Privacy enhanced intelligent search method and system based on multi-round iteration

Abstract

The invention discloses a privacy enhanced intelligent search method and system based on multi-round iteration. The method comprises the following steps: performing hierarchical semantic analysis on a query input by a user; splitting the complex query into sub-queries based on a task dependency graph algorithm; according to the sub-query, retrieving an evidence fragment from the multi-source data, constructing a semantic element coverage matrix to detect a knowledge gap, and if an uncovered element exists, generating a supplementary sub-query for iterative completion until a preset termination condition is met; integrating cross-modal data through a federated learning technology, and generating a structured knowledge graph fragment in combination with semantic vector alignment and an evidence fusion algorithm; performing dynamic desensitization processing on the retrieval result; and a closed-loop iterative updating mechanism is formed based on a user explicit and implicit feedback optimization retrieval strategy. The problems of traditional intelligent search in the aspects of semantic understanding depth, complex problem reasoning, search result accuracy and integrity, user privacy security and the like are effectively solved.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence, natural language processing, and data search technology, specifically to the interdisciplinary field of intelligent query and knowledge reasoning technology, and more specifically to a privacy-enhanced intelligent search method and system based on multiple rounds of iteration. Background Art

[0002] With the rapid development of artificial intelligence (AI), intelligent search has been widely used in various information retrieval scenarios, becoming an important tool for users to acquire knowledge and information. However, current intelligent search technology still faces many challenges, including significant deficiencies in search accuracy, complex problem handling, and user privacy protection.

[0003] In terms of semantic understanding and complex problem handling, traditional natural language processing technology has limited depth in semantic analysis of long texts such as patents, making it difficult to effectively mine implicit information and suffering from semantic gaps. When faced with complex problems, existing search technologies usually adopt a single "search-generate" process, lacking the ability to break down complex problems into sub-problems and conduct multiple rounds of reasoning, resulting in poor accuracy and completeness of answers, and unable to meet users' needs for accurate answers to complex and professional questions. Moreover, in terms of the evaluation of search results, existing technologies often lack a comprehensive and rigorous evaluation system, making it difficult to effectively control the factual accuracy, content completeness, and logical coherence of the answers, and are prone to factual errors, information omissions, or logical contradictions.

[0004] In the area of ​​user privacy protection, traditional intelligent search methods directly capture network data, and user search requests interact directly with the underlying search engine. This makes private data such as users' real IP addresses and device information easily exposed, posing a serious risk of privacy leaks. Furthermore, browser fingerprinting can be used to track users, further threatening user privacy and security. Furthermore, with increasingly stringent data security regulations, users are increasingly aware of their privacy and have placed higher demands on the privacy and security of search services. Existing search technologies struggle to meet this demand.

[0005] In summary, there is an urgent need for a new intelligent search method and system that can enhance intelligent search's deep semantic understanding and complex problem-solving capabilities while simultaneously safeguarding user privacy and security. This approach can fill the gaps in existing technologies and meet users' dual needs for information retrieval and privacy protection. Summary of the Invention

[0006] In view of this, the present invention provides a privacy-enhanced intelligent search method and system based on multiple rounds of iterations, which can effectively solve the difficulties of traditional intelligent search in terms of semantic understanding depth, complex problem reasoning, search result accuracy and completeness, and user privacy security.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] In a first aspect, an embodiment of the present invention provides a privacy-enhanced intelligent search method based on multiple iterations, comprising the following steps:

[0009] S10, performing hierarchical semantic analysis on the query input by the user, extracting core semantic elements through dependency syntax analysis and entity recognition, and generating structured semantic units;

[0010] S20, based on the task dependency graph algorithm, decomposes complex queries into a sequence of subtasks that can be executed in parallel, and generates subqueries through dynamic priority sorting;

[0011] S30, performing multiple rounds of iterative retrieval: retrieving evidence fragments from multi-source data based on subqueries, constructing a semantic element coverage matrix to detect knowledge gaps, and generating supplementary subqueries to iteratively complete uncovered elements if any, until a preset termination condition is met;

[0012] S40: Integrate cross-modal data through federated learning technology, combine semantic vector alignment and evidence fusion algorithms to generate structured knowledge graph fragments;

[0013] S50, dynamically desensitizing the search results, including masking, generalization, and differential privacy, and aggregating the anonymized search results through a meta-search proxy engine and outputting the results;

[0014] S60. Optimize the retrieval strategy based on user explicit and implicit feedback to form a closed-loop iterative update mechanism.

[0015] In a second aspect, an embodiment of the present invention further provides a privacy-enhanced intelligent search system based on multiple iterations, using the privacy-enhanced intelligent search method based on multiple iterations as described in the first aspect, the system comprising:

[0016] Semantic Parsing Module: This module performs hierarchical semantic parsing on user-entered queries, extracts core semantic elements through dependency parsing and entity recognition, and generates structured semantic units. It also decomposes complex queries into a sequence of subtasks that can be executed in parallel based on a task dependency graph algorithm, and generates subqueries through dynamic priority sorting.

[0017] Multi-round reasoning engine: used to perform multiple rounds of iterative retrieval: retrieve evidence fragments from multi-source data based on subqueries, build a semantic element coverage matrix to detect knowledge gaps, and generate supplementary subqueries to iteratively complete uncovered elements until the preset termination condition is met;

[0018] Cross-modal fusion module: used to integrate cross-modal data through federated learning technology, combining semantic vector alignment and evidence fusion algorithms to generate structured knowledge graph fragments;

[0019] Privacy and security framework module: Dynamically desensitizes search results, including masking, generalization, and differential privacy, and aggregates anonymized search results through a meta-search proxy engine before outputting them.

[0020] Feedback optimization module: Optimizes retrieval strategies based on user explicit and implicit feedback to form a closed-loop iterative update mechanism.

[0021] It can be seen from the above technical solutions that compared with the prior art, the present invention has the following technical advantages:

[0022] This invention achieves deep semantic understanding and precise answers for intelligent search through hierarchical semantic analysis, cognitive reasoning chain construction, and iterative multi-round reasoning architecture. At the same time, it utilizes third-party search engine result aggregation, random browser profile generation, IP hiding and other technologies to create a safe and reliable privacy protection system, which is suitable for various application scenarios with high requirements for information search quality and user privacy protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0024] Figure 1 This is a flow chart of the privacy-enhanced intelligent search method based on multiple rounds of iterations provided by the present invention.

[0025] Figure 2 This paper provides a technical roadmap for the privacy-enhanced intelligent search method based on multiple rounds of iterations.

[0026] Figure 3 This is the core logic roadmap for display data acquisition and in-depth data search provided by the present invention.

[0027] Figure 4 This is an architecture diagram of the privacy-enhanced intelligent search system based on multiple rounds of iterations provided by the present invention. DETAILED DESCRIPTION

[0028] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0029] The cognitive reasoning search framework constructed by the present invention simulates the human cognitive reasoning process to achieve in-depth analysis and multi-round iterative retrieval of complex queries. Figure 1 As shown, the embodiment of the present invention discloses a privacy-enhanced intelligent search method based on multiple iterations, including the following steps S10 to S60:

[0030] S10, performing hierarchical semantic analysis on the query input by the user, extracting core semantic elements through dependency syntax analysis and entity recognition, and generating structured semantic units;

[0031] This step uses the Stanford Parser dependency parser to build a parsing engine. The query text is first segmented. Using a combination of a Hidden Markov Model (HMM) and a Conditional Random Field (CRF), the continuous text is divided into independent words, and each word is labeled with a part-of-speech tag. The Kashima parsing algorithm is then used to identify dependencies between words and generate a dependency tree.

[0032] The BiLSTM-CRF model is used for entity type recognition to convert the segmented words into word vectors, and the part-of-speech tags are converted into one-hot encodings, which are then concatenated to form the input feature vector.

[0033] The Bidirectional Long Short-Term Memory (BiLSTM) network captures the contextual semantics of words. The forward LSTM processes the sequence from left to right, while the reverse LSTM processes the sequence from right to left, ultimately outputting a hidden state vector containing the contextual information. The Conditional Random Field (CRF) model is used to model the hidden state output by the BiLSTM. The Viterbi algorithm is used to decode the optimal label sequence, taking into account the transition probability between labels, to achieve accurate entity type recognition.

[0034] In the specific implementation, it is necessary to perform text preprocessing on the query input by the user, including the following:

[0035] (1) Special character cleaning: Using regular expression matching technology, define the regular pattern [^\w\s\u4e00-\u9fa5] that contains non-alphanumeric characters (such as punctuation marks, special symbols, and control characters). Use the re.sub() function to match and replace the input text character by character to ensure that Chinese characters, letters, numbers, and necessary spaces are retained. During the text format unification process, full-width characters are converted to half-width characters (such as full-width spaces are converted to half-width characters). Character standardization is achieved through the Unicode character encoding mapping table to ensure that texts in different encoding formats are unified into UTF-8 encoding.

[0036] (2) Text normalization: For English text, perform lowercase conversion and convert all uppercase letters to lowercase using the str.lower() function; for Chinese text, use a simplified / traditional conversion library (such as OpenCC) to convert traditional text to simplified text to ensure the format consistency of multilingual text.

[0037] (3) Implementation of lexical and syntactic analysis

[0038] Word segmentation and part-of-speech tagging: Chinese word segmentation uses the THULAC word segmentation system, which is based on a word segmentation algorithm that combines a hidden Markov model (HMM) and a conditional random field (CRF). After preprocessing, the input text is fed into the word segmenter as a character sequence, and the output includes word boundary information. Part-of-speech tagging is performed simultaneously. THULAC has a built-in set of 58 part-of-speech tags. The optimal part-of-speech tagging sequence is determined using a dynamic programming algorithm, achieving a tagging accuracy of 97.3%.

[0039] Dependency Syntax Analysis: Using Stanford CoreNLP's graph-based dependency syntax model (such as the Kashima dependency syntax model), we convert the segmented word sequence into a dependency tree. Each word serves as a tree node, and the edges between nodes represent dependency relationships (such as nsubj subject-verb relationships and dobj verb-object relationships). By parsing the syntax tree, we can obtain the grammatical dependencies between words, providing structural support for subsequent semantic analysis.

[0040] (4) Technical Implementation of Semantic Role Labeling

[0041] Model Architecture: A BERT-based semantic role labeling model is used. The input layer converts the segmented word sequence into a BERT word embedding (768 dimensions), combining positional and segment encoding to generate a contextual representation. The model's mid-layer uses a 12-layer Transformer encoder to capture long-range semantic dependencies, and the output layer uses a multi-layer perceptron (MLP) to predict semantic role labels (e.g., ARG0 agent, ARG1 patient, ARGM-MOD modifier)) for each word.

[0042] Training and Inference: The training data uses the CoNLL-2010 Semantic Role Labeling dataset. After pre-training with a Masked Language Model (MLM), the model is fine-tuned on domain-specific data (such as technical documentation and medical guidelines) using a cross-entropy loss function. During the inference phase, the model processes the input sentence and outputs the semantic role label for each word using the IOB2 annotation format, supporting the recognition of 18 types of predicate-argument structures.

[0043] S20, based on the task dependency graph algorithm, decomposes complex queries into a sequence of subtasks that can be executed in parallel, and generates subqueries through dynamic priority sorting;

[0044] After step S10, sub-query expansion may be performed;

[0045] By adopting the entity association expansion algorithm and combining remote supervision with manual annotation, a knowledge graph containing entities, attributes, and relationships is constructed. For example, the technical field graph contains the triple relationship of "technical term → related field → application scenario".

[0046] For the identified core entities, we query their hyponym, attribute, and related concepts through the graph to generate an extended entity set. We use the SRL model to identify the argument roles of verb phrases.

[0047] In specific implementation, a production-rule-based decomposition engine is built, with the rule format being IF <syntactic structure> AND <semantic role> THEN <subquery generation template>. For example, when the predicate is "recommendation" and the patient role is "learning resource," the subquery template "{modifier features of the patient role}" is triggered, generating the subquery "Efficient evaluation criteria for learning resources."

[0048] Dynamic priority sorting: Use the multiple attribute decision making algorithm (MADM) to define the subquery priority formula:

[0049]

[0050] Among them, α = 0.6, β = 0.4 are weight coefficients, TF-IDF(q) is the term frequency of the subquery keyword minus the inverse document frequency, and Sim(q, Q) is the cosine similarity between the subquery and the original query.

[0051] S30, performing multiple rounds of iterative retrieval: retrieving evidence fragments from multi-source data based on subqueries, constructing a semantic element coverage matrix to detect knowledge gaps, and generating supplementary subqueries to iteratively complete uncovered elements if any, until a preset termination condition is met;

[0052] This step includes the implementation process of dynamic reasoning chain generation and knowledge gap completion:

[0053] Among them, dynamic reasoning chain generation includes:

[0054] 1) Initial search and evidence extraction

[0055] A distributed vector database (Milvus) is used to convert subqueries into 768-dimensional vectors using a pre-trained text embedding model. The top-K similar documents are retrieved from the document vectors using the vector database's Annoy index or the HNSW algorithm. The cosine similarity is used to calculate the similarity.

[0056] The large language model is used to generate a credibility score for the document fragment. The input is "subquery + document fragment" and the output is a score value between 0 and 1. The calculation formula is:

[0057] Score = LLM(Query, Document)∈[0,1], retaining evidence fragments with a score ≥ 0.6.

[0058] 2) Logical chain verification and gap identification

[0059] The model is fed multiple pieces of evidence and asked to identify any contradictions, outputting the location and type of contradictions (e.g., data conflicts, logical gaps). Based on the types of knowledge relationships (e.g., causal relationships, progressive relationships), it verifies whether the chain of evidence covers all necessary relationships, for example, whether the relationship from "technical principle to application case" is complete.

[0060] A semantic element coverage matrix is ​​constructed, where the rows represent the semantic elements (entities, relations, constraints) of the subquery and the columns represent the coverage status of the evidence set (0 = uncovered, 1 = covered). The gap score is calculated as the ratio of the number of uncovered elements to the total number of elements, using the formula:

[0061] When GapScore>0.3, the gap filling mechanism is triggered.

[0062] 3) Multiple rounds of iteration

[0063] Subquery generation model: A Seq2Seq architecture (encoder-decoder model) is used to generate supplementary retrieval queries. The encoder input is the current evidence gap feature, and the decoder outputs a new subquery. The training data is the historical query-completion query pair, and the loss function is the cross-entropy loss.

[0064] Stop condition judgment logic: Set dual stop conditions - evidence confidence threshold (average credibility of all evidence ≥ 0.8) and maximum number of iterations (default 5 rounds), and decide whether to terminate the loop through logical AND operation.

[0065] The implementation process of knowledge gap filling:

[0066] 1) Confidence quantification

[0067] The source authority score is scored using a multi-source cross-validation algorithm: weights are assigned to data sources (e.g., .gov domains have a weight of 0.9, .com domains have a weight of 0.6), and the calculation formula is:

[0068]

[0069] Assertion consistency check: For the same entity or relationship, if ≥2 independent data sources support the same assertion, the assertion credibility is increased by 0.3, otherwise it is reduced by 0.2. The final evidence credibility is the weighted average of the two.

[0070] Subquery coverage scoring matrix: Construct a three-dimensional scoring matrix (entity dimension, relationship dimension, constraint dimension), with each dimension scoring ranging from 0 to 1. The calculation formula is:

[0071]

[0072] The factor weights are predefined based on domain knowledge (e.g., entity weight 0.4, relationship weight 0.3, constraint weight 0.3).

[0073] 2) Dynamic gap detection semantic matching

[0074] Semantic element matching engine: Detects gaps element by element by combining regular expressions with semantic vector matching:

[0075] Entity matching: Based on NER tags, check whether the evidence set contains all entity types defined by the subquery; Relationship matching: Through the results of dependency syntax analysis, verify whether the relationship types in the evidence cover the predicate argument structure of the subquery; Constraint matching: Extract the time, scope, credibility and other constraints in the query (such as "in the past three years" and "credibility ≥ 0.7") and check whether the evidence meets them.

[0076] Gap prioritization model: The multi-attribute decision-making (MADM) method is used to comprehensively consider the gap impact (the degree of impact on the core intent) and the retrieval cost (the difficulty of obtaining the data source). The formula is:

[0077] The impact is calculated using predefined weights by domain experts, and the retrieval cost is quantified based on the data source type (public data = 1, internal data = 3, cross-domain data = 5).

[0078] In specific implementation, the technology of dynamic reasoning chain generation is as follows, evidence retrieval technology solution:

[0079] Knowledge Graph Retrieval: We use the SPARQL query language to construct search statements, using the triple pattern SELECT?entity WHERE{?entityrdf:type:Concept} to retrieve relevant entities within the knowledge graph. For complex relationship queries, we support joint and nested queries, such as using the OPTIONAL keyword to handle optional relationships, ensuring the completeness of search results.

[0080] Text database retrieval: Implement text retrieval based on the Vector Space Model (VSM), converting subqueries and documents into TF-IDF vectors:

[0081]

[0082] Where TF(t, d) is the frequency of word t in document d, df(t) is the number of documents containing word t, and N is the total number of documents. IDF stands for inverse document frequency, and the similarity between the query vector and the document vector is calculated using the cosine similarity formula:

[0083] Retrieve the top K (default K = 20) similarity-ranked documents as candidate evidence.

[0084] Sim(u q ,u d )Query vector u q With the document vector u d The cosine similarity between them ranges from [0,1] (if the vector is non-negative, such as the text TF-IDF vector).

[0085] u q Query vector (QueryVector) is the vector representation of the query content entered by the user (such as search keywords, question description, etc.) after vectorization processing (such as TF-IDF, word embedding, topic model, etc.).

[0086] u d Document vector (DocumentVector), that is, the vector representation of document content (such as text, web page, knowledge base entry, etc.) after being processed by the same vectorization method, is the same as u q Same dimension, same feature space.

[0087] K (default (K=20)) is the truncation parameter of the retrieval results, which means that the top K documents ranked by cosine similarity are selected as candidate evidence (such as candidate facts in knowledge graph construction, candidate results in information retrieval, etc.).

[0088] Reasoning model architecture: This model uses a generative encoder-decoder architecture. The encoder encodes evidence sequences into context vectors, and the decoder generates inference results based on an attention mechanism. During pre-training, the model uses massive amounts of text data to learn general reasoning capabilities. During fine-tuning, the model optimizes reasoning tasks in specific domains (such as finance and healthcare), supporting multiple types of logical reasoning, including conditional and causal reasoning.

[0089] Logical Verification Mechanism: Build a verification system that combines a rule base with a statistical model. The rule base contains more than 200 domain-specific logical rules, and uses regular expressions to match the logical structure in the inference results. The statistical model uses a logical consistency scoring function:

[0090] Among them, n is the total number of verified inference results, r iis the logical relationship in the inference result, R is the valid relationship set in the rule base, δ is the indicator function (1 when the rule is met, otherwise 0), and when Score < 0.7, the inference result correction is triggered.

[0091] Inference chain construction algorithm:

[0092] Chain structure representation: A directed acyclic graph (DAG) is used to represent the reasoning chain, with nodes representing subqueries or inference conclusions and edges representing logical relationships (e.g., "cause → result" or "condition → conclusion"). Each node contains a summary, evidence source, and a confidence score (0-1), while edges contain the relationship type and weight (calculated based on PMI point mutual information).

[0093] Iterative reasoning control: Two termination conditions are set: ① the average confidence level of the evidence is ≥ 0.8; ② the number of iterations reaches the preset upper limit (5 by default). Each iteration focuses on uncovered semantic elements through the attention mechanism, generating a complete query, and achieving dynamic expansion and optimization of the reasoning chain.

[0094] In specific implementation, the implementation method of knowledge gap filling is as follows: Gap identification technology solution

[0095] Semantic coverage calculation: Constructing the semantic element matrix M n*m , n is the number of semantic elements (entities, relations, constraints) of the original query, and m is the number of elements covered by the evidence. The coverage calculation formula is:

[0096] Among them, match(e i ,c j ) is the semantic element matching function (1 if there is a match, 0 otherwise). Gap filling is triggered when Coverage < 0.6.

[0097] Evidence relevance score: The BM25 algorithm is used to calculate the relevance of evidence to gap elements:

[0098] Where k1 = 1.2 and b = 0.75 are empirical parameters, |d| is the document length, avgdl is the average document length, and evidence with a score < 0.3 is considered invalid evidence. TF(t, d) is the frequency of word t in document d, and q is the query input by the user.

[0099] S40: Integrate cross-modal data through federated learning technology, combine semantic vector alignment and evidence fusion algorithms to generate structured knowledge graph fragments;

[0100] After the dynamic gap detection and semantic matching step in step S30, step S40 can further implement an intelligent completion strategy, as follows:

[0101] Weakly related evidence expansion algorithm: Leveraging the context generation capability of LLM, weakly related evidence and subqueries are input, and derivative queries are generated by prompting "Please generate related supplementary questions based on the following information: {evidence fragment}". For example, the completion query "What are the common types of vulnerabilities in smart contracts?" is generated from the evidence "Smart contract technical principles".

[0102] Cross-modal data retrieval interface: Design a unified cross-modal retrieval API that automatically triggers image / audio retrieval modules when text evidence is insufficient:

[0103] Image retrieval: The CLIP model is used to convert subqueries into visual semantic vectors to retrieve image data such as technical architecture diagrams and flowcharts.

[0104] Audio retrieval: Use the Whisper model to convert audio data into text and then perform semantic matching.

[0105] User interaction completion mechanism: Develop a front-end interactive component. When a gap in internal proprietary data is detected, a modal window pops up prompting "The current information requires querying the internal knowledge base. Do you want to authorize access?" After obtaining user authorization through the OAuth 2.0 protocol, access is made to the company's internal database for retrieval.

[0106] In specific implementation, the process of generating and executing the completion strategy is as follows:

[0107] Strategy Selection Model: A strategy selector is constructed based on a decision tree algorithm. It inputs gap type (missing entity, missing relationship, missing constraint) and evidence characteristics (quantity, credibility, source), and outputs the optimal completion strategy (such as "cross-modal retrieval," "knowledge graph reasoning," or "user interaction request"). The decision tree node splitting condition is based on information gain. Trained with the ID3 algorithm, the strategy selection accuracy reaches 92%.

[0108] Cross-modal completion implementation: Image completion uses the CLIP model to convert text gaps into visual semantic vectors, calculate cosine similarity with the image feature vector (1024-dimensional features extracted by ResNet), and retrieve image data with a similarity greater than 0.7. Audio completion uses the Whisper model to convert audio into text and then perform semantic matching. It supports speech recognition in 10+ languages ​​and has a word error rate (WER) of ≤5%.

[0109] Evidence fusion algorithm: DS evidence theory is used to fuse multi-source evidence, defining the basic probability assignment (BPA) of the evidence body \(E_i\) and merging the evidence through orthogonal sum operations:

[0110]

[0111] Among them, m(A) is the trust degree of the integrated evidence on proposition A. This algorithm solves the problem of evidence conflict and improves the credibility of the integrated evidence.

[0112] A1,A2......A n : represent the 1st to nth evidence bodies respectively;

[0113] A1∩...∩A n =A: means that the intersection of all the evidences supporting the proposition is equal to A, that is, these evidences jointly support proposition A;

[0114] The i-th body of evidence for proposition A i The product of the basic probability distribution of , represents the strength of the consistent evidence that multiple evidences jointly support A.

[0115] Logical consistency check: Use a symbolic logic reasoning engine (such as Drools) to formally verify the integrated information, convert the reasoning results expressed in natural language into first-order logic expressions, verify whether there are logical contradictions, and ensure the self-consistency of the completed knowledge system.

[0116] S50, dynamically desensitizing the search results, including masking, generalization, and differential privacy, and aggregating the anonymized search results through a meta-search proxy engine and outputting the results;

[0117] The privacy-safe search framework designed in this step uses multi-dimensional technical means to build a privacy protection system covering the entire process of "data input-retrieval processing-result output". It is suitable for high-risk scenarios that are sensitive to data security, such as finance, government affairs, and medical care.

[0118] 1. Multi-source aggregation anonymized retrieval technology

[0119] 1) Meta-search proxy engine

[0120] By building independent meta-search middleware, user search requests are not sent directly to underlying search engines like Baidu and Google, but are instead routed through a proxy server. The proxy server parses the request, splits it into multiple sub-requests, and sends them to different search engines (such as Bing and DuckDuckGo) simultaneously, aggregating the results and returning them to the user.

[0121] Generate a unique session-level anonymous identifier for each search session using the UUID v4 algorithm combined with a device fingerprint hash (such as browser UA, IP address hash). The formula is:

[0122] AnonID=HMAC(SALT,UUIDv4()+DeviceFinger print)

[0123] Among them, SALT is the server-side random key, and DeviceFingerprint is the device feature hash value, which ensures that the identity identifiers between different sessions are unrelated.

[0124] 2) Data transmission encryption

[0125] The TLS1.3 protocol is used to encrypt communication between users and proxy servers. The proxy server and the underlying search engine are connected via HTTPS. The AES-256-GCM algorithm is used to encrypt the payload during data transmission to ensure that the request content cannot be stolen by man-in-the-middle attacks.

[0126] Dynamically desensitize sensitive information in query parameters (such as user ID and geographic location). For example, generalize "users in Chaoyang District, Beijing" to "users in North China" using regular expression replacement or entity replacement algorithms.

[0127] 2. Hierarchical data access control system

[0128] 1) Privacy Information Classification Model

[0129] Two-dimensional evaluation system:

[0130] Universality score (U): The number of search engine results returned by querying "whether a certain information is private" is used to calculate the awareness of the information among the user group. The formula is:

[0131] Where N1 is the number of query results for specific information, and μ1 is the average value of all private information.

[0132] Sensitivity score (S): The sensitivity is assessed by the ratio of results of "privacy items + leakage" related queries. The formula is:

[0133] Among them, N2 is the number of leakage-related query results, and μ2 is the average β value of all private information.

[0134] Security level classification:

[0135] According to the U*S results, it is divided into three levels: Level 1 (0.25), Level 2 (0.5), and Level 3 (1). Different levels correspond to different access control policies (for example, level 3 information is only accessible to administrators).

[0136] 2) Fine-grained permission management

[0137] The privacy level dimension is introduced into the traditional role-authority model (RBAC), and user roles are associated with privacy levels. For example:

[0138] Ordinary users: can access level 1 information and need to apply for level 2 information permissions;

[0139] Administrator: Has level 3 information access by default.

[0140] Trigger permission verification before data retrieval, matching user roles with information security levels through the rule engine. The verification logic is as follows:

[0141] Access Permit = f(user role, information level, operation type)

[0142] Where f is a predefined permission function that supports attribute-based access control (ABAC) extension.

[0143] 3. Data full-link security protection technology

[0144] 1) Input layer security enhancement

[0145] Disable browser fingerprint tracking: By configuring the browser proxy plug-in, hide features such as User-Agent and Canvas fingerprint, and use randomized plug-in fingerprints to simulate normal user behavior.

[0146] Search intent confusion: Add interference items to high-frequency search terms (such as mixing "bank card balance inquiry" into "bank card password reset") to reduce the probability of user intent being tracked.

[0147] 2) Processing layer security mechanism

[0148] In cross-institutional collaborative retrieval scenarios, federated learning technology is used to achieve "data remains static, model moves":

[0149] Each institution trains the retrieval model locally and only uploads the model parameters (such as vector database index) to the central server to avoid the transmission of raw data.

[0150] The central server aggregates the model parameters of multiple parties, generates a global retrieval model, and returns the encrypted retrieval results to the requester.

[0151] The retrieval operation of sensitive data is recorded on the blockchain, recording "user ID-timestamp-operation type-data hash" to ensure that the operation is traceable. The SHA-256 algorithm is used to generate the data hash. The formula is:

[0152] Hash = SHA-256 (user ID + timestamp + data content)

[0153] 3) Output layer desensitization processing

[0154] Mask processing: Display the first 3 digits and last 4 digits of sensitive fields such as ID card number and mobile phone number, and fill the middle with asterisks (such as 138****1234).

[0155] Generalization processing: The precise geographic location "Chaoyang District, Beijing" is generalized to "Beijing City", and natural language processing technology is used to identify address entities and perform hierarchical processing.

[0156] Differential privacy: Add Laplace noise to the statistics, the formula is:

[0157]

[0158] Among them, e is the privacy budget parameter that controls the noise intensity.

[0159] Technical coupling of cognitive reasoning and privacy security:

[0160] Privacy enhancement during reasoning:

[0161] During the multi-round semantic decomposition phase, the privacy framework marks sensitive entities in user queries (such as "medical record number" and "financial data"), triggering dynamic desensitization rules to ensure that sub-queries do not contain original sensitive information.

[0162] When a dynamic reasoning chain is generated, the privacy framework assigns an independent virtual identity to each round of retrieval requests to prevent multiple rounds of interactions from exposing the user's true intentions.

[0163] Dynamic adaptation of security policies:

[0164] Based on the security level of the cognitive reasoning results (such as level 3 sensitive information), the privacy framework automatically increases the encryption level of the retrieval process (such as using AES-512 encryption instead of 256 encryption) and limits the output format of the results (only allowing PDF encrypted documents to be downloaded).

[0165] For reasoning tasks involving cross-institutional collaboration (such as joint retrieval of medical data), the privacy framework activates the federated learning model to ensure that the data of each participant does not leave the local area while meeting the cross-domain data requirements of cognitive reasoning.

[0166] Unified management and control of heterogeneous data sources:

[0167] Internal enterprise documents, public web pages, and third-party API data required for cognitive reasoning are all accessed through the unified security interface of the privacy framework to achieve authentication, authorization, and audit (AAA) management of the data source.

[0168] For access to the internal knowledge base, the privacy framework combines RBAC with a hierarchical model to ensure that only users with corresponding permissions can trigger internal data retrieval in cognitive reasoning.

[0169] In specific implementation, the technical implementation process of multi-source aggregation anonymization is as follows: Multi-source search request forwarding mechanism:

[0170] Request Splitting and Routing: A round-robin load balancing algorithm is used to evenly distribute user requests to over 20 connected search engines (such as Baidu, Google, and DuckDuckGo). Dynamic weighting is supported (routing strategies are optimized in real time based on search engine response speed and return result quality). When processing request parameters, query keywords are randomly shifted (for example, a random number 1-5 is added or subtracted from the ASCII code value of each character) to generate variant queries, reducing the probability of request fingerprinting.

[0171] Response aggregation processing: Design a unified result format converter to convert the results returned by different search engines (such as HTML and JSON) into a standardized structure (including title, summary, URL, and source). The deduplication module uses the SimHash algorithm to calculate the similarity of results, with a threshold of 0.9 to ensure that the merging rate of duplicate results is ≥95%, improving the efficiency of result aggregation.

[0172] The anonymization processing technology solution is as follows: Session anonymous identifier generation: The UUID v4 algorithm is used to generate the session ID. The input parameters are the namespace UUID and the device fingerprint (including the browser User-Agent, IP address hash, and device hardware information hash). The formula is:

[0173] SessionID=UUIDv4(NamespaceUUID,SHA-56(UA+IPHash+HardwareHash))

[0174] The NamespaceUUID is fixed to the search service-specific UUID, ensuring that the generated SessionID is globally unique, with an entropy value of 128 bits and a collision resistance probability of ≤1e-36.

[0175] S60. Optimize the retrieval strategy based on user explicit and implicit feedback to form a closed-loop iterative update mechanism.

[0176] A two-way feedback mechanism: Explicit feedback (users marking "add SCI database") directly updates the query generation strategy (increasing the weight of that data source by 50%), while implicit feedback (staying on a chart area for more than 2 minutes) triggers additional analysis suggestions (such as "Is it necessary to add this technical and economic cost comparison?"), resulting in a 42% increase in search result satisfaction. For example, after a business user repeatedly modified the competitive product analysis dimensions, the system automatically optimized the subtask priorities, setting "pricing strategy comparison" to the highest priority.

[0177] Domain-adaptive parameter configuration: Generates exclusive parameter templates based on academic, commercial, and consumer scenarios (the default depth for academic research is 3, focusing on deep reasoning; the default depth for consumer decision-making is 2, focusing on quick comparison). The breadth / depth parameters are automatically adjusted based on historical task data (for example, if the average duration of a complex task is >30 minutes, the depth value is automatically lowered), balancing efficiency and accuracy. This improves task completion time by more than 30% in different scenarios and adapts to diverse user habits and needs.

[0178] In specific implementation, the privacy rating model realizes:

[0179] Calculation of the universality score (U): We regularly (daily) collect query results N1 from platforms such as Baidu Knows and Zhihu through web crawlers, and calculate the average value μ1 of all private information using the piecewise function:

[0180]

[0181] Among them, σ1 is the standard deviation, which reflects the range of data fluctuation.

[0182] Sensitivity score (S) calculation, defined as: β = N2 / (N1+N2), where N2 is the number of query results related to "privacy items + leakage". Similarly, the levels are divided by the mean μ2 and standard deviation σ2.

[0183] Dynamic permission verification: Using the attribute-based access control (ABAC) model, the permission decision formula is: Allow = f(UserAttr,DataAttr,Action), where UserAttr is the user attribute (role, department, security level), DataAttr is the data attribute (security level, creation time, project), and Action is the operation type (query, download, share). The decision tree algorithm is used to achieve visual configuration and efficient matching of permission rules.

[0184] like Figure 2 As shown in FIG, a flow chart of the principle of the privacy-enhanced intelligent search method based on multiple rounds of iterations provided by the present invention is shown:

[0185] 1. User input processing:

[0186] The user's original query is analyzed through named entity recognition and keyword extraction technology, core entities are extracted and labeled, and fuzzy expressions are standardized and converted based on the domain dictionary library. At the same time, implicit conditions are completed based on the user's historical interaction data.

[0187] 2. Intent decomposition: The dual-tower comparative learning model is used to distinguish between task-based intent and conversation-based intent. The former accounts for 70% and the latter accounts for 30%.

[0188] 3. Dynamic reasoning chain task splitting:

[0189] Task-based intentions are split into subtasks that can be executed in parallel through the task dependency graph algorithm, generating triples containing the execution subject and input and output.

[0190] Subtasks are allocated computing power based on complexity and scheduled to heterogeneous LLM clusters through a load balancing algorithm, supporting user-defined parallelism.

[0191] 4. Retrieval and preprocessing:

[0192] Each LLM instance generates exclusive queries based on subtask goals, starts asynchronous crawlers and API calls for multi-source concurrent retrieval, and processes data using a credibility-first strategy.

[0193] Perform multimodal preprocessing on unstructured data, including OCR text extraction, table parsing, and sentence segmentation, to generate computable semantic units.

[0194] 5. Multiple rounds of iterative research:

[0195] Perform preliminary screening of the search results, retain highly relevant data, calculate the data source weight through the attention mechanism, and construct knowledge graph fragments.

[0196] Based on the depth parameters and the confidence level of the conclusion, it is determined whether to continue the iteration. If the conditions are met, a new subtask is generated and the priority is adjusted. Otherwise, the iteration is terminated.

[0197] 6. Result output and closed-loop optimization:

[0198] Integrate multi-source data to build a traceable evidence chain, use DS evidence theory to handle data conflicts, and generate structured reports based on academic, commercial, consumer and other scenarios.

[0199] 7. Collect explicit and implicit user feedback to dynamically adjust system parameters and update the knowledge base to form a closed-loop optimization.

[0200] Among them, in terms of user intent analysis, the present invention adopts a dual-tower comparative learning model to integrate semantic features and historical interaction data for intent classification; traditional search systems rely on static rule matching, and fixed rule tools use preset fixed classification models.

[0201] In complex task processing, the present invention automatically splits parallel subtasks through a task dependency graph algorithm; traditional systems require manual definition of subtasks, and fixed rule tools can only process single tasks and do not support splitting.

[0202] In terms of multimodal data processing capabilities, the present invention supports deep integration and preprocessing of text, images, and PDFs; traditional systems process each modality independently, and fixed rule tools can only process text data.

[0203] In terms of weight adjustment and model updating, the present invention realizes dynamic weight allocation and incremental learning. Traditional systems and fixed rule tools are limited by fixed strategies and lack of iteration capabilities, respectively.

[0204] Table 1

[0205]

[0206]

[0207] like Figure 3 As shown, the core logic of data acquisition and in-depth data search is shown in detail:

[0208] 1) Data Ingestion

[0209] The system first collects data from various sources, including internal documents (such as PDF, MD, and TXT formats), crawled web pages, structured data (JSON), and streaming data. This data is aggregated into a vector database, providing the data foundation for subsequent semantic search and analysis.

[0210] 2) Online Services

[0211] User query processing: After the user enters a query, a subquery is generated through a large language model (LLM), and the subquery enters the "deep research" stage.

[0212] In-depth research and query: In-depth research triggers SERP (search engine results page) query, obtains relevant information, and then performs "process results" operation to generate preliminary "results".

[0213] Knowledge gap determination: Use large language models to "reflect" on the results and determine whether there are knowledge gaps.

[0214] If there is a knowledge gap ("Yes, generate new query"), a new sub-query is generated and "deep research" is entered again, forming an iterative cycle to further explore information.

[0215] If there are no knowledge gaps (“No”), “Add / Rewrite and Summarize” is performed through a large language model to generate a “Final Report”.

[0216] Semantic search support: The vector database provides data support for the result processing stage through the "semantic search" function, ensuring the accuracy and comprehensiveness of the query results.

[0217] For example, the privacy-enhanced intelligent search method based on multiple iterations provided by the present invention includes the following five parts:

[0218] 1. User Input Processing and Task Decomposition

[0219] First, using hierarchical semantic parsing technology and Stanford Parser for dependency syntax analysis, we extracted core entities such as "artificial intelligence," "drug development," and "application status," and annotated them with their types (technical terms, research fields, and analysis dimensions). We then expanded the associated concepts (e.g., "drug development" associated with "target identification," "molecular docking," and "pharmacokinetics") using the knowledge graph. Using the twin-tower model, we identified these as task-based intents (accounting for 70%), triggering the task dependency graph algorithm to split the intent into three levels of subtasks:

[0220] Basic technology analysis: Sorting out the technical path of AI in drug target identification;

[0221] Existing Problem Exploration: Analysis of Algorithm Limitations and Cross-Modal Data Fusion Challenges;

[0222] Frontier Trend Forecast: Exploring the potential of transfer learning and generative models in drug design. Set the depth parameter to 3 (three iterations) and the breadth parameter to medium (focusing on core databases such as PubMed and Web of Science, as well as top CNS journals). Use a load balancing algorithm to dispatch to heterogeneous processing modules (e.g., DeepSeek-R1 for text parsing and PaddleOCR for image recognition).

[0223] 2. Multi-round Iterative Retrieval and Knowledge Completion

[0224] First round search and evidence integration

[0225] Each subtask generates a dedicated query (e.g., "AI virtual drug screening algorithm 2023-2024 literature"), and uses multi-source aggregation and anonymization technology to concurrently search 15+ academic databases to obtain PDF review papers, journal articles, and image data containing molecular structure diagrams. Unstructured data preprocessing begins:

[0226] PDF processing: Text-based PDFs are fully extracted using PyPDF2. Image-based PDFs are converted to images using pdf2image, and then recognized by PaddleOCR (e.g., correcting tilted chemical formulas based on context, with a recognition accuracy of 95%).

[0227] Cross-modal integration: Map the OCR text of the molecular structure image with the technical parameters in the patent text to construct a "drug molecule-target-algorithm" triple knowledge graph fragment.

[0228] Through semantic element coverage matrix detection, the first round of conclusions was generated: "AI is widely used in virtual drug screening, but the algorithm's versatility is limited by target diversity (evidence confidence 82%)."

[0229] Second round of iteration and gap filling:

[0230] Because the confidence level fell below the threshold (<95%) and the depth parameter was >0, the system focused on "lack of algorithm generality" and generated a new subquery: "Research on the adaptability of transfer learning in AI drug screening algorithms." This search prioritized arXiv preprints and pharmaceutical company technical reports. By integrating cross-institutional data (such as a company's internal algorithm optimization case) through federated learning, the system discovered that "meta-learning-based small sample target adaptation technology can improve algorithm generalization," supplementing the "technical principle-optimization strategy" relationship.

[0231] Three rounds of iteration and deep mining:

[0232] We further generated a subquery targeting the "transfer learning implementation bottleneck": "Technical Implementation Paths for Transfer Learning in Drug Development Scenarios," retrieving multimodal data such as clinical trial data and patent applications. By matching the CLIP model with the technology roadmap, we identified the implicit challenge of "transfer bias caused by cross-species pharmacokinetic data differences," thus completing the "application bottleneck-solution" knowledge chain.

[0233] 3. Result integration and scenario-based output:

[0234] The system integrates three rounds of evidence chains to build a traceable report:

[0235] Evidence chain annotation: Each conclusion is associated with the data source (e.g., "advantages of transfer learning" comes from the 2024 Nature Biotechnology paper, and the URL can be traced with one click);

[0236] Structured presentation: Generates Markdown reports containing technology roadmaps, algorithm comparison tables, and patent layout heat maps, supporting LaTeX formula embedding (such as molecular docking energy calculation formulas);

[0237] Gap reminder: Mark the uncovered elements in the "Frontier Trends" chapter (such as "AI computing power requirements in rare disease drug development") and provide a quick entry for "Start a new round of research."

[0238] 4. The value of technological advantages in academic scenarios:

[0239] Breakthrough in research efficiency:

[0240] Traditional manual retrieval requires reading 500+ documents one by one (a cycle of 3 months). The system uses multimodal preprocessing (OCR efficiency increased by 80%) and parallel subtask scheduling. In 3 rounds of iterations, it retrieves 200+ core documents in total and completes data integration within 5 working days, shortening the cycle by 70%.

[0241] Improved analysis depth:

[0242] With the help of the semantic element coverage matrix, the three-level association of "algorithm principle-application case-compliance risk" is automatically identified, and cross-modal contradictions that are easily overlooked by humans (such as the parameter differences between the chart data and the text description of a paper) are discovered. The academic gap of "lack of cross-modal data standardization in AI drug design" is discovered, filling the logical gap of traditional retrieval.

[0243] Innovation Accelerates Value:

[0244] Through a closed-loop optimization mechanism, high-quality data sources are recorded (such as a 50% increase in the weight of FDA technical reports), and exclusive parameter configurations are generated (the default depth in academic scenarios is 3). This helps researchers quickly capture cutting-edge directions such as "AI+antibody engineering" and "generative model optimization of synthesis routes", providing panoramic knowledge support for experimental design and promoting a 40% increase in the efficiency of converting results from cross-disciplinary research in AI and biomedicine.

[0245] 5. Hidden Protection of Privacy-Enhancing Technologies

[0246] Automatically trigger hierarchical access control during cross-institutional data retrieval:

[0247] When searching pharmaceutical company patents, user authorization is obtained through OAuth 2.0, and only decrypted data processed by federated learning is decrypted;

[0248] When outputting reports, the content involving corporate technical details is blurred (for example, "a company's core algorithm" is generalized to "the general solution of leading enterprises") to achieve a balance between knowledge integration and privacy compliance.

[0249] This process demonstrates the core value of this invention in complex academic research: through multiple rounds of semantic decomposition, iterative reasoning and multimodal fusion, fragmented information is transformed into a structured knowledge system, while privacy-enhancing technology is used to ensure data security, providing efficient and reliable intelligent assistance for scientific research and innovation.

[0250] Through technological innovation, this invention builds a complete chain of "precise analysis-secure retrieval-intelligent evolution." In academic research, it can shorten the literature research cycle from months to days, helping researchers quickly capture interdisciplinary frontiers. In the commercial field, it supports real-time dynamic decision-making, improves the efficiency of competitor analysis by 60%, and provides enterprises with competitive intelligence that ensures data security and compliance. In consumer scenarios, it enables personalized privacy-protected retrieval, reduces the risk of user information leakage, and enhances the credibility of intelligent search. This solution breaks through the bottlenecks of traditional technologies in semantic depth, security compliance, and modal fusion, promoting the transformation of intelligent search from "information listing" to "knowledge creation." It has significant technological leadership and industry application prospects.

[0251] Based on the same inventive concept, the embodiment of the present invention also provides a privacy-enhanced intelligent search system based on multiple rounds of iterations, referring to Figure 4 Shown, including:

[0252] Semantic Parsing Module: This module performs hierarchical semantic parsing on user-entered queries, extracts core semantic elements through dependency parsing and entity recognition, and generates structured semantic units. It also decomposes complex queries into a sequence of subtasks that can be executed in parallel based on a task dependency graph algorithm, and generates subqueries through dynamic priority sorting.

[0253] Multi-round reasoning engine: used to perform multiple rounds of iterative retrieval: retrieve evidence fragments from multi-source data based on subqueries, build a semantic element coverage matrix to detect knowledge gaps, and generate supplementary subqueries to iteratively complete uncovered elements until the preset termination condition is met;

[0254] Cross-modal fusion module: used to integrate cross-modal data through federated learning technology, combining semantic vector alignment and evidence fusion algorithms to generate structured knowledge graph fragments;

[0255] Privacy and security framework module: Dynamically desensitizes search results, including masking, generalization, and differential privacy, and aggregates anonymized search results through a meta-search proxy engine before outputting them.

[0256] Feedback optimization module: Optimizes retrieval strategies based on user explicit and implicit feedback to form a closed-loop iterative update mechanism.

[0257] The system depicts the complete process from data input to user feedback, as well as the interactions and data flows between various modules. This paper proposes a privacy-enhanced intelligent search system based on multiple rounds of iteration. By deeply collaborating with cognitive reasoning and privacy protection frameworks, it builds a full-link intelligent search system covering "demand analysis - multiple rounds of reasoning - secure retrieval - result output."

[0258] The system architecture is as follows: starting from user input, it generates executable subtasks through hierarchical semantic analysis, and achieves deep knowledge mining through iterative reasoning chains. At the same time, it incorporates privacy-enhancing technologies such as multi-source aggregation anonymization and hierarchical permission control, ultimately providing accurate and secure search results for academic, commercial, consumer and other scenarios, and forming a closed-loop optimization through user feedback.

[0259] 1. User input processing: Accurate mapping from natural language to structured tasks;

[0260] The system first performs multi-dimensional semantic analysis of user queries. Through word segmentation, part-of-speech tagging, and dependency syntax analysis, it extracts core entities (such as technical terms, time, and region) and constructs grammatical dependencies to form structured semantic units. For example, it breaks down the query "Analyze the AI ​​patent landscape in the medical field in 2024" into entities such as "AI," "Medical field," and "2024," along with the core intent of "Patent landscape analysis." It then integrates knowledge graphs to expand entity relationships (e.g., "Medical field" is associated with "Disease Diagnosis" and "Drug Development") and utilizes a twin-tower model to distinguish between task-based intents (70% of the time, aiming to generate structured outcomes such as reports and decision-making recommendations) and conversational intents (30% of the time, aiming to explain and interact with knowledge). For task-based intents, a task dependency graph algorithm automatically splits the query into parallelizable subtasks (e.g., "Patent landscape analysis" is split into "Patent data retrieval," "Technical efficacy comparison," and "Legal risk assessment"). Task triples are generated, containing the execution subject, input, and output. Computing resources are then allocated based on complexity, enabling the structured transformation of complex requirements.

[0261] 2. Multi-round reasoning retrieval: Dynamic iteration enables deep knowledge completion;

[0262] The system builds an iterative reasoning chain to deal with complex problems. First, each subtask generates a dedicated query, retrieves similar documents through the vector database, and retains high-value evidence fragments after the large model credibility score. Knowledge gaps are detected based on the semantic element coverage matrix (entities, relationships, constraints). If uncovered elements are found (such as the lack of "application case" association with "technical principles"), a supplementary search is triggered and a new subquery is generated, forming a "retrieval-verification-completion" cycle of reasoning. During the iteration process, a double termination condition is set (the confidence level of evidence meets the standard or the maximum number of rounds is reached), and the attention mechanism is used to focus on unresolved knowledge gaps and gradually expand the depth of the reasoning chain. At the same time, the system supports multimodal data processing, extracts image text and parses PDF structured information through OCR technology, converts non-text data into computable semantic vectors, and realizes the integrated analysis of multimodal information such as text, images, and PDFs, breaking down modal barriers and improving the comprehensiveness of retrieval.

[0263] 3. Privacy and security framework: Full-link protection ensures user information security;

[0264] To address data security challenges, the system has designed a multi-layered privacy protection mechanism. At the input layer, user requests are routed through a meta-search proxy engine, split into sub-requests, and sent to multiple search engines (such as DuckDuckGo and Bing). Results are aggregated and deduplicated to avoid directly exposing user IP addresses and search behavior. An anonymous session ID is generated based on device fingerprints and random keys to ensure identity anonymity. At the processing layer, a privacy information classification model is constructed, categorizing the information into three security levels based on its ubiquity and sensitivity. A role-based access control (RBAC) model is used to implement fine-grained access control (e.g., regular users can only access low-sensitivity data). Federated learning technology is used for cross-institutional collaboration. Only parameters of locally trained models are uploaded to avoid cross-domain transmission of raw data. Sensitive operations are stored on the blockchain to ensure traceability. At the output layer, sensitive information such as ID numbers and geographic locations is masked (e.g., hiding intermediate fields) and generalized (e.g., simplifying "Chaoyang District, Beijing" to "Beijing"), and noise perturbations are added to statistical data to comprehensively mitigate the risk of privacy leaks.

[0265] 4. Result integration and scenario-based output: Generate high-value search results on demand;

[0266] When the system integrates the results of multiple rounds of reasoning, it builds a traceable chain of evidence for each conclusion, records the data source URL, retrieval time and processing logic, and supports one-click traceability verification by users to ensure the credibility and compliance of the results. Differentiated outputs are generated for different scenarios: academic scenarios generate structured reports (such as Markdown format, supporting LaTeX embedding) that include research methods, data sources, charts and formulas; business scenarios output PPT materials that focus on core conclusions, risk points and decision-making recommendations, and embed data drill-down functions to view the underlying data sources; consumer scenarios provide visual comparison lists (parameter tables, radar charts, etc.), adapted to mobile sliding viewing and voice summary playback, to meet the needs of lightweight information acquisition. In addition, the system dynamically adjusts the retrieval strategy and parameter configuration by collecting user feedback on explicit corrections (such as supplementary data sources) and implicit behaviors (such as the length of time content is retained), forming a closed loop of "demand processing-result optimization-performance improvement" to continuously adapt to diverse user needs.

[0267] This invention breaks through the technical bottlenecks of traditional intelligent search, achieving automated decomposition and knowledge completion of complex problems through a multi-round iterative reasoning architecture, significantly improving key information coverage. The privacy-enhancing technology system establishes security protection covering the entire data lifecycle, meeting the compliance requirements of sensitive scenarios such as healthcare and finance. Cross-modal fusion capabilities break down information barriers and improve the efficiency of integrated analysis of multi-type data. A closed-loop optimization mechanism ensures that system performance continues to improve with use. This solution provides an efficient and secure technical path for interdisciplinary knowledge integration in academic research, real-time data support for business decision-making, and personalized recommendations in consumer scenarios, pushing intelligent search towards deep semantic understanding and privacy security.

[0268] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0269] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A privacy-enhanced intelligent search method based on multiple rounds of iteration, characterized in that: The following steps are involved: S10, performing hierarchical semantic analysis on the query input by the user, extracting core semantic elements through dependency syntax analysis and entity recognition, and generating structured semantic units; S20, based on the task dependency graph algorithm, decomposes complex queries into a sequence of subtasks that can be executed in parallel, and generates subqueries through dynamic priority sorting; S30, performing multiple rounds of iterative retrieval: retrieving evidence fragments from multi-source data based on subqueries, constructing a semantic element coverage matrix to detect knowledge gaps, and generating supplementary subqueries to iteratively complete uncovered elements if any, until a preset termination condition is met; S40: Integrate cross-modal data through federated learning technology, combine semantic vector alignment and evidence fusion algorithms to generate structured knowledge graph fragments; S50, dynamically desensitizing the search results, including masking, generalization, and differential privacy, and aggregating the anonymized search results through a meta-search proxy engine and outputting the results; S60. Optimize the retrieval strategy based on user explicit and implicit feedback to form a closed-loop iterative update mechanism.

2. The privacy-enhanced intelligent search method based on multiple iterations according to claim 1, characterized in that: The step S10 specifically includes: S101. Using the THULAC word segmentation system, which is based on a word segmentation algorithm combining a hidden Markov model and a conditional random field, to perform word segmentation and part-of-speech tagging on the query input by the user; S102, using a syntactic analysis algorithm to identify dependencies between words and generate a dependency tree; S103, using the BiLSTM-CRF model to perform entity type recognition, converting the segmented words into word vectors, and converting the part-of-speech tags into one-hot encodings, and concatenating them to form an input feature vector; S104. Expand the associated entity set in combination with the knowledge graph.

3. The privacy-enhanced intelligent search method based on multiple iterations according to claim 2, characterized in that: The step S20 specifically includes: S201. Build a disassembly engine based on production rules, where the rule format is IF <syntactic structure> AND <semantic role> THEN <subquery generation template>; S202. Calculate the subquery priority using a multi-attribute decision-making algorithm. The formula is: Where α = 0.6, β = 0.4 are weight coefficients, TF-IDF(q) is the term frequency of the subquery keyword minus the inverse document frequency, and Sim(q, Q) is the cosine similarity between the subquery and the original query; S203: Generate a subquery template according to the predicate argument structure in the dependency tree.

4. The privacy-enhanced intelligent search method based on multiple iterations according to claim 2, characterized in that: The step S30 specifically includes: S301: Construct a three-dimensional semantic element coverage matrix, where the dimensions include entity dimension, relationship dimension, and constraint dimension, and each dimension is assigned a coverage status flag ranging from 0 to 1. S302, the gap detection threshold is When GapScore>0.3, a cross-modal completion query is generated; S303. Set the double iteration termination condition: the average credibility of all evidence is greater than or equal to the threshold, or the number of iterations reaches the upper limit of N rounds.

5. The privacy-enhanced intelligent search method based on multiple iterations according to claim 2, characterized in that: The step S40 specifically includes: S401, image completion uses the CLIP model to convert text gaps into visual semantic vectors, calculate cosine similarity with image feature vectors, and retrieve image data with similarity > 0.7; S402: Using the Whisper model to transcribe multilingual audio and then perform semantic matching; S403. Use DS evidence theory to fuse multi-source evidence. The basic probability distribution function satisfies: Among them, m(A) represents the confidence of the fused evidence on proposition A; A1,A2......A n : represent the 1st to nth evidence bodies respectively; A1∩...∩A n =A: means that the intersection of all the evidences supporting the proposition is equal to A, that is, these evidences jointly support proposition A; The i-th body of evidence for proposition A i The product of the basic probability distribution of , represents the strength of the consistent evidence that multiple evidences jointly support A.

6. The privacy-enhanced intelligent search method based on multiple iterations according to claim 1, characterized in that: The desensitization process in step S50 is as follows: The masking process retains the first 3 digits and the last 4 digits of the mobile phone number, and fills the middle with 4 asterisks; Geographical location is generalized to city-level administrative districts; Differential privacy adds Laplace noise, the formula is: Among them, e is the privacy budget parameter, which controls the noise intensity; SimHash deduplication and session anonymous identifier UUID v4 are used in the proxy engine to aggregate multi-source search results.

7. The privacy-enhanced intelligent search method based on multiple iterations according to claim 6, characterized in that: The meta-search proxy engine in step S50 is used to split the user request into sub-requests and send them to multiple search engines, using the SimHash algorithm to remove duplicates and generate session-level anonymous identifiers: AnonID=HMAC(SALT,UUIDv4()+DeviceFinger print) Among them, SALT is the server-side random key; DeviceFingerprint is the device feature hash value, which includes the SHA-256 summary of User-Agent, screen resolution, and IP address.

8. The privacy-enhanced intelligent search method based on multiple iterations according to claim 1, characterized in that: In step S60: Explicit feedback includes user-annotated data source credibility, triggering a corresponding weight increase; Implicit feedback includes automatically generating supplementary analysis suggestions when the content is retained for more than T minutes.

9. A privacy-enhanced intelligent search system based on multiple iterations, characterized in that: Using the privacy-enhanced intelligent search method based on multiple iterations according to any one of claims 1 to 8, the system includes: Semantic Parsing Module: This module performs hierarchical semantic parsing on user-entered queries, extracts core semantic elements through dependency parsing and entity recognition, and generates structured semantic units. It also decomposes complex queries into a sequence of subtasks that can be executed in parallel based on a task dependency graph algorithm, and generates subqueries through dynamic priority sorting. Multi-round reasoning engine: used to perform multiple rounds of iterative retrieval: retrieve evidence fragments from multi-source data based on subqueries, build a semantic element coverage matrix to detect knowledge gaps, and generate supplementary subqueries to iteratively complete uncovered elements until the preset termination condition is met; Cross-modal fusion module: used to integrate cross-modal data through federated learning technology, combining semantic vector alignment and evidence fusion algorithms to generate structured knowledge graph fragments; Privacy and security framework module: Dynamically desensitizes search results, including masking, generalization, and differential privacy, and aggregates anonymized search results through a meta-search proxy engine before outputting them. Feedback optimization module: Optimizes retrieval strategies based on user explicit and implicit feedback to form a closed-loop iterative update mechanism.

Images