User privacy data protection method for store-wide intelligent management platform

Abstract

The application relates to the technical field of user privacy data protection, in particular to a user privacy data protection method for a whole-store intelligent management platform, which comprises the following steps: collecting a user's song ordering record locally at a box office end, calculating a privacy sensitivity coefficient for each song ordering record; aggregating the user's song ordering records according to song style dimensions, constructing an original preference vector, calculating an average sensitivity coefficient of each style dimension, determining a privacy budget parameter, adding Laplace noise to each style dimension, and finally obtaining a user preference vector; calculating the similarity between users to obtain a near neighbor set; and determining the privacy score of each user and encrypting the on-demand record of each user. The application aims to better protect user privacy data.

Description

Technical Field

[0001] This application relates to the field of user privacy data protection technology, specifically to methods for protecting user privacy data in a store-wide intelligent management platform. Background Technology

[0002] In a smart management platform serving the entire store, to improve operational efficiency and marketing conversion, the KTV song selection system needs to integrate personalized recommendation algorithms. This involves aggregating multi-dimensional data such as users' historical song selection records, selection time, room type, and number of companions to build user music preference profiles in the cloud. Real-time song recommendations are then implemented based on collaborative filtering technology, significantly increasing song selection rates and user engagement. However, the historical song selection records relied upon by the recommendation system are highly sensitive behavioral data, containing private information such as users' personal interests, emotional states, and social circles. Without sufficient protection, users' song selection history can easily be leaked, leading to serious privacy risks. Therefore, ensuring the security and privacy of user data while providing personalized services has become a critical issue that urgently needs to be addressed in system design.

[0003] Since a user's song request history reflects their personal behavioral characteristics and privacy data, dynamic encryption of each user's behavioral data is necessary to prevent attackers from easily obtaining this sensitive information and conducting behavioral analysis. Traditional encryption algorithms typically rely on random functions to generate encryption private keys, but these private key parameters are unrelated to the user's specific information. This design may increase the risk of brute-force attacks. Summary of the Invention

[0004] In light of the above, it is necessary to provide methods for protecting user privacy data for the whole-store intelligent management platform to solve the aforementioned problems.

[0005] One embodiment of this application provides a method for protecting user privacy data in a store-wide intelligent management platform, the method comprising:

[0006] Collect users' on-demand records, and statistically analyze the time of each on-demand record. Set a time sensitivity coefficient, analyze the room information where the user is when making a song request, determine the context sensitivity coefficient, and combine the time sensitivity coefficient and the context sensitivity coefficient to determine the privacy sensitivity coefficient of each on-demand record.

[0007] Obtain all style tags corresponding to each user's on-demand record, aggregate the style tags of all user on-demand records, combine the distribution characteristics of the on-demand time of all on-demand records with each style tag, determine the value of each style dimension of the original preference vector, and combine the privacy sensitivity coefficient of all on-demand records for each style tag to determine the privacy budget of each style dimension.

[0008] Noise is added to each style dimension of the original preference vector to obtain a noisy preference vector; based on the numerical value of the noisy preference vector between any two users in each style dimension, and combined with the privacy budget of the two users in each style dimension, the similarity between the two users is determined.

[0009] Based on the similarity, obtain the nearest neighbor set for each user, analyze the distribution characteristics of the similarity between each user and users in the nearest neighbor set, determine the privacy score for each user, and encrypt the on-demand records for each user.

[0010] The time sensitivity coefficient is preset based on the time period in which each on-demand record is played; the preset value for 22:00~2:00 is greater than the preset value for 18:00~22:00, which is greater than the preset value for 12:00~18:00.

[0011] Specifically, determining the scenario sensitivity coefficient involves:

[0012] The negative correlation mapping result of the number of people in the box where each user's on-demand record is located is positively fused with the assignment result based on the box type to obtain the context sensitivity coefficient of each on-demand record; among them, the smaller the area of ​​the box, the larger the assignment result.

[0013] Specifically, determining the privacy sensitivity coefficient for each on-demand record involves:

[0014] The time sensitivity coefficient and context sensitivity coefficient of each on-demand record are weighted by a preset weight, and the weighted sum is used to obtain the privacy sensitivity coefficient of each on-demand record; the sum of the preset weights is equal to 1, and the weight of the time sensitivity coefficient is greater than the weight of the context sensitivity coefficient.

[0015] Specifically, determining the value of each style dimension of the original preference vector involves:

[0016] For each on-demand record, calculate the negative correlation mapping result of the ratio of on-demand time to the preset time decay period. For each style tag, obtain the sum of the negative correlation mapping results obtained from all on-demand records to get the value of each style dimension of the original preference vector.

[0017] Specifically, determining the privacy budget for each style dimension involves:

[0018] The negative correlation mapping result of the average sensitivity of all on-demand records with each style tag is positively fused with the preset basic privacy budget parameter to obtain the privacy budget for each style dimension.

[0019] The process of obtaining the noise-adding preference vector includes:

[0020] Add Laplacian random noise to the values ​​corresponding to each style dimension of the original preference vector to obtain the user preference vector;

[0021] Obtain the maximum value of each style dimension and the value 0 in the user preference vector, and denote it as the first maximum value; after normalizing all the first maximum values ​​obtained from the user preference vector, the resulting vector is denoted as the noisy preference vector.

[0022] Specifically, determining the similarity between any two users involves:

[0023] The negative of the mean of the privacy budgets of any two users in each style dimension is used as the exponent of an exponential function with the natural constant as the base, and the result of the exponential function is used as the style weight of any two users in each style dimension.

[0024] Obtain the maximum and minimum values ​​of the noise preference vectors of any two users in each style dimension;

[0025] The style weights are used as the weights of the maximum values ​​obtained in the corresponding style dimensions, and the weighted sums are obtained to get the first cumulative sum.

[0026] The style weights are used as the weights of the minimum values ​​obtained for the corresponding style dimensions, and the weighted sums are obtained to get the second cumulative sum.

[0027] The similarity between any two users is obtained by calculating the ratio of the second cumulative sum to the first cumulative sum.

[0028] Specifically, obtaining the nearest neighbor set for each user is a set consisting of a preset number of users with the highest similarity to each user.

[0029] Specifically, encrypting each user's on-demand records involves:

[0030] The average absolute deviation of the similarity between the current user and all users in its nearest neighbor set is calculated and used as the privacy score for the current user. Each user's privacy score is then used as one of the parameters for generating the private key of the encryption algorithm to encrypt each user's on-demand records.

[0031] This application has at least the following beneficial effects:

[0032] The system collects user video-on-demand records and calculates time sensitivity coefficients, analyzes context sensitivity coefficients, and determines privacy sensitivity coefficients. By analyzing the time and location of each user's video-on-demand record, the privacy sensitivity of each record is determined. Changes in time and context effectively reflect different user needs and behavioral patterns, especially video-on-demand behavior during sensitive periods (such as late at night or weekdays), which may have more privacy characteristics. By setting time and context sensitivity coefficients, the privacy levels of different user behaviors can be distinguished, and an appropriate level of privacy protection can be assigned to each record. Records with high privacy sensitivity can be encrypted with stricter strategies, reducing the chances of attackers inferring user lifestyles or behaviors by analyzing time and context.

[0033] The process involves acquiring and aggregating user style tags to determine the initial preference vector. Each playback record is assigned a style tag, and statistical information from these tags is aggregated to form a user's preference vector. This process, by understanding the user's music style preferences, further refines the foundation for personalized recommendations. Style tag aggregation typically considers the distribution characteristics of playback time to avoid directly exposing specific user behavior records, instead providing a higher level of preference information. This level of abstraction helps prevent the direct leakage of user behavior data.

[0034] By combining the privacy sensitivity coefficient of each style dimension and allocating a privacy budget, different levels of noise can be added to the values ​​of each style dimension. This process achieves differential privacy protection for user data, avoiding the precise identification of users' personalized needs. The setting of the privacy budget ensures that the degree of privacy protection is not entirely the same for different style dimensions, thus maintaining personalized recommendations while avoiding the excessive exposure of certain sensitive information. After adding noise, even if an attacker obtains the user's preference vector, it is difficult to accurately reconstruct the user's true behavioral patterns.

[0035] By adding noise to the original preference vector (differential privacy mechanism), a user's on-demand records are preserved while their specific behavior remains hidden, maintaining statistical characteristics. This step reduces the privacy risks associated with the leakage of precise data. Adding noise effectively perturbs the data, making it impossible for attackers to deduce the user's precise preferences. This not only protects the privacy of individual users but also reduces potential risks across the dataset, thereby improving the overall privacy protection capabilities of the system.

[0036] The similarity between users is calculated using the noisy preference vectors, further forming a set of nearest neighbors for each user. This similarity can be used for personalized recommendations, but due to the added noise, the actual preference information is not directly revealed. Since the similarity is calculated based on noisy data, even if an attacker obtains the similarity between a user and other users, it is difficult to accurately deduce the user's specific preferences. This reduces inference attacks based on user similarity, ensuring that the system does not reveal users' true information when making personalized recommendations.

[0037] By analyzing similarity, a privacy score is determined for each user, and this score is used to encrypt the user's viewing history. The privacy score determines the encryption strength and strategy based on each user's similarity and behavioral characteristics. The introduction of the privacy score makes the encryption process more personalized for each user. For users with higher privacy scores (i.e., users whose preferences differ significantly from the general public), the system applies stronger encryption protection. This encryption strategy ensures that sensitive user data is not leaked due to system vulnerabilities or attacks. Attached Figure Description

[0038] Figure 1 A flowchart illustrating the user privacy data protection method for the store-wide intelligent management platform provided in this application;

[0039] Figure 2 A flowchart for obtaining similarity between users provided in this application. Detailed Implementation

[0040] In the description of the embodiments in this application, the words "exemplary," "or," and "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the words "exemplary," "or," and "for example" is intended to present the relevant concepts in a specific manner.

[0041] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this application's specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0042] It should also be noted that the terms "first" and "second" in this application and its accompanying drawings are used to distinguish similar objects, rather than to describe a specific order or sequence. The methods disclosed in the embodiments of this application or the methods shown in the flowcharts include one or more steps for implementing the method. Without departing from the scope of protection of this application, the execution order of multiple steps can be interchanged, and some steps can also be deleted.

[0043] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0044] This application proposes a method for protecting user privacy data in a store-wide intelligent management platform, applied in the field of user privacy data protection technology. (See attached document.) Figure 1 The method includes the following steps:

[0045] S1: Collect users' on-demand records, and statistically analyze the time of each on-demand record. Set a time sensitivity coefficient, analyze the room information where the user is when making a song request, determine the context sensitivity coefficient, and combine the time sensitivity coefficient and the context sensitivity coefficient to determine the privacy sensitivity coefficient of each on-demand record.

[0046] This application takes into account that song request scenarios can reflect emotional state and identity characteristics to a certain extent. Therefore, it makes quantitative distinctions based on song request scenarios and performs privacy-sensitive quantification on each request record. This allows subsequent noise injection, budget allocation, and similarity calculation to be adaptively adjusted according to the real risk of leakage, thereby satisfying strong differential privacy protection while avoiding the loss of data availability.

[0047] The system locally collects users' playback records at the private room end, including song identifiers. On-demand timestamp Private room types Number of people in the same private room Calculate a privacy sensitivity coefficient for each video-on-demand record. Specifically:

[0048] The system collects users' playback records locally in the private room, without uploading the original data, including song identifiers. : No. The song ID for the next playback; playback timestamp Accurate to the minute, in the following format: Private room type : Includes small, medium, and large private rooms; number of people in the same private room .

[0049] Based on the collected data, a privacy sensitivity coefficient is calculated for each on-demand record. Specifically: First, a time sensitivity coefficient is preset according to the time period in which each on-demand record is played; then, the negative correlation mapping result of the number of people in the box where each user's on-demand record is located is positively fused with the assignment result based on the box type to obtain the context sensitivity coefficient for each on-demand record; the time sensitivity coefficient and context sensitivity coefficient of each on-demand record are weighted by preset weights, and the weighted sum is obtained to obtain the privacy sensitivity coefficient for each on-demand record.

[0050] The specific method for obtaining the time sensitivity coefficient is as follows: In this embodiment, the KTV operating hours are considered to be from 12:00 to 2:00 the next day. When the on-demand time is from 22:00 to 2:00 in the late night period, the preset value of 1 is used as the time sensitivity coefficient; when the on-demand time is from 18:00 to 22:00 in the evening period, the preset value of 0.6 is used as the time sensitivity coefficient; when the on-demand time is from 12:00 to 18:00 in the daytime period, the preset value of 0.3 is used as the time sensitivity coefficient. The preset value for the late night period is greater than the preset value for the evening period, which is greater than the preset value for the daytime period. The implementer can set the preset value according to the actual situation, and this application does not impose any restrictions on it.

[0051] The principle for obtaining the time sensitivity coefficient is as follows: different time sensitivity coefficients are assigned according to the time period corresponding to the playback record: during the late night, the activity level of the crowd is low and there is less external interference, so the song request behavior is more distinguishable and is assigned the highest baseline value; during the daytime, it is a period of highly structured social activities, group behavior tends to be similar and individual characteristics are weak, so the song request behavior is less distinguishable and is assigned the lowest baseline value; the evening time is in between the two and is assigned a medium baseline value.

[0052] The negative correlation mapping result of the number of people in the box where each user's on-demand record is located is as follows: Let the number of people in the box where each on-demand record is located be denoted as... The sum of the natural number 1 and the number of people in the private room is calculated. This sum is then used as the logarithm of a logarithmic function, and its reciprocal is calculated. The values ​​assigned to different private room types are as follows: small private rooms are assigned 1.2, medium private rooms are assigned 1.0, and large private rooms are assigned 0.7. The smaller the area of ​​the private room, the larger the assigned value. Implementers can adjust the assigned values ​​for different private room types according to this condition; this application does not impose any restrictions on this. In this embodiment, multiple variables are positively fused using a multiplication method. The scenario sensitivity coefficient is calculated using the formula... We obtained, among which, This indicates the result of the assignment of the private room type. This indicates the number of people in the private room where each video-on-demand record is located. Represents a logarithmic function.

[0053] The principle behind obtaining the contextual sensitivity coefficient is as follows: Based on the number of people in the private room and the type of room, the impact of the social context on privacy is calculated: small private rooms offer strong privacy, so the sensitivity is increased; medium-sized private rooms are mainly for regular gatherings, so they are set as the baseline; large private rooms are typically for large gatherings, and song selection focuses more on popular preferences, so the sensitivity is decreased. In the denominator... This indicates that as the number of participants increases, the degree of personalization in song requests decreases logarithmically. This is because of the law of diminishing marginal returns; the more people there are, the smaller the impact of adding more people. Therefore, logarithmic decrease is used, and adding 1 is to avoid division by zero.

[0054] Since timestamps can directly infer a user's daily routine and emotional state, while social contexts require additional information for inference, the risk of direct leakage is low. Therefore, in this embodiment, the time sensitivity coefficient is weighted. Set to 0.6, context sensitivity coefficient weight Setting it to 0.4 requires ensuring... And the two together are one.

[0055] S2: Obtain all style tags corresponding to each user's on-demand record, aggregate the style tags of all user on-demand records, combine the distribution characteristics of the on-demand time of all on-demand records with each style tag, determine the value of each style dimension of the original preference vector, and combine the privacy sensitivity coefficient of all on-demand records for each style tag to determine the privacy budget of each style dimension.

[0056] Traditional song request recommendation systems typically upload raw playback records directly, which may expose users' private data, allowing the server to access all user information. Another approach is to encrypt the uploaded playback records using encryption algorithms, thus protecting user privacy. However, this method requires managing key distribution and ciphertext computation, and also necessitates data decryption during the recommendation process, potentially leading to significant delays and impacting user experience. This application addresses this by injecting noise to transform the raw playback records into a fuzzy preference vector, ensuring the server can only access statistical features.

[0057] Aggregate users' playback history by song genre to build The original preference vector is defined in dimensional form, and the average sensitivity coefficient for each style dimension is calculated to determine the privacy budget parameter for that dimension, specifically:

[0058] The system obtains all user playback records and their privacy sensitivity coefficients, and aggregates them into an M-dimensional original preference vector according to the style tags of the song library, where M represents the number of style tags, which are obtained through song library classification. In one embodiment, M is set to 10, and the style tags include: pop, rap, electronic, light music, folk, rock, traditional Chinese style, jazz, metal, and classical.

[0059] First, all user playback records are mapped to an M-dimensional style tag space by song ID, obtaining the style dimension corresponding to each song. For each playback record, the negative correlation mapping result of the ratio of playback time to a preset time decay period is calculated. For each style tag, the sum of the negative correlation mapping results obtained from all playback records is obtained to get the value of each style dimension of the original preference vector. In this embodiment, the specific formula for the value of the j-th style dimension of the original preference vector is: ;in, This represents the set of all play records with the j-th style tag. It should be noted that a song may have multiple style tags. This represents the number of days since today for the i-th on-demand record with the j-th style tag. This indicates the time decay period, with a value of 30 days. This represents the negative correlation mapping result of the ratio.

[0060] The principle behind obtaining the values ​​of each dimension of the original preference vector is as follows: Each style dimension is accumulated based on playback time and frequency. Multiple playbacks of the same style reflect the overall strength of the style preference. The vector is then decayed over time, giving higher weight to recent playbacks to reflect the time-varying nature of user preferences—that is, users' musical tastes change over time. Normalize the result and add 1 to prevent division by zero.

[0061] For each style dimension, a privacy budget is determined based on the average privacy sensitivity of the songs it contains. The process for obtaining the privacy budget of the j-th component of the original preference vector is as follows: the negative correlation mapping result of the average sensitivity of all playback records with the j-th style tag is positively fused with a preset basic privacy budget parameter to obtain the privacy budget for the j-th style dimension. In this embodiment, the specific formula for the privacy budget of the j-th style dimension is: ,in, This represents the actual privacy budget for the j-th style tag; This represents the preset basic privacy budget parameter, with a value of 0.5. This represents the average sensitivity of the j-th style tag; This represents the negative correlation mapping result of the average sensitivity of all on-demand records with the j-th style tag.

[0062] The principle behind obtaining the privacy budget is as follows: Within the framework of differential privacy, a smaller privacy budget parameter value results in stronger privacy protection, meaning more noise is added, but lower data availability; conversely, a larger value leads to higher data availability but also a higher risk of privacy leakage. When the average privacy sensitivity of the j-th style tag is higher, the privacy budget is smaller, adding more noise to enhance privacy protection; conversely, it reduces noise addition, enhancing data availability and ensuring recommendation accuracy. Differential privacy is a well-known existing technology and will not be elaborated upon in this application.

[0063] S3: Add noise to each style dimension of the original preference vector to obtain a noisy preference vector; based on the numerical value of the noisy preference vector between any two users in each style dimension, and combined with the privacy budget of the two users in each style dimension, determine the similarity between the two users.

[0064] Furthermore, Laplace noise is added to each style dimension of the original preference vector, resulting in a noisy and normalized user preference vector. Since the privacy exposure of the same user varies across different style dimensions, applying uniform noise intensity to all styles can lead to insufficient protection in highly sensitive areas and low data availability in less sensitive areas. This application addresses this by using a hierarchical injection approach, allowing each style dimension to receive an independent privacy budget based on its average sensitivity, maximizing privacy protection while ensuring data availability.

[0065] The probability density of the Laplace distribution decreases exponentially, and Laplace noise is a random variable drawn from it. The noise level can be adjusted by controlling the scaling parameter to meet the requirements of differential privacy. Furthermore, its sampling only requires one scalar operation, resulting in low computational overhead. It does not require complex homomorphic or key management, effectively reducing recommendation latency.

[0066] Random noise is added to each style dimension of the original preference vector. For example, consider the value corresponding to the j-th style dimension of the user preference vector: ,in, Indicates from position parameter Scale parameters Sampling random noise in a Laplace distribution; This represents the numerical value corresponding to the j-th style dimension of the original preference vector. At this point, for any two user datasets that differ by only one song request record... and user dataset The probability distribution of the output satisfies the definition of differential privacy, meaning that even if an attacker knows all the information except for one record, they cannot obtain more than [a certain amount of information]. The confidence level is used to infer the record.

[0067] Furthermore, the maximum value of each style dimension and the value 0 in the user preference vector is obtained and denoted as the first maximum value; the vector formed by normalizing all the first maximum values ​​obtained from the user preference vector is denoted as the noisy preference vector.

[0068] The principle behind obtaining the noisy preference vector is as follows: negative values ​​are truncated using a maximization function. Since the Laplace distribution is symmetrical, it may produce negative values, but the values ​​in the noisy preference vector cannot be negative. Furthermore, to avoid additional errors introduced by taking absolute values, the vector is truncated to zero. Because the values ​​for the same style dimension can vary significantly among different users, normalization is performed to facilitate subsequent similarity calculations.

[0069] After receiving the noisy preference vectors of all users, the server needs to calculate the similarity between users, find the nearest neighbors of the target user, and generate a recommendation list based on the historical song requests of the nearest neighbors. Traditional collaborative filtering uses cosine similarity, but this method is sensitive to noise; the added random noise will significantly reduce the accuracy of the similarity. Therefore, the similarity measurement needs to tolerate noise.

[0070] Calculate users and users The weighted Jaccard similarity between users is used, where the weights are style weights. Specifically, this application uses a Jaccard similarity improved based on style weights for measurement, and the similarity between user u and user v is... The specific formula is: ,in, , These represent the numerical values ​​of the noisy preference vectors for users u and v in the j-th style dimension, respectively. This indicates the second cumulative sum; This represents the first cumulative sum; The style weights of the noisy preference vectors of users u and v in the j-th style dimension can be expressed by the formula... We obtained, among which, , Let represent the average sensitivity of user u and user v on the j-th style dimension, respectively. This represents an exponential function with the natural constant as the exponent.

[0071] The flowchart for obtaining similarity between users is as follows: Figure 2 As shown.

[0072] The principle behind weighted Jaccard similarity is as follows: the numerator is the minimum value in each style dimension of the two users, representing the common preference of the two users in the j-th style dimension, similar to the intersection in the Jaccard similarity formula; the denominator is the maximum value in each style dimension of the two users, representing the total preference of the two users in the j-th style dimension, similar to the union in the Jaccard similarity formula; the average of the average sensitivity of the two users in the j-th style dimension is used as the exponentially decaying weight. The higher the sensitivity, the greater the injected noise, and the lower the credibility of that dimension, so the weight is reduced; conversely, the higher the sensitivity, the smaller the injected noise, and the higher the credibility of that dimension, so the weight is increased.

[0073] S4: Based on the similarity, obtain the nearest neighbor set of each user, analyze the distribution characteristics of the similarity between each user and users in the nearest neighbor set, determine the privacy score of each user, and encrypt the on-demand records of each user.

[0074] Further filtering based on current users The users with the highest similarity are selected as the current user's nearest neighbors. Specifically:

[0075] Calculate the similarity between each pair of users, and find the K users most similar to the target user as the nearest neighbor set. The number of nearest neighbors is... V represents the number of users. In one embodiment, when a KTV has 2,500 members, K is 50.

[0076] The average absolute deviation of the similarity between the current user and all users in its nearest neighbor set is calculated as the current user's privacy score. It should be noted that a stronger privacy score indicates a higher degree of personalization, meaning a greater deviation from the general public. In other words, a higher privacy score indicates more important privacy characteristics of the user. This value is then applied to the private key generation parameters of the encryption algorithm to encrypt the user's song request history. In this embodiment, the elliptic cryptography algorithm is used; in other embodiments, the SM2 encryption algorithm can also be used. All encryption algorithms are well-known technologies, and this application will not elaborate on them.

[0077] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different blocks may also occur in a different order than disclosed in the description; sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. Each block in a block diagram and / or flowchart, and combinations of blocks in a block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0078] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for protecting user privacy data in a store-wide intelligent management platform, characterized in that: The method includes the following steps: Collect users' on-demand records, and statistically analyze the time of each on-demand record. Set a time sensitivity coefficient, analyze the room information where the user is when making a song request, determine the context sensitivity coefficient, and combine the time sensitivity coefficient and the context sensitivity coefficient to determine the privacy sensitivity coefficient of each on-demand record. Obtain all style tags corresponding to each user's on-demand record, aggregate the style tags of all user on-demand records, combine the distribution characteristics of on-demand time of all on-demand records with each style tag, determine the value of each style dimension of the original preference vector, and combine the privacy sensitivity coefficient of all on-demand records for each style tag to determine the privacy budget of each style dimension. Noise is added to each style dimension of the original preference vector to obtain a noisy preference vector; based on the magnitude of the noisy preference vector between any two users in each style dimension, and combined with the privacy budget of the two users in each style dimension, the similarity between the two users is determined. Based on the similarity, obtain the nearest neighbor set for each user, analyze the distribution characteristics of the similarity between each user and users in the nearest neighbor set, determine the privacy score for each user, and encrypt the on-demand records for each user; The determination of the similarity between any two users specifically involves: The negative of the mean of the privacy budgets of any two users in each style dimension is used as the exponent of an exponential function with the natural constant as the base, and the result of the exponential function is used as the style weight of any two users in each style dimension. Obtain the maximum and minimum values ​​of the noise preference vectors of any two users in each style dimension; The style weights are used as the weights of the maximum values ​​obtained in the corresponding style dimensions, and the weighted sums are obtained to get the first cumulative sum. The style weights are used as the weights of the minimum values ​​obtained for the corresponding style dimensions, and the weighted sums are obtained to get the second cumulative sum. The similarity between any two users is obtained by calculating the ratio of the second cumulative sum to the first cumulative sum.

2. The method for protecting user privacy data in a store-wide intelligent management platform as described in claim 1, characterized in that, The time sensitivity coefficient is preset based on the time period in which each on-demand record is played; the preset value for 22:00~2:00 is greater than the preset value for 18:00~22:00 and greater than the preset value for 12:00~18:

00.

3. The method for protecting user privacy data in a store-wide intelligent management platform as described in claim 1, characterized in that, The determination of the scenario sensitivity coefficient specifically involves: The negative correlation mapping result of the number of people in the box where each user's on-demand record is located is positively fused with the assignment result based on the box type to obtain the context sensitivity coefficient of each on-demand record; among them, the smaller the area of ​​the box, the larger the assignment result.

4. The method for protecting user privacy data in a store-wide intelligent management platform as described in claim 1, characterized in that, The determination of the privacy sensitivity coefficient for each on-demand record is specifically as follows: The time sensitivity coefficient and context sensitivity coefficient of each on-demand record are weighted by preset weights, and the weighted sum is used to obtain the privacy sensitivity coefficient of each on-demand record; the sum of the preset weights is equal to 1, and the weight of the time sensitivity coefficient is greater than the weight of the context sensitivity coefficient.

5. The method for protecting user privacy data in a store-wide intelligent management platform as described in claim 1, characterized in that, The determination of the value of each style dimension of the original preference vector is specifically as follows: For each on-demand record, calculate the negative correlation mapping result of the ratio of on-demand time to the preset time decay period. For each style tag, obtain the sum of the negative correlation mapping results obtained from all on-demand records to get the value of each style dimension of the original preference vector.

6. The method for protecting user privacy data in a store-wide intelligent management platform as described in claim 1, characterized in that, The determination of the privacy budget for each style dimension is as follows: The negative correlation mapping result of the average privacy sensitivity coefficient of all on-demand records with each style tag is positively fused with the preset basic privacy budget parameter to obtain the privacy budget for each style dimension.

7. The method for protecting user privacy data in a store-wide intelligent management platform as described in claim 1, characterized in that, The obtained noise-adding preference vector includes: Add Laplace random noise to the values ​​corresponding to each style dimension of the original preference vector to obtain the user preference vector; The maximum value between the corresponding value of each style dimension in the user preference vector and the value of 0 is obtained and denoted as the first maximum value. The vector formed by normalizing all the first maximum values ​​obtained from the user preference vector is denoted as the noisy preference vector.

8. The method for protecting user privacy data in a store-wide intelligent management platform as described in claim 1, characterized in that, The process of obtaining the nearest neighbor set for each user specifically refers to a set consisting of a preset number of users with the highest similarity to each user.

9. The method for protecting user privacy data in a store-wide intelligent management platform as described in claim 1, characterized in that, The encryption of each user's on-demand records is specifically as follows: The average absolute deviation of the similarity between the current user and all users in its nearest neighbor set is calculated and used as the current user's privacy score. Each user's privacy score is then used as one of the parameters for generating the private key of the encryption algorithm to encrypt each user's on-demand records.

Images