subsequent networks that conforms to human perceptual priors.

[0067] 2. A spatiotemporal dual-stream architecture is constructed to model the spatial appearance degradation and temporal dynamic change features of privacy-preserving videos respectively. Feature alignment and fusion are achieved through cross-modal contrastive learning and cross-attention mechanisms, thereby improving the integrity of privacy representation.

[0068] 3. A learnable ordered privacy level prototype space is introduced. By representing spatiotemporal fusion features of different degrees as learnable level prototypes and imposing strict geometric arrangement constraints, the loss of privacy level progression sequence is effectively prevented, laying a spatial topological foundation for subsequent accurate scoring.

[0069] 4. A boundary-aware reasoning mechanism based on differentiable fuzzy rough sets is proposed. On the basis of an ordered prototype space, the upper and lower approximations of rough sets and the boundary domain are used to accurately characterize the transition boundaries between different privacy levels. The loss weights are adaptively adjusted according to the uncertainty of the samples in the boundary domain, realizing a flexible transition from hard classification with black-and-white criteria to a smooth probability distribution, thereby outputting a more robust, stable and fine-grained continuous privacy protection score. Attached Figure Description

[0070] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0071] Figure 1 This is a schematic flowchart of the method of the present invention;

[0072] Figure 2 Comparison chart of the effects of different privacy protection levels in videos in this invention;

[0073] Figure 3 This is a convergence curve diagram of the VLM cognitive feedback mechanism in this invention;

[0074] Figure 4 This is a probability distribution diagram of the boundary domain prediction for each privacy protection level of the video sample set in this invention;

[0075] Figure 5 This is a schematic diagram of the continuous privacy protection score based on probability distribution mapping in this invention. Detailed Implementation

[0076] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0077] Please see Figures 1-5 The present invention provides the following technical solution:

[0078] Example 1: This example provides a method for evaluating home video privacy by integrating VLM and differentiable rough set optimization. It includes the following steps:

[0079] Step 1, input privacy-protected home video data:

[0080] In this embodiment, the input video clips are real-world private home scenes based on the Charades dataset, covering typical private environments such as bedrooms, kitchens, and living rooms. They contain rich information on human behavior and interactions between people and their environment, realistically reflecting the visual privacy leakage risks in applications such as smart monitoring and home services. Privacy perturbation processing is applied to the videos using multi-layer compressed sensing to generate video sets with varying levels of privacy protection.

[0081] Step 2: Using expert subjective evaluation samples as a cognitive benchmark, the VLM is iteratively calibrated through privacy evaluation prompts and feedback optimization mechanisms to construct a VLM model aligned with human cognition; then, video samples with different levels of privacy protection obtained in Step 1 are input into the calibrated VLM model to generate corresponding privacy protection level labels.

[0082] Since traditional manual privacy assessments are limited by high costs, poor subjective consistency, and difficulty in scaling, and because there is a significant semantic deviation between the original visual cognition of general visual language models and the judgment standards of human experts in specific privacy protection scenarios, this embodiment introduces a lightweight visual language model based on human cognition alignment, the Qwen-VL privacy measurement calibration mechanism, as a cognitive prior agent for automated privacy assessment.

[0083] First, an expert subjective evaluation dataset is constructed. One hundred representative videos are selected from the training set and processed at different privacy protection levels (CS0-CS4) to form a cognitive calibration dataset. For a given set containing... Calibration dataset of 500 video samples Among them, invitation Each video sample was scored for its level of privacy protection by a team of experts with relevant experience. .in, Indicates the first The expert commented on the first The rating of each video sample.

[0084] Considering the potential for significant bias from individual experts, this paper employs an extreme value removal and averaging strategy to obtain the final average opinion score. The specific formula is as follows: ;

[0085] in Indicates the first The human subjective privacy score corresponding to each video is used as a real reference benchmark in the subsequent VLM cognitive calibration process to constrain the consistency between model output and human privacy perception.

[0086] Because the level of privacy protection exhibits a clear and continuous variation, directly using discrete level labels is insufficient to accurately represent the subtle differences between different protection strengths. Therefore, a Virtual Model (VLM) is employed to probabilistically model video privacy levels. Let the set of privacy levels be... This indicates different levels of privacy protection. (Regarding input videos...) VLM based on the current prompt word Output the probability distribution corresponding to each privacy level. Sum of probabilities Unlike traditional methods that directly output the highest probability category, this paper preserves the complete probability distribution information of the VLM and transforms the discrete 0-4 levels into a continuous 0-10 privacy score through expectation mapping. The specific formula is as follows:

[0087] , ;

[0088] Although VLM possesses strong visual understanding capabilities, its initial evaluation criteria are derived from general visual knowledge, which may deviate from human evaluation standards in privacy-preserving scenarios. To address this, a cue word calibration mechanism based on scoring error feedback was designed to gradually align the VLM evaluation process with expert cognition.

[0089] For the In the evaluation process, the VLM prediction error is defined as... ,in This indicates the expert MOS score. Indicates the first VLM prediction scores were then calculated. High-biased samples were then selected based on the magnitude of the error. , The error threshold is set to 1.

[0090] The erroneous samples and their corresponding expert evaluation discrepancies were then converted into structured feedback information. ,in This represents the feedback generation function, which updates the prompt words for the next round based on the generated feedback information. ,in This indicates a context fusion operation.

[0091] To evaluate whether VLM has achieved cognitive alignment, a correlation index is used to measure the consistency between the model's predictions and expert MOS (Mean Opinion of Qualification). The Spearman rank correlation coefficient is used to evaluate the consistency of the predicted rankings. To represent. The Pearson correlation coefficient is used to evaluate the linear consistency of the predicted scores, using... Indicates. When and ( (Set to 0.95) and in multiple iterations At this point, VLM is considered to have reached a stable state of alignment with human cognition, meaning that a mapping relationship has been established between visual privacy protection and expert subjective evaluation, enabling automated classification of privacy levels for videos with different protection strengths based on the laws of human privacy perception. During this iterative calibration process, the convergence curve of the VLM cognitive feedback mechanism is as follows: Figure 3 As shown.

[0092] The video samples with different levels of privacy protection obtained in step 1 are then input into the calibrated VLM model to generate corresponding privacy protection level labels. These labels comprise five levels (0–4), corresponding to no, weak, medium, good, and excellent privacy protection levels, respectively, representing the overall privacy protection level of the video. This labeling result demonstrates the VLM's ability to transfer expert privacy perception scales and serves as a priori human cognition in the subsequent spatiotemporal dual-stream network training process, guiding the model to learn video privacy representations consistent with expert evaluations.

[0093] Step 3: Input the privacy-preserving video obtained in Step 2 into a spatiotemporal dual-stream network to extract features.

[0094] In video privacy protection evaluation tasks, privacy leakage is manifested not only in the decrease in spatial recognizability of single-frame images but also in the disruption of the continuity of temporal information across frames. Therefore, relying solely on a single spatial or temporal modeling approach is insufficient to comprehensively characterize the degradation process of privacy information. To address this, a spatiotemporal dual-stream feature extraction network is introduced. The spatial branch models the structural information and recognizability changes of single-frame images, characterizing the impact of privacy protection operations on visual appearance; the temporal branch models the motion dependencies and behavioral evolution features between consecutive frames, reflecting the degree to which privacy protection disrupts the temporal semantic consistency.

[0095] Feature modeling is performed on the input multi-layer compressed sensing video. Specifically, the parallel receiving batch size is... Collection of privacy-protected video clips and the corresponding set of intermediate video frames extracted from it. As input. To ensure the effectiveness and lightweight nature of feature extraction, this paper uses 3DResNet-18 as the spatiotemporal feature encoder ( ), 2D ResNet-18 as an appearance feature encoder ( By removing the fully connected layers at the end of the original network and applying global average pooling, the spatiotemporal encoder captures feature vectors containing information about motion continuity and temporal distortion from video clips. Simultaneously, the appearance encoder extracts appearance feature vectors from a single frame image, including spatial resolution and pixel occlusion conditions. .

[0096] Step 4: Perform cross-modal alignment on the spatial and temporal features obtained in Step 3, and use the cross-attention mechanism to achieve spatiotemporal information interaction and fusion.

[0097] Because privacy protection mechanisms often compromise the spatial and temporal dimensions of videos to varying degrees, and their impact on physical semantics differs, the extracted... and Features are typically distributed across different feature hyperspheres. Direct cross-domain fusion makes it difficult for the model to effectively capture joint feature distortions under different privacy levels. Therefore, a contrastive learning mechanism is introduced before feature fusion to maximize the performance of the same privacy-preserving video. and Mutual information between them. Specifically, this involves: sharing spatiotemporal features from the same protected video segment. Appearance features Positive sample pairs are formed, and features from other videos within the same batch are formed into negative sample pairs. By optimizing the InfoNCE loss function, two-dimensional features under the same privacy state are forced to be mapped to the same hypersphere. Its symmetric alignment loss function is defined as follows:

[0098] ;

[0099] ;

[0100] in, This represents the set of all target modal feature vectors within the same batch. For temperature hyperparameters, This represents the mathematical expectation of the distribution.

[0101] Traditional fusion strategies neglect the deep correlations between cross-modal features. To fully exploit the spatiotemporal correlation disruptions caused by privacy protection, a Transformer decoding layer incorporating a cross-attention mechanism is introduced to enable deep interaction and fusion of features. In this module, aligned 3D spatiotemporal features... Serving as query vectors, they guide the model to actively focus on and extract 2D appearance features as keys and values. The key spatial distortion information is as follows. Its linear mapping and attention calculation formulas are as follows:

[0102] ;

[0103] ;

[0104] ;

[0105] ;

[0106] in, represent , represent ; For layer normalization; The weight matrix is ​​a learnable matrix. The scaling dimension of the key vector.

[0107] The complementary interaction features extracted by the cross-attention module are then linked with the original spatiotemporal features via residual connections. The mixture is then added together and fused before being fed into a multilayer perceptron for processing.

[0108] ;

[0109] ;

[0110] Ultimately, the output fusion features Dimensionally reduced by MLP Compact feature space of dimensionality .

[0111] Step 5: Based on the spatiotemporal fusion features obtained in Step 4, construct an ordered privacy level prototype space.

[0112] Extracting dual-stream fusion features Next, we first construct a learnable, ordered privacy level prototype space and introduce... Each privacy level prototype vector: Each prototype vector corresponds to a semantic center state with a privacy level, which increases progressively from low to high. Unlike traditional discrete classification, this design transforms the privacy assessment problem into a continuously ordered prototype learning problem, enabling the model to characterize the gradual relationship between different privacy protection strengths in a unified feature space. To ensure the topological consistency and semantic monotonicity of privacy levels in the feature space, this paper further introduces a chain-order constraint to explicitly constrain the relative distance relationship between prototypes, formally defined as follows:

[0113] ;

[0114] This constraint ensures that adjacent privacy levels have a progressive distribution relationship in the feature space, thereby avoiding spatial overlap or disordered ordering of different level prototypes and giving the privacy level space a stable geometric structure.

[0115] Step 6: For the privacy level prototype space constructed in Step 5, use the upper and lower approximate adaptive characterization of differentiable fuzzy rough sets to describe the transition boundary between different privacy protection levels.

[0116] After constructing an ordered prototype space, in order to characterize sample features To establish a continuous association between samples and privacy levels, a fuzzy membership modeling method based on Gaussian kernel function is introduced, defining the relationship between samples and the first privacy level. The similarity between the prototypes of each level is: ;

[0117] in, This is a temperature parameter used to control the smoothness of the membership distribution.

[0118] Here, similarity is strictly mapped to Within the interval, it represents the degree to which the sample is close to the core of each evaluation level in the feature space. This similarity function extends the hard determination of a sample belonging to a certain category to a representation of multiple privacy levels with different degrees of membership, thus forming a fuzzy membership vector:

[0119] ;

[0120] This is used to represent the transitional characteristics that can effectively characterize different levels of privacy protection.

[0121] To further describe the determinism of the samples at the target privacy level, a lower approximation concept from fuzzy rough sets is introduced. For the true privacy level... The following approximation is defined as:

[0122] ;

[0123] This indicator measures whether a sample is far from all non-target levels. A larger value indicates that the sample more clearly belongs to a certain privacy level; conversely, a smaller value indicates that the sample is in the intersection of multiple levels.

[0124] Step 7: Based on the boundary domain size obtained in Step 6, perform boundary-aware privacy level inference and output the continuous privacy protection level score corresponding to the video.

[0125] Furthermore, to explicitly characterize the fuzzy boundary structure between privacy levels, a boundary domain function is introduced:

[0126] ;

[0127] in, Used to control boundary sensitivity.

[0128] This function maps the lower approximation to Intervals, thus forming continuous boundary metrics. When A value close to 0 indicates the sample is in a deterministic region; a value close to 1 indicates the sample is in a fuzzy transition region with varying privacy levels. Based on this boundary metric, a boundary-aware fuzzy inference mechanism is constructed to adaptively handle samples with different levels of uncertainty. During training, the overall optimization objective is defined as:

[0129] ;

[0130] in, Cross-entropy loss is used to constrain discriminative learning of deterministic samples; Label smoothing loss is used to enhance the probabilistic representation of boundary samples.

[0131] By introducing boundary weights The model can adaptively adjust the supervision strength based on sample uncertainty, achieving a dynamic transition from hard classification to soft inference. Furthermore, to avoid over-reliance on fuzzy regions, a boundary regularization term is introduced: This constraint is used to control the overall boundary region size, thereby improving the stability and generalization ability of the model.

[0132] The average probability distribution of each privacy protection level output on the test set is as follows: Figure 4 As shown.

[0133] To achieve a continuous characterization of privacy protection levels, this invention further maps the privacy level probability distribution output by the model to a continuous privacy protection score. Specifically, the model first outputs probability distribution vectors for different privacy levels. Subsequently, based on the ordered semantic relationships according to privacy levels, a linear weighted mapping strategy is introduced to map the discrete probability space to a continuous rating space. The mapping process is shown in the formula:

[0134] ;

[0135] To verify the sensitivity and discriminative ability of the proposed method to different compressed sensing privacy perturbation intensities, a continuous privacy protection score visualization analysis was performed on the original image and video samples under different privacy protection levels (CS1-CS4). The schematic diagram of the continuous privacy protection score based on probability distribution mapping in this invention is shown below. Figure 5 As shown, the score distribution corresponding to different privacy protection levels exhibits a significant hierarchical structure and progressive trend.

[0136] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for evaluating home video privacy by integrating VLM and differentiable rough set optimization, characterized in that, Includes the following steps: Step 1: Input the privacy-protected home video data and process the original home video with different intensities to generate video samples with different levels of privacy protection; Step 2: Using expert subjective evaluation samples as a cognitive benchmark, the VLM is iteratively calibrated through privacy evaluation prompts and feedback optimization mechanisms to construct a VLM model aligned and calibrated with human cognition; then, video samples with different levels of privacy protection are fed into the VLM model aligned and calibrated with human cognition to generate corresponding privacy protection level labels. Step 3: Input the privacy-protected video into the spatiotemporal dual-stream network to extract features; Step 4: Perform cross-modal alignment on the extracted features and use an attention mechanism to perform spatiotemporal information interaction and fusion to obtain spatiotemporal fusion features; Step 5: Represent the spatiotemporal fusion features of different privacy protection levels as learnable hierarchical prototypes with ordered structures, and constrain the arrangement relationship between adjacent hierarchical prototypes to construct a privacy level prototype space with hierarchical constraints, so that the privacy level representation conforms to the continuous change law from low to high. Step 6: Based on the ordered privacy level prototype space, use the upper and lower approximate adaptive characterization of differentiable fuzzy rough sets to depict the transition boundary between different privacy protection levels. Step 7: Based on the size of the boundary domain, perform boundary-aware privacy level inference for video privacy protection level, output the probability distribution of each privacy level, and finally map it into a continuous video privacy protection level score.

2. The method for evaluating home video privacy by integrating VLM and differentiable rough set optimization as described in claim 1, characterized in that, Step 2 specifically includes: S201. Construct a set of expert subjective evaluations, selecting them from the training set. A cognitive calibration dataset is composed of representative videos, processed with different levels of privacy protection. ; then invited Each video sample was scored for its level of privacy protection by a team of experts with relevant experience. The final average opinion score is obtained by employing an extreme value removal averaging strategy. ; in, Indicates the first The human subjective privacy score corresponding to each video is used as a real reference benchmark in the subsequent VLM cognitive calibration process to constrain the consistency between model output and human privacy perception. S202. Using VLM to probabilistically model video privacy levels: Let the set of privacy levels be These represent different levels of privacy protection for input videos. VLM based on the current prompt word Output the probability distribution corresponding to each privacy level. ,in Subsequently, the discrete 0-4 ratings were converted to a continuous 0-10 privacy score using an expectation mapping method. ； S203. Design a cue word calibration mechanism based on scoring error feedback: For the In the evaluation process, the VLM prediction error is defined as: ,in This indicates the expert MOS score. Indicates the first VLM prediction scores were then used; subsequently, high-biased samples were selected based on the magnitude of the error. ,in The error threshold is used to convert high-bias samples and their corresponding expert evaluation differences into structured feedback information. ,in This represents the feedback generation function; finally, the next round of prompts is updated based on the generated feedback information. ,in This indicates a context fusion operation; S204. Use correlation metrics to measure the consistency between model predictions and expert MOS to evaluate whether VLM has achieved cognitive alignment: Spearman's rank correlation coefficient was used to evaluate the consistency of predicted rankings. The Pearson correlation coefficient was used to evaluate the linear consistency of the predicted scores. Then when and And in multiple iterations At that time, it is considered that VLM has reached a stable state of human cognitive alignment; S205. Input video samples with different levels of privacy protection into the calibrated VLM model to generate corresponding privacy protection level labels; wherein, the privacy protection level labels include five levels from 0 to 4, which correspond to five privacy protection levels: none, weak, medium, good, and excellent, respectively, and are used to characterize the overall privacy protection level of the video.

3. The method for evaluating home video privacy by integrating VLM and differentiable rough set optimization as described in claim 1, characterized in that, Step 3 specifically includes: S301, The spatiotemporal dual-stream network includes: Spatial branching is used to model the structural information and identifiability changes of a single frame image, and to characterize the impact of privacy protection operations on visual appearance. The temporal branch is used to model the motion dependencies and behavioral evolution characteristics between consecutive frames, reflecting the degree to which privacy protection undermines the consistency of temporal semantics. S302. Perform feature modeling on the input multi-layer compressed sensing processed video: Parallel receive batch size is Collection of privacy-protected video clips and the corresponding set of intermediate video frames extracted from it. As input, a spatiotemporal encoder is used to capture feature vectors containing information on motion continuity and temporal distortion from video clips. The appearance encoder is used to extract appearance feature vectors containing spatial resolution and pixel occlusion from a single frame image. .

4. The method for evaluating home video privacy by integrating VLM and differentiable rough set optimization as described in claim 1, characterized in that, Step 4 specifically includes: S401. Introduce a contrastive learning mechanism to analyze the spatiotemporal features from the same protected video segment. Appearance features Positive sample pairs are formed, and features from other videos within the same batch are formed negative sample pairs. By optimizing the InfoNCE loss function, two-dimensional features under the same privacy state are forced to be mapped to the same hypersphere. The symmetric alignment loss function is defined as follows: ； ； in, This represents the set of all target modal feature vectors within the same batch. This indicates temperature hyperparameters. It represents the mathematical expectation of the distribution; S402. A Transformer decoding layer with a cross-attention mechanism is introduced to perform deep interaction and fusion of features. In this module, the aligned 3D spatiotemporal features... Serving as query vectors, they guide the model to actively focus on and extract 2D appearance features as keys and values. The key spatial distortion information in the image; its linear mapping and attention calculation formulas are as follows: ； ； ； ； in, express , express ; Representation layer normalization; This represents the learnable weight matrix. Indicates the scaling dimension of the key vector; The complementary interaction features extracted by the cross-attention module are then linked with the original spatiotemporal features via residual connections. The mixture is then added together and fused before being fed into a multilayer perceptron for processing. ； ； Ultimately, the output fusion features Dimensionally reduced by MLP Compact feature space of dimensionality .

5. The method for evaluating home video privacy by integrating VLM and differentiable rough set optimization as described in claim 1, characterized in that, Step 5 specifically includes: S501, Extracting dual-stream fusion features Next, we first construct a learnable, ordered privacy level prototype space and introduce... Each privacy level prototype vector: Each prototype vector corresponds to a semantic center state with a privacy level, and the privacy levels increase progressively from low to high. S502. Introduce chained ordering constraints to explicitly constrain the relative distance relationships between prototypes. The privacy level prototype space is defined as follows: 。 6. The method for evaluating home video privacy by integrating VLM and differentiable rough set optimization as described in claim 1, characterized in that, Step 6 specifically includes: S601. After constructing the ordered prototype space, in order to characterize the sample features To establish a continuous association between samples and privacy levels, a fuzzy membership modeling method based on Gaussian kernel function is introduced, defining the relationship between samples and the first privacy level. The similarity between the prototypes at each level is: ;in, Indicates temperature parameter; The similarity function extends the hard determination of a sample belonging to a certain category to a representation of multiple privacy levels with different degrees of membership, thus forming a fuzzy membership vector: ; S602. Introducing the concept of lower approximation in fuzzy rough sets, then for the true privacy level... The following approximation is defined as: ; This metric measures whether a sample is simultaneously far from all non-target levels.

7. The method for evaluating home video privacy by integrating VLM and differentiable rough set optimization as described in claim 1, characterized in that, Step 7 specifically includes: S701. To explicitly characterize the fuzzy boundary structure between privacy levels, a boundary domain function is introduced: ； in, Used to control boundary sensitivity; this function maps the lower approximation to... Intervals, thus forming continuous boundary measures; at this time when When the value is close to 0, it indicates that the sample is in the deterministic region; when it is close to 1, it indicates that the sample is in the fuzzy transition region of the privacy level. Based on the aforementioned boundary domain function, a boundary-aware fuzzy inference mechanism is constructed. During training, the overall optimization objective is defined as: ； in, For cross-entropy loss, Label smoothing loss; Then, a boundary regularization term is introduced: , used to control the overall boundary area size; S702. Map the privacy level probability distribution output by the model to a continuous privacy protection score. Specifically, the model first outputs probability distribution vectors for different privacy levels. Subsequently, based on the ordered semantic relationships of privacy levels, a linear weighted mapping strategy is introduced to map the discrete probability space to the continuous scoring space.

Images