Privacy protection advertisement monitoring method and device and storage medium

Abstract

The invention relates to the technical field of advertisement monitoring methods, in particular to a privacy-protection advertisement monitoring method and device and a storage medium, and the method specifically comprises the following steps: step 1, multi-source monitoring data collection and localized desensitization: an advertisement media party locally collects advertisement monitoring basic data at a user side, data desensitization is completed in a user equipment local or media party server; 2, differential privacy noise injection and data fragmentation: injecting differential privacy noise into the desensitized data, and performing data fragmentation processing to ensure that a single piece of data cannot reversely deduce the real behavior of a user; according to the method, the risk of user privacy disclosure is effectively reduced, sensitive information acquisition is avoided from a data acquisition link, an association path of user identity and behavior data is cut off from the source through multiple means of localized desensitization, differential privacy noise injection and the like, concentrated transmission and storage of original data are avoided, and user experience is improved. The problem that sensitive information is exposed in traditional centralized monitoring is thoroughly solved.

Description

Technical Field

[0001] This invention relates to the field of advertising monitoring methods, and in particular to a privacy-preserving advertising monitoring method, device, and storage medium. Background Technology

[0002] Currently, the advertising monitoring field mainly relies on a centralized data collection and analysis model. This means that advertising platforms and media outlets need to centrally upload data such as users' browsing history, click behavior, and device information to a third-party monitoring server. By integrating multi-source data, core indicators such as ad exposure, click-through rate, and conversion rate are calculated.

[0003] While existing technologies can achieve quantitative monitoring of advertising effectiveness, they suffer from serious privacy protection flaws: on the one hand, centralized data collection exposes sensitive user information (such as unique device identifiers, geographical location, and consumption preferences) during multi-stage data transmission and storage, making it vulnerable to hacking or abuse by third parties; on the other hand, existing privacy protection methods are mostly simple data anonymization (such as deleting names and phone numbers), but users can still be re-identified through indirect information such as device identifiers and behavioral characteristics, resulting in limited privacy protection effectiveness. Summary of the Invention

[0004] In response to the technical problems mentioned in the background art, the present invention provides a privacy-protecting advertising monitoring method, device, and storage medium.

[0005] The technical solution adopted in this invention is: a privacy-protecting advertising monitoring method, specifically including the following steps:

[0006] Step 1: Multi-source monitoring data collection and localized anonymization. The advertising media collects basic advertising monitoring data locally on the user's device and completes data anonymization on the user's device or the media's server.

[0007] Step 2: Differential privacy noise injection and data fragmentation. Differential privacy noise is injected into the de-identified data, and data fragmentation is performed to ensure that a single data entry cannot be used to infer the user's true behavior.

[0008] Step 3: Distributed model training based on federated learning. Using a federated learning architecture, the media, advertising platform, and monitoring party jointly train the advertising effect prediction model without sharing the original data, so as to achieve collaborative analysis of multi-source data.

[0009] Step 4: Cross-platform advertising behavior association and attribution. Based on the trained federated learning model and de-identified data, cross-platform and cross-device advertising behavior association is achieved, and advertising conversion attribution results are calculated, while avoiding user identification.

[0010] Step 5: Blockchain storage and verification of monitoring data. To prevent monitoring data from being tampered with, blockchain technology is introduced to store key monitoring results, ensuring data credibility and traceability.

[0011] Step 6: Generate a monitoring report with privacy protection. Based on the monitoring data output by the federated learning model and stored on the blockchain, generate an advertising monitoring report that meets privacy protection requirements and avoids disclosing any sensitive user information in the report.

[0012] Step 7: Monitor system privacy compliance audit and optimization. Regularly conduct privacy compliance audits and model optimizations of the entire monitoring system to ensure that the system continues to comply with data security regulations.

[0013] In one embodiment, step 1 specifically includes the following:

[0014] Content collected;

[0015] The advertising media outlet collects basic advertising monitoring data locally on the user's device, including ad impression data and impression time T. e Advertising IDA i d. Media IDM i d. User behavior data, including click status C s Click time T c Conversion behavior identifier V s Basic equipment data, including equipment type D t It does not collect sensitive data such as device unique identifiers, user identity information, or precise geographical location.

[0016] Data anonymization is performed locally on the user's device or on the media provider's server. The core operations include:

[0017] Equipment information fuzzification: For equipment type D t With operating system O s The classification process is simplified to avoid precise identification by the equipment;

[0018] Timestamp desensitization: The exposure time T is desensitized using the time window offset method. e Click time T c The process is performed, and the formula is T. des =T org +Δt, where T des For the desensitized timestamp, T org Δt is the original timestamp, and Δt is the random offset.

[0019] In one embodiment, step 2 is specifically as follows:

[0020] Differential privacy noise injection:

[0021] The Laplace mechanism is used to inject noise into count-type metrics in user behavior data. The formula is as follows:

[0022] Where: X noise For the data after injecting noise, X true The original real data is used; Lap(·) is the Laplace distribution function, and its probability density function is... Δf represents the function sensitivity; ∈ represents the privacy budget;

[0023] Data sharding processing:

[0024] The monitoring data after noise injection is segmented according to the dimension of "ad ID + time window", and each data segment is S. A-t It includes anonymized monitoring data for a specific advertisement within a 1-hour time window, and each segment contains at least 50 user behavior data points to avoid inferring individual user behavior from small sample segments.

[0025] In one embodiment, step 3 is specifically as follows:

[0026] Federated learning architecture setup:

[0027] Advertising media:

[0028] Stores locally anonymized monitoring data and is responsible for local training of the model;

[0029] Server-side: A third-party advertising monitoring platform responsible for initializing model parameters, aggregating model gradients uploaded by clients, and updating global model parameters; it does not store any raw client data.

[0030] Local model training: A logistic regression model is used to predict the probability of ad clicks. The model loss function is:

[0031]

[0032] Where: n is the number of local training samples on the client; y i The true label for the sample; p i The click probability (p) of the i-th sample predicted by the model. i =σ(w T x i +b)), where σ is the Sigmoid function, w is the feature weight, and b is the bias term;

[0033] The client minimizes the loss function using gradient descent and calculates the gradient of the local model. Gradients are also clipped to prevent them from leaking user privacy.

[0034] Global model aggregation: The gradients from each client are aggregated using a weighted average method, with the following formula:

[0035]

[0036] Where: m is the number of clients participating in federated learning; n j Let N be the number of local samples from the j-th client; N be the total number of samples from all clients. g localj The local gradient uploaded by the j-th client; g global This represents the aggregated global gradient; the server uses g. global Update global model parameters w new =w old -η×g global η is the learning rate, which defaults to 0.01, and the updated model parameters are then sent to each client.

[0037] In one embodiment, step 4 is specifically as follows:

[0038] Behavioral association feature extraction: This involves extracting behavioral association features from anonymized data, including:

[0039] Ad content characteristics: Ad IDA i d. Advertising Category A c ;

[0040] Behavioral temporal characteristics: Exposure time T after desensitization des Click time T des Time difference ΔT = T c-des -T e-des ;

[0041] Equipment environment characteristics: fuzzy equipment type D t Operating System O s ;

[0042] Cross-platform behavior correlation calculation: Cosine similarity is used to calculate the correlation between user behaviors on different platforms. The formula is as follows:

[0043]

[0044] in: These are behavioral feature vectors for different platforms; It is the dot product of two vectors; Let be the magnitudes of the two vectors, respectively.

[0045] Sim behav The value range is 0-1. If Sim behav If the value is ≥0.7, it is determined to be cross-platform behavior of the same user, and correlation analysis can be performed; if Sim behavIf the value is less than 0.7, it is determined to be different user behavior and no association is made.

[0046] Transformation attribution calculation:

[0047] A multi-touchpoint attribution model is used to calculate the contribution of each advertisement to the conversion, using the following formula: Among them: Contrib A The conversion contribution of ad A; k is the total number of ads the user was exposed to before conversion; W i Let A be the weight of advertisement A in the i-th contact order.

[0048] In one embodiment, step 5 is specifically as follows:

[0049] Data generation for evidence preservation:

[0050] The core advertising metrics calculated in step 4 are: Impressions (Imp), Clicks (Click), Conversion Rate (CVR), and Contrib (Contrib). A With data fragment identifier S A-t Desensitization timestamp T des Integrate and generate evidence storage data blocks. data ;

[0051] Data hash calculation: for evidence storage data blocks data The formula for calculating the SHA-256 hash value is Hash = SHA256(Block). data +Nonco);

[0052] Where: Hash is a 256-bit hash value, uniquely identifying the evidence storage data block; Nonco is a random number, used to meet the difficulty requirements of blockchain proof-of-work; SHA256(·) is the SHA-256 hash function, which has one-wayness and collision resistance;

[0053] Blockchain-based evidence storage: This involves calculating the hash and storing the evidence in a block. data The index is uploaded to the consortium blockchain, and each node synchronously stores the blockchain data to achieve distributed evidence storage;

[0054] Data verification mechanism: When it is necessary to verify the authenticity of the monitoring data, the hash value of the current monitoring data block is recalculated. new =SHA256(Block) data-new +Nonce), and compare it with the hash stored on the blockchain: if the hash new =Hash, then the data has not been tampered with; if Hash new If the value is not equal to Hash, then the data has been tampered with, triggering an error alert.

[0055] In one embodiment, step 6 specifically includes the following:

[0056] Report data aggregation: Monitoring metrics are aggregated by the dimension of "advertiser-ad campaign-time period", and only statistical data is displayed. No segmented data for single users or small samples is displayed.

[0057] Outlier Handling: The Z-score method is used to identify and handle outliers in the monitoring data. The formula is as follows: Among them: Z i X is the Z-score value of the i-th data point; i Let Z be the monitored value of the i-th data point; μ be the mean of the monitored index; σ be the standard deviation of the monitored index; if |Z i If the value is greater than 3 (outside the 99.7% confidence interval), it is considered an outlier, and the median of the indicator is used to replace X. i ;

[0058] Report access control: Set hierarchical access for monitoring reports. Advertisers can only view monitoring data for their own ads, media outlets can only view ad exposure data for their own platforms, and monitoring parties have the highest level of access but must comply with the data access log recording system.

[0059] In one embodiment, step 7 is specifically as follows:

[0060] Privacy compliance audit: The system is audited quarterly by a third-party compliance agency. The audit includes the following:

[0061] Data collection scope: Check whether there is any collection of sensitive user data beyond the permitted scope;

[0062] Desensitization and noise injection effects: Calculating privacy protection strength using differential privacy auditing tools;

[0063] Data access logs: Check for any unauthorized data access records;

[0064] Blockchain Evidence Integrity: Verifying whether the hash value on the blockchain matches the monitoring data.

[0065] Model optimization and adjustments: Based on compliance audit results and advertising monitoring data, the core model and parameters of the system were optimized, including:

[0066] Privacy budget adjustment: If the audit finds insufficient privacy protection, the privacy budget will be reduced and the noise injection strength will be increased; if the availability of monitoring data is found to be too low, the budget will be increased to balance privacy and availability.

[0067] Federated learning model optimization: If the global model prediction accuracy is below 90%, increase the number of training samples or adjust the model structure, and repeat the model training process in step 3.

[0068] Behavior association threshold optimization: If the cross-platform behavior association accuracy is below 85%, adjust the association threshold Sim. behav ;

[0069] Optimization effect verification: After optimization, test and verify the effect: run the optimized monitoring system and the unoptimized system simultaneously, and collect the privacy protection strength, monitoring indicator accuracy and cross-platform correlation accuracy of the two groups. If the privacy protection strength of the experimental group increases by 210% and the monitoring accuracy decreases by ≤3%, the optimization is deemed effective and the optimized system is officially launched. Otherwise, readjust the optimization parameters until the expected effect is achieved.

[0070] The beneficial effects of this invention are as follows: Compared with the prior art, this invention effectively reduces the risk of user privacy leakage, avoids the collection of sensitive information from the data collection stage, and cuts off the association path between user identity and behavioral data from the source through multiple means such as localized desensitization and differential privacy noise injection, avoiding centralized transmission and storage of raw data, and completely solving the problem of sensitive information exposure in traditional centralized monitoring. It also introduces blockchain technology to distribute and store core monitoring indicators, avoiding the risk of data being tampered with at a single point in traditional centralized storage. While strengthening privacy protection, it ensures that the deviation of monitoring indicators is controlled within a reasonable range and avoids data distortion caused by privacy protection measures. Attached Figure Description

[0071] Figure 1 This is a flowchart illustrating the present invention; Detailed Implementation

[0072] In the description of this invention, it should be noted that the terms "front", "up", "down", "left", "right", "vertical", "horizontal", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0073] refer to Figure 1 In order to solve the problems existing in the background technology, this application proposes the following technical solution: a privacy-protected advertising monitoring method, device and storage medium;

[0074] Specifically, the following steps are included:

[0075] Step 1: Multi-source monitoring data acquisition and localized anonymization;

[0076] Content collected;

[0077] Basic advertising monitoring data is collected locally on the user's device by the advertising media provider (such as an app or website), including ad exposure data (exposure time T). e Advertising IDA i d. Media IDM i d) User behavior data (click status C) s (Click counts as 1, no click counts as 0), Click time T c Conversion behavior identifier V s (Converted to 1, not converted to 0)), Equipment Basic Data (Equipment Type D) t (such as mobile phones, PCs), operating system O s It does not collect sensitive data such as device unique identifiers (e.g., IMEI, MAC address), user identity information (e.g., name, mobile phone number), or precise geographical location.

[0078] Localized desensitization processing

[0079] Data anonymization is performed locally on the user's device or on the media provider's server. The core operations include:

[0080] Equipment information fuzzification: For equipment type D t With operating system O s Perform fuzzy classification processing, such as blurring "iPhone15iOS" to "smartphone iOS" to avoid precise device identification;

[0081] Timestamp desensitization: The exposure time T is desensitized using the time window offset method. e Click time T c The process is performed, and the formula is T. des =T org +Δt, where T d es is the de-identified timestamp, T o rg is the original timestamp, and Δt is a random offset (ranging from -300 seconds to 300 seconds, and all behavioral data of the same user uses the same Δt to ensure consistency of behavioral timing).

[0082] The function of the above technical solution is: T des =T org The +Δt formula avoids associating user identity with precise timestamps and behavioral sequences by using random time offsets, while maintaining the temporal logic of the same user behavior data. This ensures that the correlation between behaviors is not disrupted in subsequent monitoring and analysis, thus solving the privacy leakage risk caused by the precise retention of time information in traditional desensitization methods.

[0083] Step 2: Differential privacy noise injection and data fragmentation;

[0084] To further enhance data privacy, differential privacy noise is injected into the anonymized data, and data fragmentation is performed to ensure that a single data entry cannot be used to infer a user's actual behavior.

[0085] Differential privacy noise injection

[0086] The Laplace mechanism is used to inject noise into count-based metrics in user behavior data (such as the number of clicks or conversions for a particular advertisement). The formula is as follows:

[0087] Where: X noise For the data after injecting noise, X true The original real data is used; Lap(·) is the Laplace distribution function, and its probability density function is... Δf is the function sensitivity (set to 1 here, since the maximum impact of a single action on the counting metric is 1); ∈ is the privacy budget (range 0.1-1, the smaller ∈ is, the stronger the privacy protection and the lower the data availability, the default value is 0.5, balancing privacy and availability).

[0088] For example, the actual number of clicks on an advertisement is X. true =100, after injecting noise The final data falls between 96 and 104 with a 95% confidence level, thus both concealing the true data and maintaining the statistical trend of the data.

[0089] Data fragmentation processing

[0090] The monitoring data after noise injection is segmented according to the dimension of "ad ID + time window", and each data segment is S. A-t It includes anonymized monitoring data (such as impressions and clicks) for a specific advertisement within a 1-hour time window, and each segment contains at least 50 user behavior data (satisfying the k-anonymity principle, k=50), avoiding the inference of individual user behavior from small sample segments.

[0091] The function of the above technical solution is as follows:

[0092] The formula employs a rigorous differential privacy mechanism to ensure that the probability of changes in monitoring results after deleting or adding any user data is controlled within a reasonable range, mathematically guaranteeing that user privacy is not leaked. Data sharding combined with the k-anonymity principle further reduces the risk of identifying individual data, resolving the privacy risks associated with the direct transmission of raw data in traditional monitoring.

[0093] Step 3: Distributed model training based on federated learning;

[0094] By adopting a federated learning architecture, media outlets, advertising platforms, and monitoring entities jointly train an advertising effectiveness prediction model without sharing the original data, thereby enabling collaborative analysis of multi-source data.

[0095] Federated learning architecture setup:

[0096] The advertising media (such as APP server, website server) stores the locally anonymized monitoring data and is responsible for the local training of the model; - Server side: third-party advertising monitoring platform, responsible for initializing model parameters, aggregating model gradients uploaded by clients, updating global model parameters, and does not store any original client data.

[0097] Local model training (client-side): The client uses a logistic regression model to predict ad click probabilities. The model loss function is:

[0098] Where: n is the number of local training samples on the client (i.e., the number of user behavior data entries after anonymization); y i The true labels for the sample are (1 for clicks, 0 for no clicks); p i The click probability (p) of the i-th sample predicted by the model. i =σ(w T x i +b)), where σ is the Sigmoid function, w is the feature weight, and b is the bias term).

[0099] The client minimizes the loss function using gradient descent and calculates the gradient of the local model. And the gradient is clipped (gradient clipping threshold G). th r =1.0), to avoid gradient leakage of user privacy.

[0100] Global model aggregation (server-side)

[0101] The server uses a weighted average method to aggregate the gradients from each client, with the following formula:

[0102]

[0103] Where: m is the number of clients participating in federated learning (i.e., the number of advertising media); n j Let N be the number of local samples from the j-th client; N be the total number of samples from all clients. g localj The local gradient uploaded by the j-th client; g global This represents the aggregated global gradient. The server uses g. global Update global model parameters w new =w old -η×g global(η is the learning rate, defaulted to 0.01), and then distribute the updated model parameters to each client, repeating the "local training - gradient upload - global aggregation - parameter distribution" process until the model converges (the loss function decreases by less than 10). -5 ).

[0104] The function of the above technical solution is as follows:

[0105] The loss function (Loss) measures the deviation between the model's predictions and the true values, providing an optimization objective for local model training; the global gradient aggregation formula (g) global By weighting the sample size, clients with larger sample sizes have a greater impact on the global model, thus improving model accuracy. The entire federated learning process achieves "data remains stationary while the model moves," solving the privacy leaks and data silos caused by multi-source data sharing in traditional centralized monitoring.

[0106] Step 4: Cross-platform advertising behavior correlation and attribution;

[0107] Based on a pre-trained federated learning model and anonymized data, it enables cross-platform and cross-device association of advertising behavior, accurately calculates advertising conversion attribution results, and avoids user identification.

[0108] Behavioral association feature extraction: This involves extracting behavioral association features from anonymized data, including:

[0109] Ad content characteristics: Ad IDA i d. Advertising Category A c (such as clothing and home appliances);

[0110] Behavioral temporal characteristics: Exposure time T after desensitization des Click time T des Time difference ΔT = T c-des -T e-des ;

[0111] Equipment environment characteristics: fuzzy equipment type D t Operating System O s .

[0112] 2. Cross-platform behavioral correlation calculation uses cosine similarity to calculate the correlation between user behaviors on different platforms (such as mobile apps and PC websites). The formula is:

[0113]

[0114] in: These are behavioral feature vectors for different platforms (composed of advertising content features, behavioral time sequence features, and device environment features, with each feature having been standardized). It is the dot product of two vectors; These are the magnitudes of the two vectors, respectively.

[0115] Sim behav The value range is 0-1. If Sim beh av If the value is ≥0.7, it is determined to be cross-platform behavior of the same user, and correlation analysis can be performed; if Sim beh av If the value is less than 0.7, it is determined to be different user behavior and no association is made.

[0116] Transformation attribution calculation

[0117] A multi-touchpoint attribution model is used to calculate the contribution of each advertisement to the conversion, using the following formula: Among them: Contrib A Conversion contribution of ad A (value range 0-1); k is the total number of ads the user was exposed to before conversion; W i The weight of ad A in the i-th contact sequence (using time decay weight, i.e., the closer the contact time is to the conversion time, the greater the weight, such as the last contact ad W). i =0.4, the second to last W i =0.3, the rest W i =0.3 average score).

[0118] For example, if a user is exposed to ad A (the first ad), ad B (the second ad), and ad C (the third ad) before conversion, then the content of ad C... C =0.4, Contrib of Ad B B =0.3, Contrib of Ad A A =0.3, to ensure that the attribution results fairly reflect the role of each advertisement.

[0119] The function of the above technical solution is as follows:

[0120] Sim behav The formula achieves cross-platform association through behavioral characteristics rather than user identity information, avoiding privacy leaks caused by traditional association methods based on unique device identifiers; Contrib A The formula uses multi-touchpoint attribution, which solves the problem that traditional "last click attribution" ignores the contribution of earlier advertising. At the same time, it is based on desensitized data after association to ensure that the attribution process does not involve user privacy.

[0121] Step 5: Monitoring data blockchain storage and verification

[0122] To prevent monitoring data from being tampered with, blockchain technology is introduced to store key monitoring results, ensuring data credibility and traceability.

[0123] Data generation for evidence preservation

[0124] The core advertising metrics calculated in step 4 (Imp, Click, CVR, Conversion Contribution of Each Ad) A ) and data fragment identifier S A-t Desensitization timestamp T des Integrate and generate evidence storage data blocks. data

[0125] Data hash calculation;

[0126] For evidence storage data blocks data The formula for calculating the SHA-256 hash value is Hash = SHA256(Block). data +Nonco);

[0127] Wherein: Hash is a 256-bit hash value, uniquely identifying the evidence storage data block; Nonco is a random number (range 0-100000), used to meet the difficulty requirements of blockchain Proof-of-Work (PoW) (e.g., the first 4 bits of the hash value are 0); SHA256(·) is the SHA-256 hash function, which has one-way property (it is impossible to deduce the original data from the hash value) and collision resistance (it is difficult for different original data to generate the same hash value).

[0128] Blockchain on-chain evidence storage

[0129] Calculate the Hash and store the evidence data block. data The index (rather than the complete data, to avoid storing sensitive information on the chain) is uploaded to the consortium blockchain (the nodes are jointly maintained by advertisers, media outlets, and monitoring parties). Each node synchronously stores the blockchain data to achieve distributed evidence storage.

[0130] Data verification mechanism

[0131] When it is necessary to verify the authenticity of monitoring data, the hash value of the current monitoring data block is recalculated. new =SHA256(Block) data-new +Nonce), and compare it with the hash stored on the blockchain: if the hash new =Hash, then the data has not been tampered with; if Hash new If the value is not equal to Hash, then the data has been tampered with, triggering an error alert.

[0132] In the above technical solution, Hash = SHA256 (Block) dataThe +Nonce formula ensures that the stored data cannot be tampered with through the one-wayness and collision resistance of the hash function; the distributed storage of blockchain solves the problem of data being easily tampered with at a single point in traditional centralized storage, improves the credibility and traceability of monitoring data, and meets advertisers' needs for the authenticity of monitoring results.

[0133] Step 6: Generating a monitoring report with privacy protection

[0134] Based on the monitoring data output by the federated learning model and stored on the blockchain, an advertising monitoring report that meets privacy protection requirements is generated to avoid disclosing any sensitive user information in the report.

[0135] Report data aggregation

[0136] The monitoring metrics are aggregated by the dimensions of "advertiser-ad campaign-time period" and only statistical data (such as the total exposure and average click-through rate of an advertiser's campaign in 7 days) are displayed. No segmented data of single users or small samples (sample size <50) are displayed to avoid inferring user privacy through segmented data.

[0137] Outlier handling

[0138] The Z-score method is used to identify and process outliers in the monitoring data. The formula is as follows: Among them: Z i X is the Z-score value of the i-th data point; i σ is the monitoring value of the i-th data point (e.g., the number of clicks in a certain hour); μ is the mean of the monitoring indicator; σ is the standard deviation of the monitoring indicator.

[0139] If |Z i If the value is greater than 3 (outside the 99.7% confidence interval), it is considered an outlier, and the median of the indicator is used to replace X. i To avoid outliers affecting the accuracy of the report.

[0140] Report access control

[0141] The monitoring reports are set with tiered access permissions. Advertisers can only view the monitoring data of their own ads, media outlets can only view the ad exposure data of their own platforms, and the monitoring party has the highest permissions but must comply with the data access log recording system (recording the access personnel, access time, and access content) to ensure that the data is not accessed without authorization.

[0142] The function of the above technical solution is as follows:

[0143] The formula uses statistical methods to accurately identify abnormal data, avoiding outliers that could distort monitoring reports. Data aggregation and access control further strengthen privacy protection from the report presentation and access stages, ensuring that monitoring reports only provide necessary statistical information and do not disclose individual user data.

[0144] Step 7: Monitoring System Privacy Compliance Audit and Optimization

[0145] Regularly conduct privacy compliance audits and model optimizations of the entire monitoring system to ensure that the system continuously complies with data security regulations and improves monitoring accuracy and privacy protection.

[0146] Privacy compliance audit

[0147] The system is audited quarterly by a third-party compliance agency, and the audit content includes:

[0148] Data collection scope: Check whether there is any collection of sensitive user data beyond the permitted scope;

[0149] Desensitization and noise injection effects: Calculate the privacy protection strength (e.g., whether the privacy budget is met) using differential privacy auditing tools;

[0150] Data access logs: Check for any unauthorized data access records;

[0151] Blockchain Evidence Integrity: Verifying whether the hash value on the blockchain matches the monitoring data.

[0152] Model optimization and adjustment

[0153] Based on the compliance audit results and advertising monitoring data, the core model and parameters of the system were optimized:

[0154] Privacy Budget Adjustment: If the audit finds that the privacy protection is insufficient (e.g., user data is still at risk of being identified), the privacy budget ∈ will be reduced (e.g., from 0.5 to 0.3) to increase the noise injection strength; if the availability of monitoring data is found to be too low (e.g., the indicator deviation exceeds 5%), ∈ will be appropriately increased (e.g., from 0.5 to 0.7) to balance privacy and availability.

[0155] Federated learning model optimization: If the global model prediction accuracy is below 90%, increase the number of training samples (expand the scope of client participation) or adjust the model structure (such as introducing a random forest model to replace logistic regression), and re-execute the model training process in step 3.

[0156] Behavioral association threshold optimization: If the cross-platform behavioral association accuracy (the degree of matching between association results and actual user behavior) is below 85%, adjust the association threshold Sim. beh av (For example, adjust from 0.7 to 0.75) to reduce false or missed associations.

[0157] Optimization effect verification

[0158] After optimization, the effect was verified through A / B testing: the optimized monitoring system (experimental group) and the unoptimized system (control group) were run simultaneously, and the privacy protection strength (assessed by differential privacy auditing tools), the accuracy of monitoring indicators (such as the deviation rate between the calculated value and the actual value of click-through rate), and the cross-platform correlation accuracy were collected for the two groups. If the privacy protection strength of the experimental group increased by 210% and the monitoring accuracy decreased by ≤3%, the optimization was deemed effective and the optimized system was officially launched. Otherwise, the optimization parameters were readjusted until the expected effect was achieved.

[0159] During model optimization, the privacy budget ∈ and the correlation threshold Sim are dynamically adjusted. behav These parameters address the imbalance between privacy protection and data availability during system operation; the A / B testing verification mechanism ensures that optimization measures are scientific and effective, avoiding blind adjustments that could lead to a decline in system performance, and providing a guarantee for the long-term stable operation of the monitoring system.

[0160] In summary, the specific beneficial effects of this embodiment are as follows: First, it effectively reduces the risk of user privacy leakage. This method avoids the collection of sensitive information from the data collection stage. Through multiple means such as localized desensitization (timestamp offset, device information blurring) and differential privacy noise injection (Laplace mechanism), it cuts off the association path between user identity and behavioral data from the source. Moreover, the federated learning architecture realizes "the model moves while the data remains still," avoiding the centralized transmission and storage of raw data, and completely solving the problem of sensitive information exposure in traditional centralized monitoring. By integrating multi-source data from media outlets, advertising platforms, and monitoring parties through federated learning, a global model is trained without sharing the original data. Combined with a cosine similarity association algorithm based on behavioral features, it accurately identifies user advertising behavior across platforms and devices, solving the problem of incomplete monitoring caused by the inability to share data among multiple parties in traditional monitoring. It can accurately calculate user behavior on different platforms. The system provides advertisers with a more comprehensive basis for evaluating the conversion effects after ad exposure; thirdly, it enhances the credibility and traceability of monitoring data by introducing blockchain technology for distributed storage of core monitoring indicators. The SHA-256 hash function ensures that the data cannot be tampered with, and each node synchronously stores the storage information, avoiding the risk of single-point tampering in traditional centralized storage. Simultaneously, the hash value verification mechanism can quickly verify the authenticity of the data, meeting advertisers' needs for the credibility of monitoring results; fourthly, it ensures the availability and accuracy of monitoring data. Mechanisms such as dynamic adjustment of the privacy budget in differential privacy, continuous optimization of the federated learning model, and Z-score outlier handling strengthen privacy protection while ensuring that deviations in monitoring indicators (such as impressions, click-through rate, and conversion rate) are controlled within a reasonable range, avoiding data distortion due to privacy protection measures, and providing reliable data support for optimizing advertising strategies.

[0161] Although embodiments of the invention have been shown and described, the scope of the invention will be defined by the appended claims and their equivalents by those skilled in the art.

Claims

1. A privacy-preserving advertising monitoring method, characterized in that, Specifically, the following steps are included: Step 1: Multi-source monitoring data collection and localized anonymization. The advertising media collects basic advertising monitoring data locally on the user's device and completes data anonymization on the user's device or the media's server. Step 2: Differential privacy noise injection and data fragmentation. Differential privacy noise is injected into the de-identified data, and data fragmentation is performed to ensure that a single data entry cannot be used to infer the user's true behavior. Step 3: Distributed model training based on federated learning. Using a federated learning architecture, the media, advertising platform, and monitoring party jointly train the advertising effect prediction model without sharing the original data, so as to achieve collaborative analysis of multi-source data. Step 4: Cross-platform advertising behavior association and attribution. Based on the trained federated learning model and de-identified data, cross-platform and cross-device advertising behavior association is achieved, and advertising conversion attribution results are calculated, while avoiding user identity verification. Step 5: Blockchain storage and verification of monitoring data. To prevent monitoring data from being tampered with, blockchain technology is introduced to store key monitoring results, ensuring data credibility and traceability. Step 6: Generate a monitoring report under privacy protection. Based on the monitoring data output by the federated learning model and stored on the blockchain, generate an advertising monitoring report that meets privacy protection requirements and avoids disclosing any sensitive user information in the report. Step 7: Monitor system privacy compliance audit and optimization. Regularly conduct privacy compliance audits and model optimizations of the entire monitoring system to ensure that the system continues to comply with data security regulations.

2. The privacy-protecting advertising monitoring method according to claim 1, characterized in that, Step 1 is detailed as follows: Content collected; The advertising media outlet collects basic advertising monitoring data locally on the user's device, including ad impression data and impression time T. e Advertising IDA i d. Media IDM i d. User behavior data, including click status C s Click time T c Conversion behavior identifier V s Basic equipment data, including equipment type D t It does not collect sensitive data such as device unique identifiers, user identity information, or precise geographical location. Data anonymization is performed locally on the user's device or on the media provider's server. The core operations include: Equipment information fuzzification: For equipment type D t With operating system O s The classification process is simplified to avoid precise identification by the equipment; Timestamp desensitization: The exposure time T is desensitized using the time window offset method. e Click time T c The process is performed, and the formula is T. des =T org +Δt, where T des For the desensitized timestamp, T org Δt is the original timestamp, and Δt is the random offset.

3. The privacy-protecting advertising monitoring method according to claim 2, characterized in that, Step 2 is detailed as follows: Differential privacy noise injection: The Laplace mechanism is used to inject noise into count-type metrics in user behavior data. The formula is as follows: Where: X noise For the data after injecting noise, X true The original real data is used; Lap(·) is the Laplace distribution function, and its probability density function is... Δf represents the function sensitivity; ∈ represents the privacy budget; Data sharding processing: The monitoring data after noise injection is segmented according to the dimension of "ad ID + time window", and each data segment is S. A-t It includes anonymized monitoring data for a specific advertisement within a 1-hour time window, and each segment contains at least 50 user behavior data points to avoid inferring individual user behavior from small sample segments.

4. The privacy-protecting advertising monitoring method according to claim 3, characterized in that, Step 3 is detailed below: Federated learning architecture setup: Advertising media outlets: Store locally anonymized monitoring data and are responsible for local training of the model; Server-side: A third-party advertising monitoring platform, responsible for initializing model parameters, aggregating model gradients uploaded by clients, and updating global model parameters, but does not store any raw client data; Local model training: A logistic regression model is used to predict the probability of ad clicks. The model loss function is: Where: n is the number of local training samples on the client; y i The true label for the sample; p i The click probability (p) of the i-th sample predicted by the model. i =σ(w T x i +b)), where σ is the Sigmoid function, w is the feature weight, and b is the bias term; The client minimizes the loss function using gradient descent and calculates the gradient of the local model. Gradients are also clipped to prevent them from leaking user privacy. Global model aggregation: The gradients from each client are aggregated using a weighted average method, with the following formula: Where: m is the number of clients participating in federated learning; n j Let N be the number of local samples from the j-th client; N be the total number of samples from all clients. g localj The local gradient uploaded by the j-th client; g global This represents the aggregated global gradient; the server uses g. global Update global model parameters w new =w old -η×g global η is the learning rate, which defaults to 0.01, and the updated model parameters are then sent to each client.

5. The privacy-protecting advertising monitoring method according to claim 4, characterized in that, Step 4 is detailed below: Behavioral association feature extraction: This involves extracting behavioral association features from anonymized data, including: Ad content characteristics: Ad IDA i d. Advertising Category A c ; Behavioral temporal characteristics: Exposure time T after desensitization des Click time T des Time difference ΔT = T c-des -T e-des ; Equipment environment characteristics: fuzzy equipment type D t Operating System O s ; Cross-platform behavior correlation calculation: Cosine similarity is used to calculate the correlation between user behaviors on different platforms. The formula is as follows: in: These are behavioral feature vectors for different platforms; It is the dot product of two vectors; Let be the magnitudes of the two vectors, respectively. Sim behav The value range is 0-1. If Sim behav If the value is ≥0.7, it is determined to be cross-platform behavior of the same user, and correlation analysis can be performed; if Sim behav If the value is less than 0.7, it is determined to be different user behavior and no association is made. Transformation attribution calculation: A multi-touchpoint attribution model is used to calculate the contribution of each advertisement to the conversion, using the following formula: Among them: Contrib A The conversion contribution of ad A; k is the total number of ads the user was exposed to before conversion; W i Let A be the weight of advertisement A in the i-th contact order.

6. The privacy-protecting advertising monitoring method according to claim 5, characterized in that, Step 5 is detailed below: Data generation for evidence preservation: The core advertising metrics calculated in step 4 are: Impressions (Imp), Clicks (Click), Conversion Rate (CVR), and Contrib (Contrib). A With data fragment identifier S A-t Desensitization timestamp T des Integrate and generate evidence storage data blocks. data ; Data hash calculation: for evidence storage data blocks data The formula for calculating the SHA-256 hash value is Hash = SHA256(Block). data +Nonco); Where: Hash is a 256-bit hash value, uniquely identifying the evidence storage data block; Nonco is a random number, used to meet the difficulty requirements of blockchain proof-of-work; SHA256(·) is the SHA-256 hash function, which has one-wayness and collision resistance; Blockchain-based evidence storage: This involves calculating the hash and storing the evidence in a block. data The index is uploaded to the consortium blockchain, and each node synchronously stores the blockchain data to achieve distributed evidence storage; Data verification mechanism: When it is necessary to verify the authenticity of the monitoring data, the hash value of the current monitoring data block is recalculated. new =SHA256(Block) data-new +Nonce), and compare it with the hash stored on the blockchain: if the hash... new =Hash, then the data has not been tampered with; if Hash new If the value is not equal to Hash, then the data has been tampered with, triggering an error alert.

7. The privacy-protecting advertising monitoring method according to claim 6, characterized in that, Step 6 is detailed below: Report data aggregation: Monitoring metrics are aggregated by the dimension of "advertiser-ad campaign-time period", and only statistical data is displayed. No segmented data for single users or small samples is displayed. Outlier Handling: The Z-score method is used to identify and handle outliers in the monitoring data. The formula is as follows: Among them: Z i X is the Z-score value of the i-th data point; i Let Z be the monitored value of the i-th data point; μ be the mean of the monitored index; σ be the standard deviation of the monitored index; if |Z i If |>3, it is considered an outlier, and the median of that indicator is used to replace X. i ; Report access control: Set hierarchical access for monitoring reports. Advertisers can only view monitoring data for their own ads, media outlets can only view ad exposure data for their own platforms, and monitoring parties have the highest level of access but must comply with the data access log recording system.

8. The privacy-protecting advertising monitoring method according to claim 7, characterized in that, Step 7 is detailed below: Privacy compliance audit: The system is audited quarterly by a third-party compliance agency. The audit includes the following: Data collection scope: Check whether there is any collection of sensitive user data beyond the permitted scope; Desensitization and noise injection effects: Privacy protection strength is calculated using differential privacy auditing tools; Data access logs: Check for any unauthorized data access records; Blockchain Evidence Integrity: Verifying whether the hash value on the blockchain matches the monitoring data. Model optimization and adjustment: Based on compliance audit results and advertising monitoring data, the core model and parameters of the system were optimized, including: Privacy budget adjustment: If the audit finds insufficient privacy protection, the privacy budget will be reduced and the noise injection strength will be increased; if the availability of monitoring data is found to be too low, the budget will be increased to balance privacy and availability. Federated learning model optimization: If the global model prediction accuracy is below 90%, increase the number of training samples or adjust the model structure, and repeat the model training process in step 3. Behavioral association threshold optimization: If the cross-platform behavioral association accuracy is below 85%, adjust the association threshold Sim. behav ; Optimization effect verification: After optimization, test and verify the effect: run the optimized monitoring system and the unoptimized system simultaneously, and collect the privacy protection strength, monitoring indicator accuracy and cross-platform correlation accuracy of the two groups. If the privacy protection strength of the experimental group increases by 210% and the monitoring accuracy decreases by ≤3%, the optimization is deemed effective and the optimized system is officially launched. Otherwise, readjust the optimization parameters until the expected effect is achieved.

9. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory is used to store computer programs; and the processor is used to execute the programs stored in the memory to implement the privacy-protected advertising monitoring method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the privacy-protected advertising monitoring method according to any one of claims 1 to 8.

Images