Spatial crowdsourcing fuzzy task recommendation method and system based on differential privacy
By employing K-anonymity technology, user-based collaborative filtering, and an optimized BERT2FTR model, combined with differential privacy protection, the problems of privacy leakage and computational overhead in spatial crowdsourcing task recommendation are solved, achieving efficient and accurate task recommendation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANKAI UNIV
- Filing Date
- 2026-02-05
- Publication Date
- 2026-05-19
AI Technical Summary
Existing technologies lack dual privacy protection in spatial crowdsourcing task recommendations, cannot achieve a balance between high recommendation accuracy and low computational overhead, and lack content-based filtering technology, resulting in inaccurate result distribution.
K-anonymity is used to obfuscate task requests, user-based collaborative filtering is used to aggregate similar requests, an optimized lightweight BERT2FTR model is used for task recommendation, and results are distributed through content-based filtering. Differential privacy noise is added for protection.
It achieves dual privacy protection for task requesters, improves recommendation accuracy and efficiency, reduces computational overhead, and ensures the accuracy and personalization of recommendation results.
Smart Images
Figure CN121636830B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of spatial crowdsourcing, and more particularly to a spatial crowdsourcing fuzzy task recommendation method and system based on differential privacy. Background Technology
[0002] As a fundamental function of spatial crowdsourcing, task recommendation enables task requesters to find suitable recommendations based on their task needs and interests. However, sensitive demand information may be leaked during the recommendation process, leading to serious privacy breaches. Furthermore, in typical SC systems, the task recommendation mechanism is the key link between task requesters and recommendation results, and its performance directly impacts system participation and data quality. Existing research often focuses on task recommendation performance to achieve high task completion rates, neglecting the risk of privacy breaches. Some studies also attempt to protect the privacy of task requesters or recommendation results, but fail to achieve a balance between privacy and recommendation efficiency. Moreover, the complex models introduced to ensure recommendation accuracy incur high computational costs, making them difficult to deploy in real-world crowdsourcing environments. Therefore, existing solutions generally face a dilemma: either sacrifice recommendation accuracy and efficiency to protect privacy, or bear the risk of privacy breaches in pursuit of recommendation performance. A balanced solution that simultaneously achieves strong privacy protection, high recommendation accuracy, and low computational cost is still lacking.
[0003] Currently, Chinese invention patent application CN201710619287.X discloses a recommendation method based on differential privacy protection, including: obtaining user ratings for items from a website; adding Laplace noise to the user rating data to obtain perturbed rating data; inputting the perturbed rating data into a probabilistic matrix factorization recommendation algorithm using the Laplace mechanism to obtain user factor matrices and item factor matrices; predicting each user's ratings for all items based on the user factor matrix and item factor matrix; and recommending the top-ranked items with the highest predicted ratings to the user. However, this related technology does not utilize K-anonymity and user-based collaborative filtering techniques for dual privacy protection of task requesters, thus lacking privacy; it does not utilize an optimized lightweight BERT2FTR model to achieve fuzzy task recommendation; it does not use content-based filtering for result distribution, thus lacking recommendation accuracy. Summary of the Invention
[0004] The technical problem solved by this invention is that related technologies do not utilize K-anonymity and user-based collaborative filtering techniques to provide dual privacy protection for task requesters, thus lacking privacy; they do not utilize an optimized lightweight BERT2FTR model to achieve fuzzy task recommendation; they do not use content-based filtering to achieve result distribution, thus lacking recommendation accuracy.
[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:
[0006] A spatial crowdsourced fuzzy task recommendation method based on differential privacy includes the following steps:
[0007] Step S1: Use K-anonymity technology to obfuscate the task request based on the task requester's task request to obtain an obfuscated request and submit it to the fog node;
[0008] Step S2: The fog node receives fuzzy requests, aggregates the fuzzy requests using user-based collaborative filtering technology to obtain group requests, and sends the group requests to the cloud server.
[0009] Step S3: The cloud server uses the BERT2FTR model to recommend tasks to the group requests, obtains the recommendation results, and sends them to the fog nodes.
[0010] In step S4, the fog node distributes the recommendation results to the corresponding task requesters using content-based filtering technology.
[0011] Preferably, step S1 specifically includes:
[0012] Step S11: Obtain the task request through the task requester, and extract the sensitive information fields in the task request to obtain the quasi-identifier;
[0013] Step S12: Generalize the quasi-identifiers, group the generalized quasi-identifiers and process them using K-anonymity technology to obtain fuzzy requests.
[0014] Preferably, step S2 specifically includes:
[0015] Step S21: Use the BERT2FTR model to extract the feature values of fuzzy requests and transform each task request into a fuzzy request feature vector.
[0016] Step S22: Using user-based collaborative filtering technology, calculate the cosine similarity of the fuzzy request feature vectors. The calculation expression is as follows:
[0017] ;
[0018] in, and The fuzzy request feature vector of the task requester;
[0019] Step S23: Calculate the cosine similarity of all fuzzy request feature vectors and obtain the similarity matrix;
[0020] Step S24: Use the K-means clustering algorithm to aggregate fuzzy requests with similarity higher than the similarity threshold into a group of requests;
[0021] If any fuzzy request has a similarity to all other fuzzy requests that is less than the similarity threshold, then that fuzzy request is treated as a group request.
[0022] Preferably, step S3 specifically includes:
[0023] Fog nodes send group requests to the cloud server and use the BERT2FTR model for task recommendation. The BERT2FTR model processing includes an input layer, an embedding layer, a Transformer layer, and a classification output layer.
[0024] The input layer processing logic specifically includes:
[0025] The group requests are segmented using a preset word segmenter to obtain a sequence of sub-word units. The sub-word unit sequence is then converted to lowercase and filtered for irrelevant symbols.
[0026] A pre-defined special marker is inserted into the processed sub-word unit sequence to obtain a standardized marker sequence;
[0027] An initial tensor is generated based on a canonical tag sequence. The initial tensor includes a word tensor, a sentence vector, and an attention mask tensor.
[0028] Identify the effective word positions and filling positions in the canonical tag sequence, set the corresponding effective word position value in the attention mask tensor to 1, and set the corresponding filling position value to 0;
[0029] The generated lexical tensor and attention mask tensor are used as the original input token sequence.
[0030] Preferably, the Embedding layer processing logic specifically includes:
[0031] Obtain the original input label sequence and process it using a tail truncation strategy. The tail truncation strategy processing logic includes:
[0032] If the length of the original input label sequence exceeds the preset maximum length Then only the last one of the original input tokens is retained. 1 tag, as a sequence of tags to be processed;
[0033] Each token in the sequence of tokens to be processed is converted into a word embedding vector;
[0034] Assign a learnable position embedding vector to each position in the sequence of tags to be processed, and add the word embedding vector to the learnable position embedding vector to obtain the input representation;
[0035] Obtain all input representations to get the input representation sequence.
[0036] Preferably, the Transformer layer processing logic specifically includes:
[0037] The input representations are passed sequentially. A bidirectional Transformer encoder is used for feature extraction, and each layer of the bidirectional Transformer encoder outputs the hidden representation of the current layer.
[0038] Within each Transformer encoder layer, bidirectional dependencies between any tokens in the input representation sequence are calculated using a multi-head self-attention mechanism.
[0039] The output of the multi-head self-attention mechanism is nonlinearly transformed through a feedforward network. Residual connections and layer normalization operations are then performed on the output of the nonlinear transformation to obtain the hidden representation of the current layer.
[0040] Preferably, the classification output layer processing logic specifically includes:
[0041] The mean pooling method is applied to all hidden representations of the current layer output by the Transformer layer to generate the overall sequence representation. The mathematical expression for mean pooling is:
[0042] ;
[0043] in, For the overall representation of the sequence, To preset the maximum length, For the first The hidden representation of the i-th positional word in the layer Transformer encoder;
[0044] The sequence is represented as a whole by the task-specific classification head, and mapped to the probability distribution of each task category through a linear transformation and a softmax function. The expression for calculating the probability distribution is as follows:
[0045] ;
[0046] in, This represents the probability distribution vector for the task category. A set of tags for classifying tasks. The learnable weight matrix for the classification layer. The learnable bias vector for the classification layer;
[0047] The task category corresponding to the highest probability distribution vector is selected as the recommendation result.
[0048] Preferably, step S4 specifically includes:
[0049] Step S41: Construct a request index for the task requester in the fog node. The request index includes the distance range, time span, and request type of the request.
[0050] Step S42: Store the recommendation results in the form of a report to obtain a recommendation report, which includes the name, location, type, open time and evaluation characteristics of the results;
[0051] Step S43: The content-based filtering technology calls the request index, iterates through the matching degree between the recommendation report and the request index, and distributes the recommendation reports with a matching degree greater than the threshold to the corresponding task requesters.
[0052] Preferably, differential privacy stochastic gradient descent is applied during the training of the BERT2FTR model to obtain the same... - A differential privacy-preserving BERT2FTR model; where, To limit the loss of privacy, For privacy failure probability.
[0053] A spatial crowdsourced fuzzy task recommendation system based on differential privacy includes a fuzzy processing module, an aggregation module, a recommendation module, and a distribution module.
[0054] The fuzzing module is used to fuzz the task request from the task requester using K-anonymity technology to obtain a fuzzy request.
[0055] The aggregation module is used to aggregate fuzzy requests using user-based collaborative filtering technology to obtain group requests;
[0056] The recommendation module is used to perform task recommendation on group requests using the BERT2FTR model to obtain recommendation results;
[0057] The distribution module is used to distribute recommendation results to the corresponding task requesters using content-based filtering technology.
[0058] Compared with existing technologies, the advantages of this invention are as follows: It utilizes K-anonymity technology to fuzzify task requests and employs user-based collaborative filtering to aggregate similar requests into group requests for task recommendation, thus achieving dual privacy protection for task requesters. Furthermore, to reduce the computational overhead of the recommendation process, this invention uses an optimized lightweight BERT2FTR model to perform the fuzzy recommendation process, while simultaneously utilizing content-based filtering to distribute the group recommendation results to each user, thereby improving recommendation accuracy. Differential privacy noise is added during training to prevent the other party from inferring privacy from the recommendation results of this privacy model. Security analysis and experiments demonstrate that this scheme can improve task recommendation accuracy and efficiency while achieving fuzzy recommendation. Attached Figure Description
[0059] Figure 1 A basic flowchart of a spatial crowdsourcing fuzzy task recommendation method based on differential privacy provided in an embodiment of the present invention;
[0060] Figure 2 This is a graph showing the relationships between entities in a spatial crowdsourcing fuzzy task recommendation method based on differential privacy, provided in one embodiment of the present invention.
[0061] Figure 3 This is a flowchart of the task aggregation process for step S2;
[0062] Figure 4 The flowchart for distributing the results of step S4. Detailed Implementation
[0063] Specific embodiments of the present invention are given below. These specific embodiments are only used to further illustrate the present invention in detail and do not limit the scope of protection of the present invention.
[0064] Example 1, referring to Figure 1 and Figure 2 As an embodiment of the present invention, a spatial crowdsourcing fuzzy task recommendation method based on differential privacy is provided, comprising the following steps:
[0065] Step S1: Use K-anonymity technology to obfuscate the task request based on the task requester's task request to obtain an obfuscated request and submit it to the fog node;
[0066] Step S2: The fog node receives fuzzy requests, aggregates the fuzzy requests using user-based collaborative filtering technology to obtain group requests, and sends the group requests to the cloud server.
[0067] Step S3: The cloud server uses the BERT2FTR model to recommend tasks to the group requests, obtains the recommendation results, and sends them to the fog nodes.
[0068] In step S4, the fog node distributes the recommendation results to the corresponding task requesters using content-based filtering technology.
[0069] This invention utilizes K-anonymity to obfuscate task requests and employs user-based collaborative filtering to aggregate similar requests into group requests for task recommendation, thus achieving dual privacy protection for task requesters. Furthermore, to reduce the computational overhead of the recommendation process, this paper uses an optimized lightweight BERT2FTR model to perform the fuzzy recommendation process, while employing content-based filtering to distribute the group recommendation results to each user, thereby improving recommendation accuracy. Differential privacy noise is added during training to prevent other users from inferring privacy from the recommendation results of this privacy model. Security analysis and experiments demonstrate that this scheme can improve both the accuracy and efficiency of task recommendation while achieving fuzzy recommendation.
[0070] In a specific embodiment, step S1 specifically includes:
[0071] Step S11: Obtain the task request through the task requester, and extract the sensitive information fields in the task request to obtain the quasi-identifier;
[0072] Specifically, extracting sensitive information fields from task requests to obtain quasi-identifiers includes:
[0073] The task request is parsed to identify and separate various fields. Then, based on the preset quasi-identifier definition rules, sensitive information fields are filtered out from these fields. Sensitive information fields include location, detailed timestamp, age, gender, occupation, postal code, and combinations of contextual attributes strongly related to user identity. Finally, the filtered fields and their original value sets are output, which is the quasi-identifier set corresponding to the task request.
[0074] Step S12: Generalize the quasi-identifiers, group the generalized quasi-identifiers and process them using K-anonymity technology to obtain fuzzy requests.
[0075] Specifically, generalization processing includes:
[0076] Based on a predefined generalization hierarchy, the specific values of quasi-identifiers are replaced with more generalized values to reduce the precision of the data.
[0077] The specific processing using K-anonymity technology includes:
[0078] Group the generalized quasi-identifiers to ensure that each group contains at least one... There are 10 records, and these records have exactly the same value for the quasi-identifier. In this way, any single record cannot have a quasi-identifier that is identical to fewer than 10 records. Distinguish it from other records.
[0079] Extract the quasi-identifier attribute from the task request These quasi-identifier attributes can identify the attribute information of the task requester, and are used to distinguish task requesters. Then, regarding the quasi-identifier attributes... Generalization processing yields Then according to Task requests are grouped and processed using K-anonymity techniques to ensure that each group contains at least [number missing] task requests. Each request is initially indistinguishable; after K-anonymization, each task request is transformed into a vague request by removing specific personal sensitive information, ultimately retaining only generalized data similar to other requests. To ensure rigorous privacy, the system needs to verify that each group conforms to K-anonymity.
[0080] In a specific embodiment, step S2 specifically includes:
[0081] Step S21: Use the BERT2FTR model to extract the feature values of fuzzy requests and transform each task request into a fuzzy request feature vector.
[0082] The BERT2FTR model is obtained by optimizing the BERT model.
[0083] Extracting feature values for fuzzy requests using the BERT2FTR model specifically includes:
[0084] Each fuzzy request text is segmented and specially marked, then converted into an input tensor and fed into the BERT2FTR model. The BERT2FTR model performs bidirectional context encoding on the input through a multi-layer Transformer encoder, ultimately extracting... The hidden state vector of the marked location is used as the fuzzy request feature vector of the fuzzy request.
[0085] Step S22: Using user-based collaborative filtering technology, calculate the cosine similarity of the fuzzy request feature vectors. The calculation expression is as follows:
[0086] ;
[0087] in, and The fuzzy request feature vector of the task requester;
[0088] Specifically, for the fuzzy requests obtained in step S21, the fog node will use memory-based collaborative filtering technology to calculate the similarity of the task requests, and finally aggregate the task requests that meet the threshold to obtain the group requests.
[0089] It should be noted that memory-based collaborative filtering technology mainly includes user-based collaborative filtering technology and item-based collaborative filtering technology. In this invention, each task requester submits a task request, and the request center calculates the similarity between task requests; therefore, this invention uses user-based collaborative filtering technology.
[0090] User-based collaborative filtering treats each task requester as a user and their published fuzzy request feature vector as that user's historical behavioral preferences. By calculating the cosine similarity between the feature vectors of different requesters, the nearest neighbors of each requester are identified. Then, based on the request patterns of these nearest neighbors, the potential needs of the current requester are assessed, or the results are used for subsequent group recommendations.
[0091] Step S23: Calculate the cosine similarity of all fuzzy request feature vectors and obtain the similarity matrix;
[0092] Step S24: Use the K-means clustering algorithm to aggregate fuzzy requests with similarity higher than the similarity threshold into a group of requests;
[0093] If any fuzzy request has a similarity to all other fuzzy requests that is less than the similarity threshold, then that fuzzy request is treated as a group request.
[0094] Specifically, a similarity threshold of 0.7 can identify and aggregate requests that have overlapping core keywords but slightly different expressions, thereby filtering out redundant differences in expression while retaining the core demand characteristics.
[0095] Using the K-means clustering algorithm to aggregate fuzzy requests with similarity higher than a similarity threshold into a group of requests specifically includes:
[0096] Based on the similarity threshold and the minimum number of samples required to form a cluster, requests with similarity higher than the similarity threshold among all fuzzy requests are aggregated into different clusters, and requests that do not meet the conditions are marked as noise. For each generated cluster, the non-sensitive task content of all fuzzy requests within the cluster is summarized and extracted, and its quasi-identifier features are further generalized to a unified description that can completely cover all members within the cluster, generating a group request that represents the group and has a unified task description and generalized feature identifier.
[0097] Reference Figure 3 Regarding task requests from four task requesters , , and Calculate based on their similarity , , If the similarity between them is greater than a threshold, we aggregate them into a group of task requests. , The similarity between this request and the other three task requests is less than a threshold, so it is treated as a separate group request. Finally, the group request... and Send to the cloud server.
[0098] In a specific embodiment, step S3 specifically includes:
[0099] Fog nodes send group requests to the cloud server and use the BERT2FTR model for task recommendation. The BERT2FTR model processing includes an input layer, an embedding layer, a Transformer layer, and a classification output layer.
[0100] The specific processing of the input layer, embedding layer, Transformer layer, and classification output layer is as follows:
[0101] The input layer processing logic specifically includes:
[0102] The WordPiece word segmenter is used to segment the group request to obtain a sequence of sub-word units. The sub-word unit sequence is then converted to lowercase and filtered for irrelevant symbols.
[0103] Specifically, the lowercase conversion and irrelevant symbol filtering processes include:
[0104] Traverse each character in the sub-word unit sequence, convert all uppercase letters to their corresponding lowercase letters according to the ASCII encoding mapping relationship, and identify redundant punctuation, illegal characters and special format symbols in the sub-word unit sequence based on a preset noise symbol library;
[0105] During the filtering process, specific terms or symbol combinations representing fuzzy expressions in the sub-word unit sequence are identified and retained by matching regular expressions, non-key symbols are removed, and the processed sub-word units are recombined according to the original temporal order to generate a standardized word sequence for subsequent vectorized representation.
[0106] A pre-defined special marker is inserted into the processed sub-word unit sequence to obtain a standardized marker sequence;
[0107] Specifically, the preset special marker is as follows: and . Inserted at the beginning of the sequence, In single-sentence input, the insertion point is at the end of the sequence. In sentence-pair input, the insertion point is at the end of the first sentence and at the end of the entire sequence.
[0108] An initial tensor is generated based on a canonical tag sequence. The initial tensor includes a word tensor, a sentence vector, and an attention mask tensor.
[0109] Identify the effective word positions and filling positions in the canonical tag sequence, set the corresponding effective word position value in the attention mask tensor to 1, and set the corresponding filling position value to 0;
[0110] It should be added that padding will be performed for the padding position if the input length is insufficient.
[0111] The generated lexical tensor and attention mask tensor are used as the original input token sequence.
[0112] In a specific embodiment, the Embedding layer processing logic specifically includes:
[0113] Obtain the original input label sequence and process it using a tail truncation strategy. The tail truncation strategy processing logic includes:
[0114] If the length of the original input label sequence exceeds the preset maximum length Then only the last one of the original input tokens is retained. 1 tag, as a sequence of tags to be processed;
[0115] Specifically, key information for task classification often appears at the end of the user request. Although some preceding information may be lost, experiments have shown that using a tail truncation strategy to retain tail labels can improve model performance.
[0116] Each token in the sequence of tokens to be processed is converted into a word embedding vector;
[0117] Assign a learnable position embedding vector to each position in the sequence of tags to be processed, and add the word embedding vector to the learnable position embedding vector to obtain the input representation;
[0118] Obtain all input representations to get the input representation sequence.
[0119] Specifically, for a given tag, we construct its input representation by adding its word embeddings to its learnable positional embeddings:
[0120] ;
[0121] in, for Uyghur word embedding, for Dimensional positional embeddings. Using learnable positional embeddings allows the model to automatically adjust during training to better adapt to the specific positional distribution patterns in the task-requested corpus.
[0122] In a specific embodiment, the Transformer layer processing logic specifically includes:
[0123] The input representations are passed sequentially. A bidirectional Transformer encoder is used for feature extraction, and each layer of the bidirectional Transformer encoder outputs the hidden representation of the current layer.
[0124] Within each Transformer encoder layer, bidirectional dependencies between any tokens in the input representation sequence are calculated using a multi-head self-attention mechanism.
[0125] The output of the multi-head self-attention mechanism is nonlinearly transformed through a feedforward network. Residual connections and layer normalization operations are then performed on the output of the nonlinear transformation to obtain the hidden representation of the current layer.
[0126] Specifically, the The architecture of the layered bidirectional Transformer encoder enables the model to capture complex bidirectional dependencies between arbitrary words in a request, which is crucial for understanding incomplete or ambiguous descriptions.
[0127] In a specific embodiment, the classification output layer processing logic specifically includes:
[0128] The mean pooling method is applied to all hidden representations of the current layer output by the Transformer layer to generate the overall sequence representation. The mathematical expression for mean pooling is:
[0129] ;
[0130] in, For the overall representation of the sequence, To preset the maximum length, For the first The hidden representation of the i-th positional word in the layer Transformer encoder;
[0131] Specifically, all output tokens of the final Transformer layer are averaged and pooled to generate a holistic representation of the sequence. This is compared to using only... This tagging method can aggregate full sequence information and provides a more robust generalized representation for group requests in practice.
[0132] The sequence is represented as a whole by the task-specific classification head, and mapped to the probability distribution of each task category through a linear transformation and a softmax function. The expression for calculating the probability distribution is as follows:
[0133] ;
[0134] in, This represents the probability distribution vector for the task category. A set of tags for classifying tasks. The learnable weight matrix for the classification layer. The learnable bias vector for the classification layer;
[0135] Specifically, the learnable weight matrix of the classification layer is as follows: The matrix, is the dimension of the overall sequence representation. The total number of task categories, responsible for processing the input. Linear mapping of dimensional feature space to A dimensional category space. Each column corresponds to a discriminant vector for a task category.
[0136] The learnable bias vector of the classification layer has a length of The vector is used to introduce an independent base offset for each task category to adjust the result of the linear transformation. Its purpose is to provide a baseline score for each category without considering input features.
[0137] The task category corresponding to the highest probability distribution vector is selected as the recommendation result.
[0138] In a specific embodiment, step S4 specifically includes:
[0139] Step S41: Construct a request index for the task requester in the fog node. The request index includes the distance range, time span, and request type of the request.
[0140] Step S42: Store the recommendation results in the form of a report to obtain a recommendation report, which includes the name, location, type, open time and evaluation characteristics of the results;
[0141] Step S43: The content-based filtering technology calls the request index, iterates through the matching degree between the recommendation report and the request index, and distributes the recommendation reports with a matching degree greater than the threshold to the corresponding task requesters.
[0142] Specifically, content-based filtering is a recommendation technology that personalizes results distribution for task requesters. This method analyzes the specific characteristics of each task requester's request, focusing on key features that distinguish it from other task requests, in order to recommend results that match each task requester's specific needs and preferences.
[0143] Content-based filtering relies on a request index built into the system for task requesters. This index captures information crucial to the task requester, information that differs from other task requests. This includes key characteristics such as the request's distance range, time span, and type of need, which are then used to personalize result distribution.
[0144] Content-based filtering is based on how to represent recommendation results in the system. Each recommendation result is stored in the form of a report, and each report is a list that describes the result in detail. It lists the result's name, location, type, open time, rating, and other detailed characteristics.
[0145] By utilizing task requests stored in the index and recommendation results in reports, the content-based filtering module compares the features of each recommendation result to see if they meet the requirements of the task request, and distributes the matching results to the corresponding task requesters. In this personalized recommendation process, recommendations can be customized to meet the specific needs and preferences of each task requester.
[0146] Reference Figure 4 For the four recommendation results, content-based filtering technology performed content-driven recommendations based on their specific features. Taking the first two requesters as examples, we provided the task requester with... Recommended results and and to the task requester Recommended results and .
[0147] In a specific embodiment, differential privacy stochastic gradient descent is applied during the training of the BERT2FTR model to provide training data. - Differential privacy protection, and meets the same requirements - A differential privacy-preserving BERT2FTR model; where, To limit the loss of privacy, For privacy failure probability.
[0148] Specifically, obtain the BERT2FTR model to be trained and its training dataset. Set the target differential privacy budget parameters, including the upper limit of privacy loss. And privacy failure probability At the same time, determine the preset number of training epochs for the model and the base learning rate for the optimizer.
[0149] For each round of training, the training dataset is divided into multiple batches for iteration. For each batch, the following steps are performed:
[0150] Calculate the gradient of the loss with respect to the model parameters for each training sample in the current batch independently. Then, perform... Norm clipping: Calculate its If the norm is greater than a preset clipping threshold C, the gradient vector is scaled to a norm equal to the preset clipping threshold; otherwise, the original gradient is retained. This operation ensures that the influence of any single sample on the gradient update is capped at the preset clipping threshold.
[0151] The average gradient is obtained by averaging the gradients of all clipped samples in the current batch. Based on a preset noise multiplier σ, a random noise vector with the same dimension as the average gradient is generated. This noise vector has a mean of 0 and a covariance matrix of... The noise is sampled from a Gaussian distribution. This noise is added to the average gradient to obtain a noisy gradient estimate. Based on the noisy gradient estimate, the parameters of the BERT2FTR model are updated using a stochastic gradient descent optimizer. The above steps are repeated until all preset training rounds are completed.
[0152] Given target differential privacy budget parameters The system presets the number of training epochs, batch size, and training set size, and calculates the batch sampling probability. The required noise multiplier σ is derived using the formula from moment accounting. In practice, an open-source privacy-preserving machine learning library is used to automate this calculation and track privacy consumption.
[0153] During training, the privacy accounting module continuously monitors the accumulated privacy consumption. At the end of training, if the accumulated privacy consumption does not exceed the set target differential privacy budget parameter, it proves that the final generated BERT2FTR model meets the preset requirements. - Differential privacy protection.
[0154] This invention forms a dual defense system of K-anonymity technology and differential privacy protection. By applying differential privacy stochastic gradient descent to generalized data, the intensity of the noise to be injected can be reduced, thereby achieving higher recommendation accuracy at the same level of privacy.
[0155] Example 2, another embodiment of the present invention, provides a spatial crowdsourcing fuzzy task recommendation system based on differential privacy, characterized in that it includes a fuzzy processing module, an aggregation module, a recommendation module, and a distribution module:
[0156] The fuzzing module is used to fuzz the task request from the task requester using K-anonymity technology to obtain a fuzzy request;
[0157] The aggregation module is used to aggregate fuzzy requests using user-based collaborative filtering technology to obtain a group of requests;
[0158] The recommendation module is used to perform task recommendations on group requests using the BERT2FTR model and obtain recommendation results;
[0159] The distribution module is used to distribute recommendation results to the corresponding task requesters using content-based filtering technology.
[0160] This invention utilizes K-anonymity to obfuscate task requests and employs user-based collaborative filtering to aggregate similar requests into group requests for task recommendation, thus achieving dual privacy protection for task requesters. Furthermore, to reduce the computational overhead of the recommendation process, this paper uses an optimized lightweight BERT2FTR model to perform the fuzzy recommendation process, while employing content-based filtering to distribute the group recommendation results to each user, thereby improving recommendation accuracy. Differential privacy noise is added during training to prevent other users from inferring privacy from the recommendation results of this privacy model. Security analysis and experiments demonstrate that this scheme can improve both the accuracy and efficiency of task recommendation while achieving fuzzy recommendation.
[0161] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0162] Any aspects not covered in this invention are applicable to existing technologies.
Claims
1. A spatial crowdsourcing fuzzy task recommendation method based on differential privacy, characterized in that, Includes the following steps: Step S1: Use K-anonymity technology to obfuscate the task request based on the task requester's task request to obtain an obfuscated request and submit it to the fog node; Step S2: The fog node receives fuzzy requests, aggregates the fuzzy requests using user-based collaborative filtering technology to obtain group requests, and sends the group requests to the cloud server. Step S3: The cloud server uses the BERT2FTR model to recommend tasks to the group requests, obtains the recommendation results, and sends them to the fog nodes. The BERT2FTR model consists of an input layer, an embedding layer, a Transformer layer, and a classification output layer. The classification output layer processing logic specifically includes: The mean pooling method is applied to all hidden representations of the current layer output by the Transformer layer to generate the overall sequence representation. The mathematical expression for mean pooling is: ； in, For the overall representation of the sequence, To preset the maximum length, For the first The hidden representation of the i-th positional word in the layer Transformer encoder; The sequence is represented as a whole by the task-specific classification head, and mapped to the probability distribution of each task category through a linear transformation and a softmax function. The expression for calculating the probability distribution is as follows: ； in, This represents the probability distribution vector for the task category. A set of tags for classifying tasks. The learnable weight matrix for the classification layer. For the learnable bias vector of the classification layer; The task category corresponding to the highest probability distribution vector is selected as the recommendation result; In step S4, the fog node distributes the recommendation results to the corresponding task requesters using content-based filtering technology.
2. The spatial crowdsourcing fuzzy task recommendation method based on differential privacy as described in claim 1, characterized in that, Step S1 specifically includes: Step S11: Obtain the task request through the task requester, and extract the sensitive information fields in the task request to obtain the quasi-identifier; Step S12: Generalize the quasi-identifiers, group the generalized quasi-identifiers and process them using K-anonymity technology to obtain fuzzy requests.
3. The spatial crowdsourcing fuzzy task recommendation method based on differential privacy as described in claim 2, characterized in that, Step S2 specifically includes: Step S21: Use the BERT2FTR model to extract the feature values of fuzzy requests and transform each task request into a fuzzy request feature vector. Step S22: Using user-based collaborative filtering technology, calculate the cosine similarity of the fuzzy request feature vectors. The calculation expression is as follows: ； in, and The fuzzy request feature vector of the task requester; Step S23: Calculate the cosine similarity of all fuzzy request feature vectors and obtain the similarity matrix; Step S24: Use the K-means clustering algorithm to aggregate fuzzy requests with similarity higher than the similarity threshold into a group of requests; If any fuzzy request has a similarity to all other fuzzy requests that is less than the similarity threshold, then that fuzzy request is treated as a group request.
4. The spatial crowdsourcing fuzzy task recommendation method based on differential privacy as described in claim 3, characterized in that, Step S3 specifically includes: Fog nodes send group requests to the cloud server, which then uses the BERT2FTR model for task recommendation. The input layer processing logic specifically includes: The group requests are segmented using a preset word segmenter to obtain a sequence of sub-word units. The sub-word unit sequence is then converted to lowercase and filtered for irrelevant symbols. A pre-defined special marker is inserted into the processed sub-word unit sequence to obtain a standardized marker sequence; An initial tensor is generated based on a canonical tag sequence. The initial tensor includes a word tensor, a sentence vector, and an attention mask tensor. Identify the effective word positions and filling positions in the canonical tag sequence, set the corresponding effective word position value in the attention mask tensor to 1, and set the corresponding filling position value to 0; The generated lexical tensor and attention mask tensor are used as the original input token sequence.
5. The spatial crowdsourcing fuzzy task recommendation method based on differential privacy as described in claim 4, characterized in that, The specific processing logic of the Embedding layer includes: Obtain the original input label sequence and process it using a tail truncation strategy. The tail truncation strategy processing logic includes: If the length of the original input label sequence exceeds the preset maximum length Then only the last one of the original input tokens is retained. 1 tag, as a sequence of tags to be processed; Each token in the sequence of tokens to be processed is converted into a word embedding vector; Assign a learnable position embedding vector to each position in the sequence of tags to be processed, and add the word embedding vector to the learnable position embedding vector to obtain the input representation; Obtain all input representations to get the input representation sequence.
6. The spatial crowdsourcing fuzzy task recommendation method based on differential privacy as described in claim 5, characterized in that, The Transformer layer processing logic specifically includes: The input representations are passed sequentially. A bidirectional Transformer encoder is used for feature extraction, and each layer of the bidirectional Transformer encoder outputs the hidden representation of the current layer. Within each Transformer encoder layer, bidirectional dependencies between any tokens in the input representation sequence are calculated using a multi-head self-attention mechanism. The output of the multi-head self-attention mechanism is nonlinearly transformed through a feedforward network. Residual connections and layer normalization operations are then performed on the output of the nonlinear transformation to obtain the hidden representation of the current layer.
7. The spatial crowdsourcing fuzzy task recommendation method based on differential privacy as described in claim 6, characterized in that, Step S4 specifically includes: Step S41: Construct a request index for the task requester in the fog node. The request index includes the distance range, time span, and request type of the request. Step S42: Store the recommendation results in the form of a report to obtain a recommendation report, which includes the name, location, type, open time and evaluation characteristics of the results; Step S43: The content-based filtering technology calls the request index, iterates through the matching degree between the recommendation report and the request index, and distributes the recommendation reports with a matching degree greater than the threshold to the corresponding task requesters.
8. The spatial crowdsourcing fuzzy task recommendation method based on differential privacy as described in claim 7, characterized in that, Differential privacy stochastic gradient descent is applied during the training of the BERT2FTR model to obtain models that satisfy the same... - A differential privacy-preserving BERT2FTR model; where, To limit the loss of privacy, For privacy failure probability.
9. A spatial crowdsourcing fuzzy task recommendation system based on differential privacy, applied to the spatial crowdsourcing fuzzy task recommendation method based on differential privacy as described in any one of claims 1-8, characterized in that, It includes a fuzzy processing module, an aggregation module, a recommendation module, and a distribution module: The fuzzing module is used to fuzz the task request from the task requester using K-anonymity technology to obtain a fuzzy request. The aggregation module is used to aggregate fuzzy requests using user-based collaborative filtering technology to obtain group requests; The recommendation module is used to perform task recommendation on group requests using the BERT2FTR model to obtain recommendation results; The distribution module is used to distribute recommendation results to the corresponding task requesters using content-based filtering technology.