Recommendation system implementation method, device and equipment based on cloud disk privacy protection

Abstract

The invention discloses a recommendation system implementation method, device and equipment based on cloud disk privacy protection. The method comprises the steps that edge nodes are additionally arranged on a client side and a server side of a trusted cloud disk recommendation system; calling a multi-modal visual language model of the server to locally extract standardized file data features, completing generation of user file text description information, and performing desensitization processing on file data; the server modeling adopts a fully homomorphic encryption algorithm to encrypt user data, and user files are subjected to file hash verification; the client adopts an event-driven architecture, monitors and responds to user behaviors in real time, captures user dynamic information in real time through context sensing analysis, and updates user data and cache data of a user interest model; a Transform deep learning architecture is adopted, the relevance of entity user data is analyzed through a self-attention mechanism and a knowledge graph, high-weight core file features are searched, and a client performs personalized content dynamic recommendation based on user real-time context perception.

Description

Technical Field

[0001] This application generally relates to the field of cloud storage technology, and in particular to a method, apparatus and device for implementing a recommendation system based on cloud disk privacy protection. Background Technology

[0002] With the widespread adoption of cloud computing and mobile devices, personal cloud storage services have become a crucial platform for users to store and manage their data. Users accumulate vast amounts of personal information and files in cloud storage, data that contains rich insights into user preferences and behavioral patterns. Personalized content recommendation systems can analyze these cloud storage files to gain a deeper understanding of user interests and needs, thereby providing more accurate and personalized content recommendations.

[0003] Most recommendation systems on the market today are content customization based on user group tags, i.e., B2C functional query mode, such as precision marketing based on user content preferences, prediction based on user behavior habits, and collaborative filtering. Figure 1 This is a common recommendation system service model. The most significant characteristic of this type of system is that user data is completely exposed within the business system. The recommendation system performs detailed analysis of user data to ultimately create user tags and, combined with data modeling, formulates various operational strategies at the application layer. However, with users increasingly valuing data privacy, how to rationally utilize cloud storage data for recommendations without infringing on user privacy has become a key challenge in building recommendation systems. Furthermore, this type of tag- and data model-based recommendation system is highly dependent on sample data, especially positive samples, which can even directly affect the model's accuracy.

[0004] Typical recommendation systems generally divide their architecture into two parts: an offline system and an online system. The offline part is usually responsible for data governance and content modeling. For example, after the content ecosystem data is collected, the original data and the introduced collaborative data are integrated, managed, and consumed in a unified manner. The offline system performs content processing on this data, such as tagging, tag feature extraction, and vectorization, and then performs data modeling to ultimately form a user model. The online part is responsible for generating recommendation content, and usually uses various recall algorithms and result ranking algorithms. Common recall algorithms include content-based recall, collaborative filtering recall, and model-based recall; while common result ranking algorithms include logistic regression, gradient boosting decision trees, and neural network ranking.

[0005] Based on this system architecture, the functional architecture of existing cloud disk data-driven personalized recommendation systems (such as...) Figure 2The system is divided into a file service system cluster and a recommendation system cluster. The drawbacks of this B2C query system are obvious: user data stored in the file service system is completely exposed to the server. The cloud storage server can perform any operation on user data, including but not limited to accessing, modifying, copying, or deleting user files. This design not only raises user concerns about privacy but also brings the following problems: 1. Data security issues: Once the server is hacked or a data breach occurs, user data may be obtained by unauthorized third parties. 2. Privacy leakage risk: The server can easily access user data, which may lead to the improper use or leakage of users' personal privacy and sensitive information. 3. Decreased trust: Users may lose trust in the service provider due to concerns about their data security, which will affect the service's reputation and user stickiness. 4. Legal liability and compliance challenges: In many countries and regions, data protection regulations require companies to protect users' personal data. The server's complete access to data may expose the service provider to legal liability and compliance issues. 5. Data misuse issues: The server may misuse user data for improper business practices, such as data analysis and advertising without user consent. 6. Increased dependence on technology: Users become more dependent on the server. If the server encounters problems, users' data may be affected, resulting in data loss or unavailability.

[0006] Furthermore, while this batch query system based on user tags and recommendation algorithms relies on a user tag library to provide highly accurate personalized services, it also has some significant limitations. First, because the system depends on an offline tag library, the accuracy of its data is limited by the update frequency of the tag library, resulting in a time lag in the system's response to changes in user behavior. Second, frequent updates to the entire user tag library can impose significant system overhead on the business system. Such large-scale update operations, while intended to maintain data freshness and accuracy, can actually lead to unnecessary resource waste. Especially considering the business characteristics of cloud storage applications, many user-stored data files do not frequently undergo large-scale changes. Therefore, this full-update strategy is not very effective in many cases and may even become a bottleneck for system performance due to the large amount of computing resources consumed. Therefore, this application aims to develop a recommendation system that can effectively mine the value of user data while fully protecting user privacy. Summary of the Invention

[0007] This application addresses the problems and shortcomings of existing recommendation systems by providing a method, apparatus, and device for implementing a recommendation system based on cloud storage privacy protection. It achieves high-accuracy model prediction optimization and improvement by utilizing as little low-labeled or unlabeled data as possible. This effectively mines the value of user data while fully protecting user privacy, enhancing user experience and strengthening user trust.

[0008] In a first aspect, embodiments of this application provide a method for implementing a recommendation system based on cloud disk privacy protection, comprising: a recommendation system service model, including: Add an edge node between the client and server of the trusted cloud disk recommendation system; The server-side multimodal visual language model is called to extract standardized file data features locally. The multimodal visual language model first processes the unencrypted original file on the local client to generate user file text descriptions. Then, the user file is parsed into text features locally through edge computing to generate user file text description data. Finally, the file feature data uploaded to the server is desensitized. The server-side modeling uses a fully homomorphic encryption algorithm to encrypt the user file text description data. The encrypted text description data is uploaded to the server for modeling and recommendation. The file hash encryption algorithm is used to perform file hash verification on the user's original file to generate a hash verification value. The local client uses an event-driven architecture to capture file change hash verification, monitor user behavior in real time, and trigger the trusted cloud disk recommendation system to respond in real time. The client also analyzes real-time user behavior dynamics based on user context awareness and updates the encrypted user feature data and cached data of the user interest model. The multimodal visual language model adopts the Transformer deep learning architecture. It uses a self-attention mechanism combined with knowledge graph analysis to identify the self-relationships of entity user data, finds high-weight core file features, and combines the resulting recommended content list with machine learning recommendation algorithms to complete personalized dynamic content recommendation ranking and push.

[0009] In a second aspect, embodiments of this application provide an apparatus for implementing a recommendation system based on cloud disk privacy protection, comprising: The adjustment module is used to add edge nodes between the client and server of the trusted cloud disk recommendation system; The extraction and desensitization module is used to call the server's multimodal visual language model to extract standardized file data features locally. The multimodal visual language model first processes the unencrypted original file on the local client to generate user file text descriptions. Then, it uses edge computing to parse the user file into text features locally, generates user file text description data, and desensitizes the file feature data uploaded to the server. The encryption and verification module is used for server-side modeling. It uses a fully homomorphic encryption algorithm to encrypt the text description of user files. The encrypted text description data is uploaded to the server for modeling and recommendation. It also uses a file hash encryption algorithm to perform file hash verification on the original user files to generate a hash verification value. The update module is used by the local client to capture file change hash verification using an event-driven architecture, monitor user behavior in real time and trigger real-time responses from the trusted cloud disk recommendation system, and analyze real-time dynamic information of user behavior based on user context awareness, and update user encrypted feature data and cached data of the user interest model. The personalized recommendation module uses a multimodal visual language model with a Transformer deep learning architecture. It analyzes the self-attention mechanism and knowledge graph to identify the self-relationships of entity user data, finds high-weight core file features, and combines the resulting list of recommended content with machine learning recommendation algorithms to complete personalized dynamic recommendation, sorting, and push.

[0010] Thirdly, embodiments of the present invention provide an electronic device, which includes: one or more processors; a memory for storing one or more programs; and when one or more programs are executed by one or more processors, the one or more processors implement the recommendation system implementation method based on cloud disk privacy protection provided in any embodiment of the present invention.

[0011] Fourthly, this application discloses a computer-readable storage medium for storing a computer program; when the computer program is executed by a processor, it implements the steps of the aforementioned disclosed method for implementing a recommendation system based on cloud disk privacy protection.

[0012] Fifthly, embodiments of this disclosure provide a computer program product, including at least one of a computer program and instructions, wherein when the computer program and at least one of the instructions are executed by an electronic device, they implement the steps of the recommendation system implementation method based on cloud disk privacy protection described in the first aspect.

[0013] Compared with the prior art, the present invention achieves the following effects: This application employs fully homomorphic encryption and file hash verification to enhance the security of user privacy data, ensuring its safety and privacy during storage, processing, and transmission. Only then can users confidently utilize cloud storage services and personalized recommendation systems.

[0014] In order to effectively protect users' privacy data from leakage, the edge gateway assists users in extracting local data features and de-identifies the data uploaded to the server, thereby reducing the exposure of the original data and protecting user data security.

[0015] This paper employs a multimodal visual language model based on the Transformer architecture to solve the sequence labeling problem. Through a self-attention mechanism, user media files are converted into textual description data. Then, based on the self-association analysis of knowledge graph entity user data, high-weight core features are identified to obtain a list of recommended content. Finally, machine learning recommendation algorithms are combined to complete the ranking and push of recommended content. Finally, A / B testing is used to collect performance evaluation data for fine-tuning of model parameters and freezing of effective parameters to achieve continuous model improvement.

[0016] The recommendation system employs an event-driven architecture and context-aware analysis to capture real-time user dynamic information. It can respond in real-time to changes in user behavior and dynamically update recommended content to provide recommendations that are more closely aligned with the user's current interests.

[0017] The overall architecture of the recommendation system was adjusted to make resource utilization more rational and better suited to business scenarios with small-scale, high-frequency data updates. An edge gateway layer was added between the client and server to handle tasks such as extracting text description information from media files uploaded by cloud disk clients. This decomposes the modeling process, allowing personal data preprocessing to be completed locally on the user terminal, improving resource utilization and reducing unnecessary data computation on the server. It effectively separates high-frequency and low-frequency tasks, improving system efficiency and scalability. Attached Figure Description

[0018] Figure 1 A schematic diagram of the service model for an existing recommendation system; Figure 2 Functional architecture diagram of a cloud disk data-driven personalized recommendation system; Figure 3 This is a flowchart illustrating the cloud disk data-driven personalized recommendation method according to an embodiment of this application. Figure 4 This is a schematic diagram of the functional architecture of the cloud disk data-driven personalized recommendation method in an embodiment of this application; Figure 5 This is a diagram illustrating the multimodal visual language model used in the cloud disk data-driven personalized recommendation method of this application embodiment. Figure 6 This is a diagram illustrating the iterative compression process of the cloud disk data-driven personalized recommendation method in an embodiment of this application. Figure 7 This is a general knowledge graph example diagram of the cloud disk data-driven personalized recommendation method in the embodiments of this application; Figure 8 This is a classification flowchart of the cloud disk data-driven personalized recommendation method according to an embodiment of this application; Figure 9 This is an entity relationship completion diagram for the cloud disk data-driven personalized recommendation method in the embodiments of this application; Figure 10This is a schematic diagram of the cloud disk data-driven personalized recommendation device according to an embodiment of this application; Figure 11 This is a structural diagram of a cloud disk data-driven personalized recommendation device according to an embodiment of this application. Detailed Implementation

[0019] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and not intended to limit the scope of the application. Furthermore, it should be noted that, for ease of description, only the parts relevant to the present application are shown in the accompanying drawings, not the entire structure.

[0020] based on Figure 3 This application describes an implementation method for a recommendation system based on cloud disk privacy protection, comprising: S1. Add an edge node between the client and server of the trusted cloud disk recommendation system; S2. The server-side multimodal visual language model, which is transferred by the edge node module, extracts standardized file data features locally. The multimodal visual language model first processes the unencrypted original file on the local client to generate user file text descriptions. Then, it parses the user file into text features locally through edge computing, generates user file text description data, and performs desensitization processing on the file feature data uploaded to the server. S3. The server-side modeling uses a fully homomorphic encryption algorithm to encrypt the user file text description data. The encrypted text description data is uploaded to the server for modeling and recommendation. The original user file is hashed using a file hash encryption algorithm to generate a hash verification value. This ensures that the data remains in an encrypted state from the moment it leaves the user's device.

[0021] S4. The local client uses an event-driven architecture to capture file change hash verification, monitor user behavior in real time and trigger the trusted cloud disk recommendation system to respond in real time. The client also analyzes real-time user behavior dynamics based on user context awareness and updates the encrypted user feature data and cached data of the user interest model. S5, the multimodal visual language model adopts the Transformer deep learning architecture. It uses a self-attention mechanism combined with knowledge graph analysis to identify the self-relationships of entity user data, find high-weight core file features, and combine the resulting recommended content list with machine learning recommendation algorithms to complete personalized dynamic content recommendation ranking and push.

[0022] This recommendation system is adaptable to the needs of different user groups and business scenarios, exhibiting excellent flexibility and scalability. By analyzing users' cloud storage data, it provides accurate and personalized recommendation services. This not only improves the user experience but also brings higher user engagement and satisfaction to enterprises, potentially increasing user stickiness and conversion rates.

[0023] In one embodiment of the first aspect, edge computing technology significantly improves response efficiency by shifting data processing and analysis tasks from the central server to the edge of the network, i.e., closer to the data source. Simultaneously, through the unified allocation of edge nodes via edge gateways, workloads can be distributed across multiple edge nodes, achieving load balancing. Furthermore, due to the modular nature of edge computing, edge nodes can be flexibly added or removed as needed, enabling a scalable architecture to adapt to constantly changing business requirements and system loads.

[0024] In one embodiment of the first aspect, the data used for server-side modeling consists of anonymized feature data. The original data and sensitive data are all stored with fully homomorphic encryption. User file text information is generated locally by calling the model. Subsequent data modeling uses this standardized file feature data. This reduces the exposure of the original data and protects user data security. Subsequent data modeling is for a recommendation system.

[0025] In one embodiment of the first aspect, the fully homomorphic encryption algorithm sets three ciphertext types. Specifically, the fully homomorphic encryption algorithm is a scheme that completes the data encryption and decryption process based on six functions.

[0026] In one embodiment of the first aspect, the user's original file undergoes file hash verification to generate a hash digest verification value. The generated hash digest information is then compared to determine whether the user file has been modified. This enables data modeling and personalized content recommendation of user data under a fully encrypted state. Real-time monitoring of user dynamic changes is also implemented.

[0027] In one embodiment of the first aspect, the user data cache and the update data cache are data caches provided by edge nodes, and server-side parameters are updated in batches when incremental changes in user data cause changes in recommended content. In another embodiment of the first aspect, the hash function is based on its resistance to second preimages: that is, given an input, it produces the same hash value.

[0028] In one embodiment of the first aspect, the file hash encryption algorithm uses the national standard SIM3 algorithm. The SIM3 algorithm adopts an MD structure, and the input message (length L < 2) 64 After padding, expansion, and iterative compression, a hash value of 256 bits is generated. The SM3 algorithm includes padding blocks, message expansion, and iterative compression.

[0029] In one embodiment of the first aspect, an event-driven architecture is employed to monitor and respond to user behavior in real time, rapidly updating the user interest model. The timing of user behavior is captured, triggering a recommendation system response, while context awareness provides environmental information at the time of the behavior, enabling dynamic recommendations. The event-driven architecture is responsible for capturing file change hash events and user behavior events in real time, triggering subsequent encrypted feature updates and cloud-based multimodal visual language model recalculation processes.

[0030] like Figure 4 This paper describes a comprehensive architectural adjustment to the trusted cloud storage recommendation system using an edge computing (EC) model architecture. User feature engineering algorithms and recommendation model updates are performed locally on the cloud storage client. An edge gateway layer is added between the client and server to extract text description information from media files uploaded by cloud storage clients. This decomposes the modeling process, with personal data preprocessing completed locally on the user's terminal; the client and edge nodes handle the data preprocessing off-site. This reduces data transmission requirements and unnecessary server-side computation, lowers bandwidth usage and latency, while enhancing data privacy and improving resource utilization, response speed, and scalability. The user feature engineering algorithm is based on federated learning.

[0031] The cloud disk server performs pre-training of visual language models, while edge nodes synchronously update recommendation model algorithms and homomorphic encryption algorithms; for example... Figure 5 The user generates their file text information locally by calling a multimodal visual language model. Subsequent data modeling can then utilize this standardized file feature data. The cloud storage server generates customized content by transmitting the de-identified file information and encrypted sensitive information from the uploaded encrypted file.

[0032] Based on the updated cached data provided by the edge nodes, the server parameters are only updated in batches when incremental user data causes changes in the recommended content, thus reducing the frequency and amount of data transmission.

[0033] In one embodiment of the first aspect, edge nodes are deployed as data preprocessing nodes and data caching nodes to both the client and edge nodes to perform tasks such as uploading user feature text description data required for modeling the trusted cloud disk recommendation system and incrementally uploading data by comparing cached data, thereby reducing data transmission. This relies on edge nodes for user data caching in customized content delivery.

[0034] In one embodiment of the first aspect, such as Figure 6The user extracts stylistic features locally by calling the server-side visual language model transferred from the edge node. The data used for server-side modeling is the de-identified feature data. Both the original data and the sensitive data are stored in homomorphic encryption, which realizes the full-process protection of the user's sensitive data, reduces the risk of data being intercepted during transmission, and enhances the privacy protection of the data.

[0035] In one embodiment of the first aspect, edge computing via an edge gateway significantly improves response speed and efficiency by shifting data processing (including data preprocessing) and analysis tasks from the central server to the edge of the network, i.e., closer to the data source. This distributed architecture shortens data transmission distances and reduces data transmission latency, enabling applications to respond to user requests more quickly. Simultaneously, through the unified allocation of edge nodes by the edge gateway, workloads can be distributed across multiple edge nodes, achieving load balancing. Furthermore, due to the modular nature of edge computing, edge nodes can be flexibly added or removed as needed, enabling a scalable architecture to adapt to constantly changing business requirements and system loads.

[0036] In one embodiment of the first aspect, a toroidal fully homomorphic encryption algorithm is used to encrypt user data. Simultaneously, the trusted cloud storage recommendation system performs file hash verification on user files, generating hash digest verification values. These hash digests are compared to quickly determine if the user files have been modified, thus preventing tampering. In cloud processing, this significantly reduces the risk of improper access or misuse of data on the server side, avoiding the risk of data and key leakage caused by frequent encryption and decryption of data files. The cloud storage data-driven personalized recommendation system uses fully homomorphic encryption technology to achieve data modeling and personalized content recommendation of cloud storage user data under a fully encrypted state. This ensures the security and privacy protection of user data, enhancing user data privacy protection. Users can confidently use cloud storage services and the personalized recommendation system.

[0037] In one embodiment of the first aspect, the user data is encrypted using a computationally efficient Fully Homomorphic Encryption over the Torus (TFHE) algorithm. Bootstrap binary gates are provided to implement the functions required by the developer. The TFHE algorithm is based on the torus mathematical structure, denoting the real torus used as T = Rmod1, and the N-dimensional polynomial with coefficients being elements on the torus as T. N [X] After each logic gate operation, a bootstrapping technique is used to refresh the ciphertext to reduce noise. The TFHE algorithm sets three ciphertexts. TLWE Ciphertext: Plaintext space is T, Ciphertext space is T n+1 The key space is B nCiphertext c = (a, b) ∈ T n+1 , where b = s·a + μ + e. Here, s is the key, a is the mask (a random element on the torus), μ is the plaintext message, and e is noise.

[0038] TRLWE Ciphertext: Plaintext space is T N [X], the ciphertext space is T N [X] 2 The key space is B N [X], ciphertext c = (a, b) ∈ T N [X] 2 , where b = s·a + μ + e.

[0039] TRGSW Ciphertext: Plaintext space is Z[X] / (X) n+1 The ciphertext space is T N [X] 2 / ×2 The key space is B N [X], Ciphertext C=Z+m·G2∈T N [X] 2 / ×2 , where Z is a TRLWE ciphertext of length 2 / column with zeros, and G2 is a gadget matrix.

[0040] The TFHE algorithm uses functions to perform data encryption and decryption: TFHE.KeyGen(params): Generates the private key sk, public key pk, key conversion key KS, and bootstrapping key BK.

[0041] TFHE.SymEnc(μ): Randomly samples a noise e from a Gaussian distribution, and randomly samples n elements from the torus to form a random mask: a∈T n Returns a TLWE ciphertext c = (a, b), where b = sk·a + μ + e.

[0042] TFHE.SymDec(c, sk): Takes a TLWE ciphertext c and a key sk as input, and returns the message μ.

[0043] TFHE.KeySwitch(C1,…,Cp,f,KS i j): Input p TLWE ciphertexts c i ∈TLWE K (μ) i For i ∈ {1,2,…,p}, an R-Lipschitz mapping f and a key transformation key KS i ,j, output a TRLWE sample C∈TRLWE K (f(μ1,…,μp)).

[0044] TFHE.BlindRotate(c,a1,…,a p ,b,C1,…,C p ): Input a TRLWE ciphertext c, p+1 integer coefficients a1, ..., a p b∈Z / (2NZ) and p TRGSW samples C1,…,C p Returns a TRLWE ciphertext of X - ρ·v, where ρ = b - a i Modulo 2N.

[0045] TFHE.Bootstrapping(μ1,C): Input a constant μ1∈T, a TLWE ciphertext C= (χ· ,e), where χ∈B is a binary vector, e is the noise term and bootstrap key sk boot Returns a TLWE ciphertext C'=(χ·μ1,e') where e' is the noise term in the new ciphertext.

[0046] In the TFHE algorithm, the homomorphic evaluation of binary gates is achieved through logical operations on the TLWE ciphertext followed by a gate bootstrap function. By using a fully homomorphic encryption algorithm, all basic binary gates can be implemented using a single gate bootstrap function. The bootstrap function is as follows: TFHE.BootsNAND(c1,c2)=GateBootstrap((0,-1 / 8)+c1+c2); TFHE.BootsAND(c1,c2)=GateBootstrap((0,5 / 8)-c1-c2); TFHE.BootsOR(c1,c2)=GateBootstrap((0,1 / 8)+c1+c2); TFHE.BootsXOR(c1,c2)=GateBootstrap(2·(c1-c2)); TFHE.BootsNOT(c)=(0,1 / 4)-c.

[0047] File hash verification: To avoid the risk of data and key leakage caused by frequent encryption and decryption of data files, while users perform homomorphic encryption on their files, the trusted cloud storage recommendation system also generates a hash digest for the user's files. Subsequently, the system can quickly determine whether the user's files have been modified by comparing the digest information. (Based on the second preimage resistance of the hash function: that is, given input, it produces the same hash value).

[0048] In one embodiment of the first aspect, the national cryptographic algorithm SIM3 is used as the file hash encryption algorithm. The SIM3 algorithm adopts an MD structure, and the input message (length L < 2)64 After padding, expansion, and iterative compression, a hash value of length 256 bits is generated. The implementation process of the SM3 algorithm includes padding groups: padding groups are the process of adding bits of any length to the end of the input message according to certain rules to a multiple of 512 bits, and then dividing the padded input string into several groups of 512 bits each. For an input message of length L, first, add bits of "1" to the end of the message, then add k bits of "0", where k is the smallest non-negative integer that makes L+k+1=448mod 512, and then add a 64-bit binary string, which is the binary representation of length L. Then, divide the padded input message into n groups of 512 bits each: B0, B1, ..., B n-1 , where n = (L + k + 65) / 512.

[0049] Message expansion: Each 512-bit input message block is expanded before the iterative compression function, generating 132 message words W0, W1, ..., W 132 And are used as inputs to the compression function. Where W0, W1, ..., W 15 The input message group is the input message group, and the remaining extended message word expressions are: W j =P1(W j-16 ⊕W j-9 ⊕(W j-3 <<<15))⊕(W j-13 <<<7)⊕W j-16 ≤j≤67 W' j =W j ⊕W j+4 ,0≤j≤63 In the formula, <<< represents a circular left shift; P1 is the permutation function in message expansion, and its expression is: P1(X)=X⊕(X<<<15)⊕(X<<<23).

[0050] Iterative Process: Let the compression function CF. Each iteration's input consists of a 512-bit input message block and the 256-bit output from the previous iteration. The output length of each iteration is 256 bits. Executing the SM3 algorithm once requires n iterations, and the iteration process is V. i+1 =CF(B i V i ), 0≤i≤n-1, where the initial iteration value V0 is a specified 256-bit constant IV. After all n message input packets have undergone iterative compression, the resulting V... n It is the 256-bit hash value output by the SM3 hash algorithm.

[0051] Compression function: In the SM3 algorithm, a single-iteration compression process is defined, where A, B, C, D, E, F, G, and H are 32-bit variable registers, and SS1, SS2, TT1, and TT2 are intermediate variables. A single iteration involves 64 rounds of compression, as follows: Figure 6 The execution process of the function CF shown is as follows: ABCDEFGH=V i Forj=0 to 63 SS1=((A<<<12)+E+(T j <<<(j mod 32)))<<<7 SS2 = SS1 ⊕ (A << < 12) TT1 = FF j (A,B,C)+D+SS2+ W' j TT2 = GG j (E,F,G)+H+SS1+ W' j A = TT1, B = A, C = B << < 9, D = C E = P0(TT2), F = E, G = F << < 19, H = G EndFor V i+1 = ABCDEFGH⊕V i Among them, FF j GG j It is a Boolean function, T j Let P0 represent the constant for each round, and let P0 be the permutation function in the compression function. The expressions for these functions are as follows: FF j (X,Y,Z)=

[0052] GG j (X,Y,Z)=

[0053] P0(X)=X⊕(X<<<9)⊕(X<<<17) In the formula, ∧, ∨, and ┐ represent bitwise AND, OR, and NOT operations, respectively, and ⊕ represents bitwise XOR.

[0054] In one embodiment of the first aspect, an event-driven architecture is employed to respond to changes in user behavior in real time. Context-aware analysis captures dynamic user information in real time, dynamically updates recommended content based on the user's current interests, and combines this with user file hash verification (rapid comparison of summaries) to monitor dynamic changes in user behavior in real time. This allows for rapid updates to the user interest model, enabling more detailed and dynamic recommendation segmentation, and providing a faster, more timely response to changes in user behavior.

[0055] You can embed the event real-time listening and reporting SDK toolkit into the cloud drive client.

[0056] In one embodiment of the first aspect, the user's real-time context-aware information enables dynamic recommendations; the user context information is divided into two parts: user state context information and user file content context information; the user state context information is monitored and collected in real time through the client SDK, and the collected information includes user system information, user location information, user application information, user device information, and user device status information; User file content context information is monitored in real time using hash aggregation. The operation steps are as follows: When a user uploads a single file to the cloud storage system, a hash value K is generated for that single file. n ; Let a set C be defined. n Used to store user file hash values; The set is encrypted using the AES symmetric encryption algorithm to generate the summary ciphertext EM. h ; Summary of EM at different times h Compare the files; if they match, it means the user files have not been changed. For example, summaries of different time periods EM h If there is a difference, use the symmetric key pair for EM. h Decrypt the data and find the file hash value K that differs by comparing the elements in the set. x ; Based on the difference file hash value K x Locate the corresponding file.

[0057] In one embodiment of the first aspect, traditional user-tag-based precision marketing systems require a large amount of sample data for their data models. Therefore, a data analysis and modeling algorithm is designed to achieve more effective data analysis and reduce reliance on sample data volume. A multimodal visual language model based on the Transformer architecture is employed. This model converts user media files into text description data through a self-attention mechanism, then analyzes entity relationships based on a knowledge graph to obtain a list of recommended content. Finally, a machine learning recommendation algorithm is used to rank and push the recommended content. A / B testing is used to collect performance evaluation data for fine-tuning the multimodal visual language model parameters and freezing effective parameters to achieve continuous model improvement. Entity relationship analysis based on the knowledge graph is part of the recommendation system process. The multimodal visual language model processes unencrypted raw files on the local client to generate text descriptions, encrypts the text descriptions, and generates hash checksums for the raw files. The text descriptions and hash checksums together serve subsequent processes.

[0058] Knowledge graphs include general knowledge graphs and user profile graphs. Machine learning recommendation algorithms include collaborative filtering (CF) and matrix factorization. Specifically, collaborative filtering finds similarities (similarities between users or between items) based on the behavior of a group, and uses these similarities to make decisions and recommendations for users.

[0059] In one embodiment of the first aspect, various model algorithms and data analysis, such as self-attention mechanisms, are employed to reduce reliance on large amounts of labeled data. Even when user data labels are incomplete or missing, this improves the model's hit rate and provides more accurate content recommendations. This achieves high-accuracy content recommendation with low dependence on labeled data.

[0060] Semi-supervised learning: pre-training the model using a small amount of labeled data and a large amount of unlabeled data; Model pre-training: Pre-training using a large-scale encrypted and de-identified dataset improves model performance; Supervised learning: Even with limited labeled data, effective content recommendation can be achieved through feature engineering and machine learning algorithms; Self-attention mechanism: Using deep learning techniques based on self-attention mechanism to process user behavior sequence data.

[0061] A / B testing: This method uses user grouping and comparison testing to evaluate the existing model. After evaluating the model, the behavioral features of the test users are vectorized, which can be used as parameters to supplement supervised machine learning algorithms for model tuning and training.

[0062] Freeze the model: This allows the recommendation model to learn from new data incrementally without retraining the entire model each time. It is suitable for scenarios with a continuous influx of data, allowing the model to be constantly updated to adapt to new user preferences and behavioral patterns, thus improving the model's adaptability and stability.

[0063] In one embodiment of the first aspect, the trusted cloud storage recommendation system employs a multimodal visual language model: reducing reliance on large amounts of labeled data, and without relying on external data, the multimodal visual language model algorithm architecture is a Transformer deep learning architecture based on a self-attention mechanism. The core formula of the self-attention mechanism is: the scaling dot product attention matrix calculation formula: Calculate the attention score: Calculate the dot product of the query vector and the key vector to obtain the attention score. This is achieved using the following formula, where Q is the query matrix, K is the key matrix, and Ki is the key matrix. T It is the transpose of the key matrix: Attention_scores=QK T Next, calculate the scaling fraction: To stabilize the training process, the point integral is usually scaled. The scaling factor is the square root of the key vector dimension.

[0064] Scaled_scores=

[0065] Applying softmax: The softmax function is applied to the scaled scores to obtain attention weights, which represent the importance of each value vector. Attention_weights=softmax

[0066] Weighted summation: The value vector is summed using attention weights to obtain the final attention output: Attention_output = Attention_weights V, where V is the original text matrix.

[0067] Knowledge Graph: Combining general knowledge graphs and user profile graphs to improve the accuracy of system-recommended content.

[0068] 1) General knowledge graph analysis: concept layer, entity layer, and representation layer. Figure 7 This is a simple, general-purpose knowledge graph example.

[0069] Representation Layer: An entity can have multiple representations, commonly referred to as entity aliases or synonyms. Due to the diversity of natural language, one representation can also refer to multiple entities. Therefore, there is a many-to-many mapping relationship between entity names and entities. A machine learning-based classification model is used to organize entity information in the representation layer; the classification model used is the FastText model. Figure 8 Machine learning-based classification flowchart; Entity Layer: The core of the general knowledge graph is the entity layer. The entity layer contains most of the knowledge in the general knowledge graph, consisting of entities (nodes) and the relationships (edges) between entities. It also includes several entity attributes for each entity and the corresponding attribute values. To construct the entity layer reasonably, the collected data is first processed for entity alignment, that is, identifying the same entity representation in the dataset. After completing entity identification, an entity relationship completion process is performed. Figure 9 ); Conceptual layer: Concepts are abstract generalizations of real things, helping to understand the intrinsic essence of entities. Conceptualization is a uniquely human ability. How to endow computers with concepts and the ability to understand concepts is crucial for computers to understand deep human semantics and to build cognitive and common sense graphs.

[0070] 2) User Profile Graph: A user profile is a representation of a specific user's personal information, typically represented by a series of precise attributes and numerical values. Traditional user profiles are usually built using tag-based modeling methods, with user tag design relying heavily on the experience of business personnel and primarily involving manual compilation and summarization. While tag creation is not particularly difficult, semantic understanding of the tags and subsequent deep applications such as associative reasoning are challenging, which is precisely where knowledge graphs excel. Knowledge graphs abstractly represent and store knowledge of the objective world through structures such as entities and attributes, entities and relationships, and entities and concepts in a machine-understandable format. This allows users to understand profile tags beyond the textual representation (expression layer) itself, enabling deeper profile applications through the entity and concept layers of the underlying knowledge base. The steps for building a user profile graph are: text classification, sequence labeling, and keyword extraction; Text classification: The document text classification model uses the fastText neural network model for text classification; Sequence labeling: Classified text files utilize a deep learning algorithm based on the Transformer architecture to generate vector representations of features. These vector representations are then decoded using a CRF model with conditional constraints, outputting sequence labeling information for each word in the input text. This method, which extracts features from large-scale corpora instead of traditional manually constructed feature templates, effectively solves the problem of text sequence labeling.

[0071] Keyword extraction: The core of the keyword extraction method based on statistical features is to select some statistical features as key indicators of words, and then rank the candidate words according to the scores.

[0072] Commonly used statistical features include those representing the weight of words in a document, such as part-of-speech tagging, term frequency, inverse document frequency, relative term frequency, and word length; those representing the position of words in a document, such as the first N words, the last N words, the beginning of a paragraph, the end of a paragraph, whether they appear in the title or introduction, and word span; and some metrics representing the relevance between words and between words and documents, such as mutual information, Hits value, co-occurrence probability, and TF-IDF value. Among these, TF-IDF (Term Frequency / Inverse Document Frequency) is a very representative metric. It is based on the bag-of-words model and represents the importance of a word w to a document. The calculation formula is as follows: Term Frequency (TF) =

[0073] Inverse document frequency (IDF) = log

[0074] TF-IDF = Term Frequency (TF) × Inverse Document Frequency (IDF) For short texts, a document can be considered a sentence, and the corpus consisting of several sentences represents the total number of documents. As the formula shows, the more times a word appears and the fewer documents it appears in, the stronger its representational power. While the idea behind TF-IDF is very simple, it is highly universal, applicable to both short and long texts. It can also be combined with features such as parts of speech to weight words in a sentence, making it suitable for various domain scenarios.

[0075] Example 2 Based on Figure 1 This application describes a recommendation system implementation apparatus based on cloud storage privacy protection, such as... Figure 10 As shown, it includes: Adjust module 01, used to add an edge node module between the client and server of the trusted cloud disk recommendation system; The extraction and desensitization module 02 is used to call the server's multimodal visual language model to extract standardized file data features locally. The multimodal visual language model first processes the unencrypted original file on the local client to generate user file text descriptions. Then, it uses edge computing to parse the user file into text features locally, generates user file text description data, and desensitizes the file feature data uploaded to the server. The encryption verification module 03 is used for server-side modeling. It uses a fully homomorphic encryption algorithm to encrypt the user file text description data. The encrypted text description data is uploaded to the server for modeling and recommendation. It also uses a file hash encryption algorithm to perform file hash verification on the user's original file to generate a hash verification value. Update module 04 is used by the local client to capture file change hash verification using an event-driven architecture, monitor user behavior in real time and trigger real-time response from the trusted cloud disk recommendation system, and the client analyzes real-time dynamic information of user behavior based on user context awareness, and updates the encrypted user feature data and cached data of the user interest model. The personalized recommendation module 05 uses a multimodal visual language model with a Transformer deep learning architecture. It analyzes the self-attention mechanism and knowledge graph to identify the self-relationships of entity user data, finds high-weight core file features, and combines the resulting list of recommended content with machine learning recommendation algorithms to complete personalized dynamic recommendation, sorting, and push.

[0076] In one embodiment of the second aspect, the edge node module achieves load balancing by transferring data processing and analysis tasks from the central server to the edge of the network, and by distributing the workload across multiple edge nodes through unified allocation of edge nodes via the edge gateway.

[0077] In one embodiment of the second aspect, the data used for server-side modeling is de-identified feature data, and both the original data and the sensitive data are stored in fully homomorphic encrypted form.

[0078] In one embodiment of the second aspect, the fully homomorphic encryption algorithm sets three ciphertext types; specifically, the fully homomorphic encryption algorithm is a scheme based on six functions to complete the data encryption and decryption process.

[0079] User data caching and update data caching rely on data caching provided by edge nodes. When incremental changes in user data lead to changes in recommended content, server-side parameters are updated in batches.

[0080] In one embodiment of the second aspect, the file hash encryption algorithm is a file hash encryption algorithm based on a hash function. The file hash encryption algorithm uses the national cryptographic algorithm SIM3. The SM3 algorithm adopts an MD structure. After the input message is padded, expanded, and iteratively compressed, a hash value with a length of 256 bits is generated. The SM3 algorithm includes padding blocks, message expansion, and iterative compression.

[0081] In one embodiment of the second aspect, an edge computing model architecture is used to adjust the overall architecture of the recommendation system. The cloud drive client completes the user feature engineering algorithm and recommendation model algorithm update locally. An edge gateway layer is added between the client and the server to handle the task of extracting text description information from media files uploaded by cloud drive clients, thus decomposing the modeling process. Data preprocessing is offloaded to the client and edge nodes, with personal data preprocessing completed locally by the user terminal. The user feature engineering algorithm is a federated learning-based user feature engineering algorithm.

[0082] The cloud storage server's visual language model is pre-trained, and the edge nodes synchronously update the recommendation model algorithm and homomorphic encryption algorithm. Users generate their file text information locally by calling the multimodal visual language model, and subsequent data modeling can use these standardized file feature data. The cloud storage server generates customized content by passing through the de-identified file information and encrypted sensitive information of the uploaded encrypted files.

[0083] Based on the data cache provided by the edge nodes, when incremental user data causes changes in the recommended content, the server parameters are updated in batches, thus reducing the data transmission frequency and the amount of data transmitted.

[0084] Edge nodes serve as data preprocessing nodes and data caching nodes, uploading user feature text description data required for recommendation system modeling and achieving incremental data upload by comparing cached data.

[0085] Users can extract stylistic features locally by calling the server's multimodal visual language model. The data used for modeling on the server side is anonymized feature data. Both the original data and the sensitive data are stored in homomorphic encryption, which realizes the protection of users' sensitive data throughout the process, reduces the risk of data interception during transmission, and enhances data privacy protection.

[0086] Edge computing via an edge gateway shifts data processing (including data preprocessing) and analysis tasks from the central server to the edge of the network. At the same time, it distributes workloads across multiple edge nodes through unified allocation of edge nodes by the edge gateway, thus achieving load balancing.

[0087] In one embodiment of the second aspect, user data is encrypted using a toroidal fully homomorphic encryption algorithm. Simultaneously, the trusted cloud storage recommendation system performs file hash verification on user files, generating a hash digest verification value. The generated hash digest information is compared to quickly determine if the user file has been modified, thus preventing tampering. This achieves data modeling and personalized content recommendation of user data under a fully encrypted state throughout the entire process.

[0088] Fully homomorphic encryption algorithms are based on the torus mathematical structure. They provide bootstrapping binary gates to implement the required functions. The real torus used is denoted as T = Rmod1, and the N-dimensional polynomial with coefficients being elements of the torus is denoted as T. N [X] uses bootstrapping to refresh the ciphertext after each logic gate operation to reduce noise in the ciphertext.

[0089] In one embodiment of the second aspect, the TFHE algorithm completes the data encryption and decryption process through a function; In one embodiment of the second aspect, in the TFHE algorithm, the homomorphic evaluation of the binary gate is achieved by performing logical operations on the TLWE ciphertext and then passing it through a gate bootstrap function.

[0090] In one embodiment of the second aspect, an event real-time listening and reporting SDK toolkit can be embedded in the cloud disk client.

[0091] In one embodiment of the second aspect, the event-driven architecture responds to changes in user behavior in real time, and user context-aware analysis captures user dynamic information in real time, dividing user context information into user state context information and user file content context information; user state context information is monitored and collected in real time through the client SDK, and the collected information includes user system information, user location information, user application information, user device information, and user device status information. In one embodiment of the second aspect, user file content context information is monitored in real time through hash aggregation, as follows: When a user uploads a single file to the cloud storage system, a hash value K is generated for that single file. n ; Let a set C be defined. n Used to store user file hash values; The set is encrypted using the AES symmetric encryption algorithm to generate the summary ciphertext EM. h ; Summary of EM at different times h Compare the files; if they match, it means the user files have not been changed. For example, summaries of different time periods EM h If there is a difference, use the symmetric key pair for EM. h Decrypt the data and find the file hash value K that differs by comparing the elements in the set. x ; Based on the difference file hash value K x Locate the corresponding file.

[0092] In one embodiment of the second aspect, the trusted cloud disk recommendation system adopts a multimodal visual language model. The multimodal visual language model algorithm architecture is a Transformer deep learning architecture based on a self-attention mechanism. The scaling dot product attention matrix calculation formula of the self-attention mechanism is as follows: first calculate the attention score, then calculate the scaling score, apply the softmax function to obtain the attention weights, and perform weighted summation: use the attention weights to perform weighted summation on the value vector to obtain the final attention output.

[0093] Knowledge graphs include general knowledge graphs and user profile graphs. Combining general knowledge graphs and user profile graphs can improve the accuracy of system recommendations. Machine learning recommendation algorithms include collaborative filtering and matrix factorization.

[0094] The following steps are required to build a user profile graph: Text classification: The document text classification model uses the fastText neural network model for text classification; Sequence labeling: Classified text files utilize a deep learning algorithm based on the Transformer architecture to generate vector representations of features. These vector representations are then decoded using a CRF model with conditional constraints, outputting sequence labeling information for each word in the input text. This method, which extracts features from large-scale corpora instead of traditional manually constructed feature templates, effectively solves the problem of text sequence labeling.

[0095] Keyword extraction: Keyword extraction based on statistical features involves selecting some statistical features as key metrics for words, and then ranking candidate words according to the scores.

[0096] Example 3 Figure 11 A schematic diagram of the structure of a recommendation system implementation device based on cloud disk privacy protection provided in an embodiment of the present invention is shown below. Figure 11 As shown, the cloud-based privacy-preserving recommendation system implementation device includes at least one processor 310, a memory 320, an input device 330, and an output device 340; the number of processors 310 in the cloud-based privacy-preserving recommendation system implementation device can be one or more. Figure 11 Taking a processor 310 as an example; the processor 310, memory 320, input device 330, and output device 340 in the cloud disk privacy-protected recommendation system implementation device can be connected via a bus or other means. Figure 11 Taking the example of a connection between China and Israel via a bus.

[0097] The memory 320, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the cloud-based privacy-preserving recommendation system implementation method in this embodiment of the invention (e.g., adjustment module, extraction and desensitization module, encryption and verification module, update module, and personalized recommendation module). The processor 310 executes the software programs, instructions, and modules stored in the memory 320 to perform various functional applications and data processing of the cloud-based privacy-preserving recommendation system implementation device, thereby implementing the aforementioned cloud-based privacy-preserving recommendation system implementation method.

[0098] The memory 320 may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function; the data storage area may store data created based on terminal usage. Furthermore, the memory 320 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 320 may further include memory remotely configured relative to the processor 310, which can be connected via a network to a cloud-based privacy-preserving recommendation system to implement the device. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0099] Input device 330 can be used to receive input numeric or character information, and to generate key signal inputs related to user settings and function control of the cloud-based privacy-protected recommendation system implementation device. Output device 340 may include a display device such as a screen.

[0100] This application also discloses a computer-readable storage medium storing a computer program. When the computer program is loaded and executed by a processor, it implements the steps of the recommendation system implementation method based on cloud disk privacy protection disclosed in any of the foregoing embodiments.

[0101] In an exemplary embodiment, a computer program product is also provided, including at least one of a computer program and instructions, wherein when the computer program and instructions are executed by an electronic device, they implement the steps of the above-described method for implementing a recommendation system based on cloud disk privacy protection.

[0102] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0103] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0104] The above description is merely illustrative of the embodiments of the present invention and is not intended to limit the present invention. For those skilled in the art, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A cloud disk privacy protection-based recommendation system implementation method, comprising: adding an edge node between the client and the server of a trusted cloud disk recommendation system; calling a multi-modal visual language model of the server to locally extract standardized file data features, the multi-modal visual language model generating user file text description data by processing unencrypted original files locally on the client and parsing the user files into text features through edge computing, and performing desensitization processing on the file feature data uploaded to the server; the server modeling uses a fully homomorphic encryption algorithm to encrypt the user file text description data, the encrypted text description data is uploaded to the server for modeling and recommendation, and a file hash encryption algorithm is used to perform file hash verification on the user original file to generate a hash check value; the local client uses an event-driven architecture to capture file change hash verification, monitors and triggers the trusted cloud disk recommendation system in real time, and updates the user encrypted feature data and cache data of the user interest model based on real-time user behavior dynamic information of user context perception analysis; the multi-modal visual language model uses a Transformer deep learning architecture to analyze the self-association of entity user data through a self-attention mechanism combined with a knowledge graph, find high-weight core file features, and obtain a recommended content list combined with a machine learning recommendation algorithm to complete personalized content dynamic recommendation sorting and pushing.

2. The method of claim 1, wherein: The data used by the server modeling is desensitized feature data, and the original data and sensitive data are fully homomorphic encryption storage.

3. The method of claim 2, wherein: The file hash encryption algorithm is a file hash encryption algorithm based on a hash function, using the SM3 algorithm; the SM3 algorithm includes padding grouping, message expansion, and iteration compression.

4. The method of claim 3, wherein: The user feature engineering algorithm and the recommendation model algorithm are updated on the local client.

5. The method of claim 4, wherein: The user feature engineering algorithm is a user feature engineering algorithm based on federated learning.

6. The method of claim 5, wherein: The edge node serves as a data preprocessing node and a data caching node, uploads user feature text description data required for modeling of the trusted cloud disk recommendation system, and realizes data incremental uploading by comparing cached data.

7. The method of claim 6, wherein: The user original file is subjected to file hash verification to generate a hash check value of a hash digest, the generated hash digest information is compared to determine whether the user file has changed; data modeling and personalized content recommendation are realized in a full-process encryption state.

8. The method of claim 7, wherein: The fully homomorphic encryption algorithm sets three types of ciphertext; the fully homomorphic encryption algorithm is based on six functions to complete the data encryption and decryption process.

9. The method of claim 8, wherein: The fully homomorphic encryption algorithm is based on a torus mathematical structure.

10. The method of claim 9, wherein: The user context perception information is divided into user state context information and user file content context information; The user state context information is collected through real-time monitoring by the client SDK; The user file content context information is monitored in real time through hash summary.

11. The method of claim 10, wherein: The user file content context information is monitored in real time through hash summary, comprising: When a user uploads a single file to a cloud disk system, a hash value K is generated for the single file n ; Set a collection C n For storing user file hash values; Encrypting the set using the AES symmetric encryption algorithm to generate the summary ciphertext EM h ; Aggregated EMs for different times h Comparison is made, and if the same, it indicates that the user file has not changed. As the summary EM of different times h If there is a difference, use the symmetric key to EM h Decryption and find the file hash value K that exists difference through the comparison of set elements x ; According to the difference file hash value K x Find the corresponding file.

12. The method of claim 11, wherein: The trusted cloud disk recommendation system adopts a multi-modal visual language model, and the multi-modal visual language model algorithm architecture is a Transformer deep learning architecture based on a self-attention mechanism. A scaled dot-product attention matrix calculation formula of the self-attention mechanism is as follows: attention scores are calculated first, then scaled scores are calculated, a softmax function is applied to obtain attention weights, and weighted summation is performed: the value vector is weighted and summed by using the attention weights to obtain a final attention output.

13. The method of claim 12, wherein the method further comprises: The knowledge graph includes a general knowledge graph and a user portrait graph. The user portrait graph is constructed by: text classification: a file text classification model uses a fastText neural network model to perform text classification; Sequence labeling: the classified file text uses a deep learning algorithm based on a Transformer architecture to form a vector representation of features, and then decodes by using a CRF model to output sequence labeling information corresponding to each word of the input text.

14. A cloud disk privacy protection-based recommendation system implementation device, comprising: An adjustment module configured to add an edge node module between a client and a server of a trusted cloud disk recommendation system; An extraction and desensitization module configured to call a multi-modal visual language model of the server to locally extract standardized file data features, and to generate user file text descriptions by processing unencrypted original files locally on the client, to parse the user files into text features by edge computing, to generate user file text description data, and to desensitize file feature data uploaded to the server; An encryption verification module configured to use a homomorphic encryption algorithm to encrypt user file text descriptions on the server side, to upload the encrypted text description data to the server side for modeling and recommendation, and to use a file hash encryption algorithm to generate a hash check value by performing file hash verification on the original user files; An update module configured to use an event-driven architecture on the local client to capture file change hash checks, to monitor and trigger real-time responses of the trusted cloud disk recommendation system, and to update user encryption feature data and cache data of a user interest model based on real-time user behavior dynamic information of user context perception analysis on the client; A personalized recommendation module configured to use a Transformer deep learning architecture of the multi-modal visual language model to analyze the self-relevance of entity user data by using a self-attention mechanism in combination with a knowledge graph, to find high-weight core file features, and to complete personalized content dynamic recommendation sorting and pushing by combining a machine learning recommendation algorithm with a recommended content list.

15. An electronic device, comprising: An electronic device includes one or more processors, a memory for storing one or more programs, and when the one or more programs are executed by the one or more processors, the one or more processors implement a cloud disk privacy protection-based recommendation system implementation method as claimed in any one of claims 1-13.

16. A computer readable storage medium characterized by: A computer program is stored, and when the computer program is executed by a processor, the steps of the cloud disk privacy protection-based recommendation system implementation method as claimed in any one of claims 1-13 are implemented.

17. A computer program product comprising at least one of a computer program, instructions, characterized in that, The computer program, at least one of the instructions, is executed by the electronic device to implement the method of claim 1 13. The steps of the method of claim 1-13 for implementing a cloud disk privacy protection-based recommendation system.

Images