Cross-domain reasoning method based on federated learning

Abstract

The invention relates to the technical field of big data, provides a federated learning-based cross-domain inference method, is applied to financial and medical health care service scenes, can realize cross-domain data logic alignment based on a lightweight semantic index mechanism, and gets rid of ETL dependency. Encryption processing is carried out by adopting a multi-layer privacy protection mechanism, so that the data security is effectively improved; random disturbance is added to the gradient ciphertext which does not pass sensitivity detection, and verifiable privacy constraint is achieved; realizing targeted configuration of an aggregation strategy according to different node attributes based on a dynamic node weight optimization algorithm; model aggregation is carried out by adopting a multi-layer semantic oligomerization algorithm, so that the adaptability of a global model at an edge node is improved; cross-domain knowledge fusion is carried out by adopting a cross-domain knowledge distillation mechanism, so that the generalization ability of the global model is enhanced; and performing targeted parameter fine adjustment on the initial model of each edge node so as to meet local personalized business requirements of the node, thereby assisting local execution of intelligent reasoning at each edge node.

Description

Technical Field

[0001] This invention relates to the field of big data technology, and in particular to a cross-domain reasoning method based on federated learning. Background Technology

[0002] With the continuous development of artificial intelligence, data security is receiving increasing attention.

[0003] In existing technologies, a centralized modeling approach using ETL (Extract, Transform, Load) is typically employed to aggregate data from various institutions or nodes to a central server. However, this approach has the following drawbacks: (1) High privacy and compliance risks: Data from various institutions (such as user data of banks, patient information of hospitals, etc.) must be desensitized and transmitted to the central warehouse. There is still a risk of reverse re-identification, which cannot meet the compliance requirement of "data not leaving the domain".

[0004] (2) High training and inference latency: Due to cross-node transmission and centralized training, the model update cycle is long and the real-time performance is poor, making it unsuitable for high-frequency business needs (such as real-time financial monitoring and IoT (Internet of Things) streaming inference).

[0005] (3) Insufficient model generalization ability: Significant differences exist in the data distribution of each node (such as the Non-IID (Non-Independently and Identically Distributed) problem), resulting in the centralized model performing poorly on some edge nodes.

[0006] (4) Lack of dynamic collaborative optimization mechanism: The fixed-weighted aggregation method does not take into account multiple factors such as node bandwidth, latency, and data quality, resulting in insufficient performance optimization.

[0007] Therefore, there is an urgent need for a new collaborative learning framework that can achieve intelligent reasoning and modeling across institutions and domains without performing data extraction (Zero-ETL (Zero-Extract, Transform, Load)), which can both ensure privacy and security and improve model convergence speed and generalization performance. Summary of the Invention

[0008] In view of the above, it is necessary to provide a cross-domain reasoning method based on federated learning, which aims to solve the problem of not being able to achieve highly secure intelligent reasoning across institutions and domains without data extraction.

[0009] A cross-domain reasoning method based on federated learning, the method comprising: In response to cross-domain inference commands triggered by different business domains corresponding to each edge node, business data and system operation data are collected at each edge node as local data, and the local data collected at each edge node is preprocessed based on a lightweight semantic indexing mechanism to obtain the data to be processed at each edge node. The model is trained at each edge node using the data to be processed at each edge node, and the local model, local model parameters and local model gradient of each edge node are obtained. A multi-layered privacy protection mechanism is used to encrypt the local model parameters and local model gradients of each edge node, resulting in ciphertext of the parameters and gradients of each edge node. Sensitivity detection is performed on the gradient ciphertext of each edge node, and random perturbation is added to the gradient ciphertext that fails the sensitivity detection to update the gradient ciphertext; Collect multidimensional state data of each edge node, and generate an aggregation strategy based on the dynamic node weight optimization algorithm and the multidimensional state data; In the cloud, a multi-layer semantic alignment aggregation algorithm is used to aggregate models based on the aggregation strategy, the parameter ciphertext, and the updated gradient ciphertext to obtain an initial global model. At each edge node, a cross-domain knowledge distillation mechanism is used to perform cross-domain knowledge fusion based on the initial global model to obtain the initial model of each edge node; In the cloud, the parameters of the initial model of each edge node are fine-tuned based on the data to be processed at each edge node to obtain the sub-model of each edge node. The sub-models of each edge node are used to execute inference tasks locally on each edge node, and the inference results are sent to the trigger of the corresponding inference task.

[0010] A cross-domain inference device based on federated learning, the device comprising: The preprocessing unit is used to respond to cross-domain inference instructions triggered based on different business domains corresponding to each edge node, collect business data and system operation data at each edge node as local data, and preprocess the local data collected by each edge node based on a lightweight semantic indexing mechanism to obtain the data to be processed at each edge node. The training unit is used to train the model at each edge node using the data to be processed at each edge node, and to obtain the local model, local model parameters and local model gradient of each edge node. The encryption unit is used to encrypt the local model parameters and local model gradients of each edge node using a multi-layer privacy protection mechanism, so as to obtain the parameter ciphertext and gradient ciphertext of each edge node. The update unit is used to perform sensitivity detection on the gradient ciphertext of each edge node, and add random perturbation to the gradient ciphertext that fails the sensitivity detection to update the gradient ciphertext; The generation unit is used to collect multidimensional state data of each edge node and generate an aggregation strategy based on the dynamic node weight optimization algorithm and the multidimensional state data. An aggregation unit is used to perform model aggregation in the cloud using a multi-layer semantic alignment aggregation algorithm based on the aggregation strategy, the parameter ciphertext, and the updated gradient ciphertext to obtain an initial global model. The fusion unit is used to perform cross-domain knowledge fusion based on the initial global model at each edge node using a cross-domain knowledge distillation mechanism to obtain the initial model of each edge node. The fine-tuning unit is used in the cloud to fine-tune the parameters of the initial model of each edge node based on the data to be processed at each edge node, so as to obtain the sub-model of each edge node. The inference unit is used to perform inference tasks locally on each edge node using the sub-models of each edge node, and send the obtained inference results to the triggerer of the corresponding inference task.

[0011] A computer device, the computer device comprising: A memory for storing at least one instruction; and a processor for executing the instructions stored in the memory to implement the federated learning-based cross-domain inference method.

[0012] A computer-readable storage medium storing at least one instruction that is executed by a processor in a computer device to implement the federated learning-based cross-domain inference method.

[0013] As can be seen from the above technical solutions, this invention can preprocess the local data collected by each edge node based on a lightweight semantic indexing mechanism, achieving cross-domain data logical alignment rather than physical migration, thus eliminating ETL dependence; it employs a multi-layer privacy protection mechanism to encrypt the local model parameters and gradients of each edge node, effectively improving data security; it adds random perturbations to the gradient ciphertext that fails sensitivity detection, achieving verifiable privacy constraints; based on a dynamic node weight optimization algorithm and multi-dimensional state data, it implements targeted configuration of aggregation strategies according to different node attributes; it uses a multi-layer semantic alignment aggregation algorithm for model aggregation, improving the adaptability of the global model on edge nodes; it employs a cross-domain knowledge distillation mechanism for cross-domain knowledge fusion, strengthening the generalization ability of the global model; and it fine-tunes the parameters of the initial model of each edge node based on the data to be processed at each edge node, meeting the personalized business needs of the node's locality, thereby assisting in the execution of intelligent inference locally at each edge node. Attached Figure Description

[0014] Figure 1 This is a flowchart of a preferred embodiment of the cross-domain reasoning method based on federated learning of the present invention.

[0015] Figure 2 This is a functional block diagram of a preferred embodiment of the cross-domain inference device based on federated learning of the present invention.

[0016] Figure 3 This is a schematic diagram of the structure of a computer device that implements a preferred embodiment of the cross-domain reasoning method based on federated learning according to the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1 The diagram shown is a flowchart of a preferred embodiment of the cross-domain inference method based on federated learning of the present invention. The order of the steps in this flowchart can be changed, and some steps can be omitted, depending on different requirements.

[0019] The federated learning-based cross-domain inference method is applied to one or more computer devices. The computer device is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions. Its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0020] The computer device can be any electronic product that can interact with the user, such as a personal computer, tablet computer, smartphone, personal digital assistant (PDA), game console, interactive network television (IPTV), smart wearable device, etc.

[0021] The computer equipment may also include network equipment and / or user equipment. The network equipment includes, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of hosts or network servers.

[0022] The server can be a standalone server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0023] Artificial intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0024] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0025] The network in which the computer device is located includes, but is not limited to, the Internet, wide area network, metropolitan area network, local area network, and virtual private network (VPN).

[0026] S10, in response to cross-domain inference instructions triggered by different business domains corresponding to each edge node, collects business data and system operation data at each edge node as local data, and preprocesses the local data collected by each edge node based on a lightweight semantic indexing mechanism to obtain the data to be processed at each edge node.

[0027] In this embodiment, the local data may include various business data and system operation data of each edge node.

[0028] In this embodiment, the preprocessing of local data collected from each edge node based on the lightweight semantic indexing mechanism to obtain the data to be processed for each edge node includes: For each edge node, obtain the structured data and unstructured data of the edge node; The structured data is converted into a business-related first feature vector using the lightweight feature extractor built into the edge node; The unstructured data is transformed into a second feature vector of a specified dimension using a pre-trained lightweight model; Define a semantic dimension and construct a semantic mapping dictionary based on the semantic dimension; wherein the semantic mapping dictionary is used to reflect the mapping relationship between the semantic dimension and heterogeneous features; The first feature vector and the second feature vector are mapped to a unified semantic space according to the semantic mapping dictionary to obtain the data to be processed for the edge node.

[0029] The structured data may include transaction volumes in the financial sector and testing indicators in the medical sector.

[0030] The unstructured data may include user behavior text in the financial field, medical images in the medical field, etc.

[0031] The first feature vector may include a monthly average repayment ability vector, a disease feature vector, etc.

[0032] The lightweight model may include a lightweight convolutional neural network model, etc.

[0033] The second feature vector may include a fixed-dimensional feature vector obtained by transforming unstructured data such as financial user behavior text and medical images.

[0034] The semantic mapping dictionary can store the mapping relationship between the transaction anomaly degree dimension and heterogeneous features such as "cross-border transaction records", and the mapping relationship between the pathological feature dimension and heterogeneous features such as "pathological slide features".

[0035] Through the above embodiments, a unified semantic mapping can be established for heterogeneous data sources (such as transaction and wealth management system data of banks, and medical record and imaging system data of hospitals) at various edge nodes. This maps feature vectors of different formats and dimensions to the same semantic space (such as "financial risk semantic layer" and "clinical diagnosis semantic layer"), achieving semantic logical alignment of data rather than physical migration. This completely eliminates the dependence on traditional ETL (Extract, Transform, Load) and avoids the risk of privacy leakage during the original data transmission process. Through a lightweight semantic indexing mechanism, a unified semantic layer structure is established for heterogeneous data sources, solving the problem of data format and dimension differences between different nodes and laying a data foundation for subsequent collaborative modeling. By extracting only local feature vectors and data statistical summaries, the amount of data transmission is reduced, and the overall system processing efficiency is improved.

[0036] S11: Use the data to be processed at each edge node to train the model at each edge node, and obtain the local model, local model parameters and local model gradient of each edge node.

[0037] In this embodiment, each edge node completes its own model training locally, and the data does not need to leave the domain.

[0038] S12 employs a multi-layered privacy protection mechanism to encrypt the local model parameters and local model gradients of each edge node, resulting in ciphertext of the parameters and gradients of each edge node.

[0039] In this embodiment, to ensure data security, the local model parameters and local model gradients of each edge node need to be encrypted.

[0040] Specifically, the method of employing a multi-layered privacy protection mechanism to encrypt the local model parameters and local model gradients of each edge node, resulting in the following ciphertexts of the parameters and gradients of each edge node: By adding controllable random noise to the local model gradient of each edge node using the Differential Privacy (DP) algorithm, the desensitized model gradient of each edge node is obtained. The desensitized model gradient and local model parameters of each edge node are homomorphically encrypted using the homomorphic encryption (HE) algorithm to obtain the encrypted gradient and parameter ciphertext of each edge node. The Secure Multi-Party Computation (MPC) algorithm is used to split the encryption gradient of each edge node into multiple fragments to obtain the gradient ciphertext of each edge node.

[0041] The differential privacy algorithm enables gradient-sensitive desensitization, preventing gradient back-inference of the original data.

[0042] The homomorphic encryption algorithm ensures that the gradient is "computable in ciphertext and invisible in plaintext" during transmission and aggregation.

[0043] The secure multi-party computation algorithm can distribute fragments to designated nodes in the cloud, thus preventing a single node from possessing the complete gradient.

[0044] Through the above embodiments, a triple protection mechanism can be formed based on multiple layers of privacy protection to ensure privacy and security throughout the entire process from data processing and transmission to computation.

[0045] Furthermore, encrypted gradients and parameters can be uploaded only to the central node or the cloud to avoid the leakage of sensitive information during model training.

[0046] S13, perform sensitivity detection on the gradient ciphertext of each edge node, and add random perturbation to the gradient ciphertext that fails the sensitivity detection to update the gradient ciphertext.

[0047] In this embodiment, the step of performing sensitivity detection on the gradient ciphertext of each edge node and adding random perturbations to the gradient ciphertext that fails the sensitivity detection to update the gradient ciphertext includes: The sensitive feature recognition rule base is invoked; wherein, the sensitive feature recognition rule base is used to store the correspondence between sensitive features, deviation thresholds, and disturbance noise information; The gradient ciphertext of each edge node is matched against the sensitive feature recognition rule base; The matched gradient ciphertext is identified as a gradient ciphertext that fails the sensitivity detection, and random perturbation is added to the gradient ciphertext that fails the sensitivity detection based on the matched perturbation noise information to update the gradient ciphertext.

[0048] For example, the sensitive feature recognition rule base may include: in the financial scenario, "user income gradient fluctuation > 50%, corresponding to Gaussian noise with an intensity of 0.5", and in the medical scenario, "tumor marker gradient deviates from the mean by 3 times, corresponding to Laplace noise with an intensity of 1".

[0049] In the above embodiments, privacy constraints can be verified by gradient sensitivity control and random perturbation to prevent malicious nodes from back-engineering the original data through gradients.

[0050] S14. Collect multidimensional state data of each edge node, and generate an aggregation strategy based on the Adaptive Federated Weighting (AFW) algorithm and the multidimensional state data.

[0051] In this embodiment, the multidimensional state data may include: 1) Network dimension: bandwidth and latency of node upload gradients; 2) Data dimension: Distributional dissimilarity of node data (deviation from the global data distribution, i.e., the degree of Non-IID (Non-Independently and Identically Distributed)). 3) Computational dimensions: CPU (Central Processing Unit) / GPU (Graphics Processing Unit) computing power and memory usage of the node; 4) Historical dimension: the contribution of nodes to the model in the past (such as the contribution rate of gradients to the improvement of global model accuracy). 5) Task dimension: high-frequency real-time task nodes (such as financial real-time abnormal behavior detection tasks, medical emergency diagnosis tasks, etc.) and high-precision requirement task nodes (such as financial risk deep modeling tasks, rare disease diagnosis model modeling tasks, etc.).

[0052] In this embodiment, the step of collecting multidimensional state data from each edge node and generating an aggregation strategy based on a dynamic node weight optimization algorithm and the multidimensional state data includes: Collect multi-dimensional status data of each edge node according to network dimension, data dimension, computing dimension, historical dimension, and task dimension; Identify strong and weak nodes in each edge node based on the multidimensional state data of each edge node; Aggregate weights are assigned based on the strong and weak nodes among the edge nodes; wherein the weight of the strong nodes is higher than that of the weak nodes. Identify high-frequency real-time task nodes and high-precision-requirement task nodes in each edge node based on the multi-dimensional state data of each edge node. Configure the high-frequency real-time task node to asynchronous aggregation mode, and configure the high-precision requirement task node to synchronous aggregation mode.

[0053] The strong nodes can include nodes with high bandwidth, low latency, uniform data distribution, and large historical contributions, which need to be assigned higher aggregation weights; the weak nodes can include organizations in remote areas, such as rural banks and community hospitals, which need to be assigned basic weights and have their gradient aggregation priority reduced, so as to avoid slowing down the overall progress.

[0054] Specifically, an asynchronous aggregation mode is configured for the high-frequency real-time task nodes, so that the high-frequency real-time task nodes do not need to wait for all nodes, thus reducing waiting latency.

[0055] Specifically, configuring a synchronous aggregation mode for the high-precision task nodes allows for unified aggregation after all node gradients have been uploaded, thereby ensuring model accuracy.

[0056] In the above embodiments, the dynamic node weight optimization algorithm can allocate aggregation weights according to the multi-dimensional state of nodes, avoiding low-quality nodes (such as low data volume and high latency) from dragging down the global model performance and improving aggregation efficiency; the automatic switching between synchronous and asynchronous aggregation strategies solves the global waiting problem caused by high latency of some nodes, thereby reducing the model update cycle.

[0057] S15. In the cloud, a multi-layer semantic-aligned aggregation (SAA) algorithm is used to aggregate models based on the aggregation strategy, the parameter ciphertext, and the updated gradient ciphertext to obtain an initial global model.

[0058] In this embodiment, it is also necessary to address the issue of data distribution differences across multiple domains.

[0059] Specifically, in the cloud, the process of using a multi-layer semantic alignment aggregation algorithm to aggregate models based on the aggregation strategy, the parameter ciphertext, and the updated gradient ciphertext to obtain an initial global model includes: The multiple fragments are restored using the secure multi-party computation algorithm to obtain the encryption gradient of each edge node; The encryption gradients of each edge node are semantically aligned according to the semantic mapping dictionary. According to the aggregation strategy, the aligned encrypted gradients are weighted and aggregated based on the parameter ciphertext to obtain the initial global model.

[0060] In the above embodiments, the multi-layer semantic alignment aggregation algorithm can specifically solve the Non-IID problem (i.e., the problem of data distribution differences among nodes), and improve the adaptability of the global model to edge nodes.

[0061] In this embodiment, after obtaining an initial global model by performing model aggregation using a multi-layer semantic alignment aggregation algorithm based on the aggregation strategy, the parameter ciphertext, and the updated gradient ciphertext in the cloud, the method further includes: The aggregation strategy is adjusted using a reinforcement learning (RL) mechanism.

[0062] For example, a reward function can be configured, and the aggregation strategy can be updated based on the reward function after aggregation to improve the aggregation strategy.

[0063] In the above embodiments, latency perception and energy consumption balance can be achieved through the self-optimization mechanism of reinforcement learning, thereby adapting to the unstable network and limited computing power of edge nodes (such as financial branch servers and medical IoT devices).

[0064] S16. At each edge node, a cross-domain knowledge fusion mechanism is used based on the initial global model to obtain the initial model of each edge node.

[0065] In this embodiment, it is also necessary to improve the generalization performance of the model.

[0066] Specifically, at each edge node, the method of using a cross-domain knowledge distillation mechanism to perform cross-domain knowledge fusion based on the initial global model to obtain the initial model for each edge node includes: The initial global model is used as the teacher model, and the knowledge of the teacher model is transformed into an encrypted knowledge graph; Using the local model of each edge node as the student model, and based on the data to be processed at each edge node, each student model is guided to learn the knowledge of the teacher model according to the encrypted knowledge graph, thus obtaining the initial model of each edge node.

[0067] The knowledge in the teacher model may include decision boundaries, feature importance, etc.

[0068] Through the above embodiments, the "local knowledge" of different nodes can be fused based on the cross-domain knowledge distillation mechanism, thereby enhancing the generalization ability of the global model.

[0069] S17, in the cloud, the parameters of the initial model of each edge node are fine-tuned based on the data to be processed at each edge node to obtain the sub-model of each edge node.

[0070] In this embodiment, in order to make the model of each edge node more suitable for the corresponding business needs, parameter fine-tuning is also required.

[0071] Specifically, in the cloud, the initial model of each edge node is fine-tuned based on the data to be processed at each edge node to obtain the sub-model of each edge node, including: Extract business requirement data from the data to be processed at each edge node; Based on the business requirements data of each edge node, the parameters of the initial model of each edge node are fine-tuned to obtain the sub-model of each edge node.

[0072] For example, for bank nodes where micro and small enterprises are concentrated, the ability to identify risks of micro and small enterprises can be strengthened; for pediatric hospital nodes, the ability to diagnose rare diseases in children can be strengthened.

[0073] Through the above embodiments, the local business needs of each edge node can be met by fine-tuning, thereby achieving the dual goals of global collaboration and local optimization.

[0074] S18: Utilize the sub-models of each edge node to execute inference tasks locally on each edge node, and send the obtained inference results to the triggerer of the corresponding inference task.

[0075] For example, financial and banking nodes can be used to infer the risks of abnormal transactions; medical diagnostic nodes can be used to infer the risks of diseases.

[0076] In this embodiment, after the method of using the sub-models of each edge node to perform inference tasks locally on each edge node and sending the obtained inference results to the triggerer of the corresponding inference task, the method further includes: A drift detection algorithm is used to detect the feature differences between the data features of each edge node and the feature differences between the data to be processed at each edge node in real time; when the feature difference corresponding to the first edge node is greater than the difference threshold, it is determined that the first edge node has experienced data distribution drift, and a retraining task is triggered; and / or The model performance of each edge node's sub-model is detected at preset time intervals; when the model performance of a second edge node's sub-model decreases by more than a preset threshold, the retraining task is triggered.

[0077] The preset time interval, the difference threshold, and the preset threshold can be optimal values ​​selected based on the experiment.

[0078] Through the above embodiments, regular performance evaluation and drift detection can be achieved, timely detection of performance degradation caused by changes in data distribution can be made, and model failure can be avoided; automatic model retraining can also ensure that the model is always adapted to the latest business data, thereby improving overall performance.

[0079] As can be seen from the above technical solutions, this invention can preprocess the local data collected by each edge node based on a lightweight semantic indexing mechanism, achieving cross-domain data logical alignment rather than physical migration, thus eliminating ETL dependence; it employs a multi-layer privacy protection mechanism to encrypt the local model parameters and gradients of each edge node, effectively improving data security; it adds random perturbations to the gradient ciphertext that fails sensitivity detection, achieving verifiable privacy constraints; based on a dynamic node weight optimization algorithm and multi-dimensional state data, it implements targeted configuration of aggregation strategies according to different node attributes; it uses a multi-layer semantic alignment aggregation algorithm for model aggregation, improving the adaptability of the global model on edge nodes; it employs a cross-domain knowledge distillation mechanism for cross-domain knowledge fusion, strengthening the generalization ability of the global model; and it fine-tunes the parameters of the initial model of each edge node based on the data to be processed at each edge node, meeting the personalized business needs of the node's locality, thereby assisting in the execution of intelligent inference locally at each edge node.

[0080] like Figure 2 The diagram shown is a functional block diagram of a preferred embodiment of the cross-domain inference device based on federated learning of the present invention. The cross-domain inference device 11 based on federated learning includes a preprocessing unit 110, a training unit 111, an encryption unit 112, an update unit 113, a generation unit 114, an aggregation unit 115, a fusion unit 116, a fine-tuning unit 117, and an inference unit 118. The module / unit referred to in this invention is a series of computer program segments that can be executed by a processor and perform a fixed function, stored in memory. In this embodiment, the functions of each module / unit will be described in detail in subsequent embodiments.

[0081] The preprocessing unit 110 is used to respond to cross-domain inference instructions triggered based on different business domains corresponding to each edge node, collect business data and system operation data at each edge node as local data, and preprocess the local data collected at each edge node based on a lightweight semantic indexing mechanism to obtain the data to be processed at each edge node.

[0082] In this embodiment, the local data may include various business data and system operation data of each edge node.

[0083] In this embodiment, the preprocessing unit 110 preprocesses the local data collected by each edge node based on a lightweight semantic indexing mechanism to obtain the data to be processed for each edge node, including: For each edge node, obtain the structured data and unstructured data of the edge node; The structured data is converted into a business-related first feature vector using the lightweight feature extractor built into the edge node; The unstructured data is transformed into a second feature vector of a specified dimension using a pre-trained lightweight model; Define a semantic dimension and construct a semantic mapping dictionary based on the semantic dimension; wherein the semantic mapping dictionary is used to reflect the mapping relationship between the semantic dimension and heterogeneous features; The first feature vector and the second feature vector are mapped to a unified semantic space according to the semantic mapping dictionary to obtain the data to be processed for the edge node.

[0084] The structured data may include transaction volumes in the financial sector and testing indicators in the medical sector.

[0085] The unstructured data may include user behavior text in the financial field, medical images in the medical field, etc.

[0086] The first feature vector may include a monthly average repayment ability vector, a disease feature vector, etc.

[0087] The lightweight model may include a lightweight convolutional neural network model, etc.

[0088] The second feature vector may include a fixed-dimensional feature vector obtained by transforming unstructured data such as financial user behavior text and medical images.

[0089] The semantic mapping dictionary can store the mapping relationship between the transaction anomaly degree dimension and heterogeneous features such as "cross-border transaction records", and the mapping relationship between the pathological feature dimension and heterogeneous features such as "pathological slide features".

[0090] Through the above embodiments, a unified semantic mapping can be established for heterogeneous data sources (such as transaction and wealth management system data of banks, and medical record and imaging system data of hospitals) at various edge nodes. This maps feature vectors of different formats and dimensions to the same semantic space (such as "financial risk semantic layer" and "clinical diagnosis semantic layer"), achieving semantic logical alignment of data rather than physical migration. This completely eliminates the dependence on traditional ETL (Extract, Transform, Load) and avoids the risk of privacy leakage during the original data transmission process. Through a lightweight semantic indexing mechanism, a unified semantic layer structure is established for heterogeneous data sources, solving the problem of data format and dimension differences between different nodes and laying a data foundation for subsequent collaborative modeling. By extracting only local feature vectors and data statistical summaries, the amount of data transmission is reduced, and the overall system processing efficiency is improved.

[0091] The training unit 111 is used to train the model at each edge node using the data to be processed at each edge node, so as to obtain the local model, local model parameters and local model gradient of each edge node.

[0092] In this embodiment, each edge node completes its own model training locally, and the data does not need to leave the domain.

[0093] The encryption unit 112 is used to encrypt the local model parameters and local model gradients of each edge node using a multi-layer privacy protection mechanism to obtain the parameter ciphertext and gradient ciphertext of each edge node.

[0094] In this embodiment, to ensure data security, the local model parameters and local model gradients of each edge node need to be encrypted.

[0095] Specifically, the encryption unit 112 employs a multi-layer privacy protection mechanism to encrypt the local model parameters and local model gradients of each edge node, obtaining the parameter ciphertext and gradient ciphertext of each edge node, including: By adding controllable random noise to the local model gradient of each edge node using the Differential Privacy (DP) algorithm, the desensitized model gradient of each edge node is obtained. The desensitized model gradient and local model parameters of each edge node are homomorphically encrypted using the homomorphic encryption (HE) algorithm to obtain the encrypted gradient and parameter ciphertext of each edge node. The Secure Multi-Party Computation (MPC) algorithm is used to split the encryption gradient of each edge node into multiple fragments to obtain the gradient ciphertext of each edge node.

[0096] The differential privacy algorithm enables gradient-sensitive desensitization, preventing gradient back-inference of the original data.

[0097] The homomorphic encryption algorithm ensures that the gradient is "computable in ciphertext and invisible in plaintext" during transmission and aggregation.

[0098] The secure multi-party computation algorithm can distribute fragments to designated nodes in the cloud, thus preventing a single node from possessing the complete gradient.

[0099] Through the above embodiments, a triple protection mechanism can be formed based on multiple layers of privacy protection to ensure privacy and security throughout the entire process from data processing and transmission to computation.

[0100] Furthermore, encrypted gradients and parameters can be uploaded only to the central node or the cloud to avoid the leakage of sensitive information during model training.

[0101] The update unit 113 is used to perform sensitivity detection on the gradient ciphertext of each edge node, and add random perturbation to the gradient ciphertext that fails the sensitivity detection to update the gradient ciphertext.

[0102] In this embodiment, the update unit 113 performs sensitivity detection on the gradient ciphertext of each edge node, and adds random perturbation to the gradient ciphertext that fails the sensitivity detection to update the gradient ciphertext, including: The sensitive feature recognition rule base is invoked; wherein, the sensitive feature recognition rule base is used to store the correspondence between sensitive features, deviation thresholds, and disturbance noise information; The gradient ciphertext of each edge node is matched against the sensitive feature recognition rule base; The matched gradient ciphertext is identified as a gradient ciphertext that fails the sensitivity detection, and random perturbation is added to the gradient ciphertext that fails the sensitivity detection based on the matched perturbation noise information to update the gradient ciphertext.

[0103] For example, the sensitive feature recognition rule base may include: in the financial scenario, "user income gradient fluctuation > 50%, corresponding to Gaussian noise with an intensity of 0.5", and in the medical scenario, "tumor marker gradient deviates from the mean by 3 times, corresponding to Laplace noise with an intensity of 1".

[0104] In the above embodiments, privacy constraints can be verified by gradient sensitivity control and random perturbation to prevent malicious nodes from back-engineering the original data through gradients.

[0105] The generation unit 114 is used to collect multidimensional state data of each edge node and generate an aggregation strategy based on the Adaptive Federated Weighting (AFW) algorithm and the multidimensional state data.

[0106] In this embodiment, the multidimensional state data may include: 1) Network dimension: bandwidth and latency of node upload gradients; 2) Data dimension: Distributional dissimilarity of node data (deviation from the global data distribution, i.e., the degree of Non-IID (Non-Independently and Identically Distributed)). 3) Computational dimensions: CPU (Central Processing Unit) / GPU (Graphics Processing Unit) computing power and memory usage of the node; 4) Historical dimension: the contribution of nodes to the model in the past (such as the contribution rate of gradients to the improvement of global model accuracy). 5) Task dimension: high-frequency real-time task nodes (such as financial real-time abnormal behavior detection tasks, medical emergency diagnosis tasks, etc.) and high-precision requirement task nodes (such as financial risk deep modeling tasks, rare disease diagnosis model modeling tasks, etc.).

[0107] In this embodiment, the generation unit 114 collects multi-dimensional state data of each edge node and generates an aggregation strategy based on a dynamic node weight optimization algorithm and the multi-dimensional state data, including: Collect multi-dimensional status data of each edge node according to network dimension, data dimension, computing dimension, historical dimension, and task dimension; Identify strong and weak nodes in each edge node based on the multidimensional state data of each edge node; Aggregate weights are assigned based on the strong and weak nodes among the edge nodes; wherein the weight of the strong nodes is higher than that of the weak nodes. Identify high-frequency real-time task nodes and high-precision-requirement task nodes in each edge node based on the multi-dimensional state data of each edge node. Configure the high-frequency real-time task node to asynchronous aggregation mode, and configure the high-precision requirement task node to synchronous aggregation mode.

[0108] The strong nodes can include nodes with high bandwidth, low latency, uniform data distribution, and large historical contributions, which need to be assigned higher aggregation weights; the weak nodes can include organizations in remote areas, such as rural banks and community hospitals, which need to be assigned basic weights and have their gradient aggregation priority reduced, so as to avoid slowing down the overall progress.

[0109] Specifically, an asynchronous aggregation mode is configured for the high-frequency real-time task nodes, so that the high-frequency real-time task nodes do not need to wait for all nodes, thus reducing waiting latency.

[0110] Specifically, configuring a synchronous aggregation mode for the high-precision task nodes allows for unified aggregation after all node gradients have been uploaded, thereby ensuring model accuracy.

[0111] In the above embodiments, the dynamic node weight optimization algorithm can allocate aggregation weights according to the multi-dimensional state of nodes, avoiding low-quality nodes (such as low data volume and high latency) from dragging down the global model performance and improving aggregation efficiency; the automatic switching between synchronous and asynchronous aggregation strategies solves the global waiting problem caused by high latency of some nodes, thereby reducing the model update cycle.

[0112] The aggregation unit 115 is used to perform model aggregation in the cloud using a multi-layer semantic-aligned aggregation (SAA) algorithm based on the aggregation strategy, the parameter ciphertext, and the updated gradient ciphertext to obtain an initial global model.

[0113] In this embodiment, it is also necessary to address the issue of data distribution differences across multiple domains.

[0114] Specifically, the aggregation unit 115 is located in the cloud and uses a multi-layer semantic alignment aggregation algorithm to aggregate models based on the aggregation strategy, the parameter ciphertext, and the updated gradient ciphertext to obtain an initial global model, including: The multiple fragments are restored using the secure multi-party computation algorithm to obtain the encryption gradient of each edge node; The encryption gradients of each edge node are semantically aligned according to the semantic mapping dictionary. According to the aggregation strategy, the aligned encrypted gradients are weighted and aggregated based on the parameter ciphertext to obtain the initial global model.

[0115] In the above embodiments, the multi-layer semantic alignment aggregation algorithm can specifically solve the Non-IID problem (i.e., the problem of data distribution differences among nodes), and improve the adaptability of the global model to edge nodes.

[0116] In this embodiment, in the cloud, after the initial global model is obtained by using a multi-layer semantic alignment aggregation algorithm based on the aggregation strategy, the parameter ciphertext, and the updated gradient ciphertext, the aggregation strategy is adjusted using a reinforcement learning (RL) mechanism.

[0117] For example, a reward function can be configured, and the aggregation strategy can be updated based on the reward function after aggregation to improve the aggregation strategy.

[0118] In the above embodiments, latency perception and energy consumption balance can be achieved through the self-optimization mechanism of reinforcement learning, thereby adapting to the unstable network and limited computing power of edge nodes (such as financial branch servers and medical IoT devices).

[0119] The fusion unit 116 is used to perform cross-domain knowledge fusion based on the initial global model at each edge node using a cross-domain knowledge distillation mechanism to obtain the initial model of each edge node.

[0120] In this embodiment, it is also necessary to improve the generalization performance of the model.

[0121] Specifically, at each edge node, the fusion unit 116 employs a cross-domain knowledge distillation mechanism to perform cross-domain knowledge fusion based on the initial global model, obtaining the initial model for each edge node, including: The initial global model is used as the teacher model, and the knowledge of the teacher model is transformed into an encrypted knowledge graph; Using the local model of each edge node as the student model, and based on the data to be processed at each edge node, each student model is guided to learn the knowledge of the teacher model according to the encrypted knowledge graph, thus obtaining the initial model of each edge node.

[0122] The knowledge in the teacher model may include decision boundaries, feature importance, etc.

[0123] Through the above embodiments, the "local knowledge" of different nodes can be fused based on the cross-domain knowledge distillation mechanism, thereby enhancing the generalization ability of the global model.

[0124] The fine-tuning unit 117 is used to fine-tune the parameters of the initial model of each edge node in the cloud according to the data to be processed of each edge node, so as to obtain the sub-model of each edge node.

[0125] In this embodiment, in order to make the model of each edge node more suitable for the corresponding business needs, parameter fine-tuning is also required.

[0126] Specifically, the fine-tuning unit 117, located in the cloud, fine-tunes the parameters of the initial model of each edge node based on the data to be processed at each edge node, resulting in a sub-model for each edge node, including: Extract business requirement data from the data to be processed at each edge node; Based on the business requirements data of each edge node, the parameters of the initial model of each edge node are fine-tuned to obtain the sub-model of each edge node.

[0127] For example, for bank nodes where micro and small enterprises are concentrated, the ability to identify risks of micro and small enterprises can be strengthened; for pediatric hospital nodes, the ability to diagnose rare diseases in children can be strengthened.

[0128] Through the above embodiments, the local business needs of each edge node can be met by fine-tuning, thereby achieving the dual goals of global collaboration and local optimization.

[0129] The inference unit 118 is used to perform inference tasks locally on each edge node using the sub-models of each edge node, and send the obtained inference results to the triggerer of the corresponding inference task.

[0130] For example, financial and banking nodes can be used to infer the risks of abnormal transactions; medical diagnostic nodes can be used to infer the risks of diseases.

[0131] In this embodiment, after the sub-models of each edge node execute inference tasks locally at each edge node and send the obtained inference results to the triggerer of the corresponding inference task, a drift detection algorithm is used to detect the feature differences between the data features of each edge node and the data to be processed at each edge node in real time; when the feature difference corresponding to a first edge node is greater than the difference threshold, it is determined that the first edge node has experienced data distribution drift, and a retraining task is triggered; and / or The model performance of each edge node's sub-model is detected at preset time intervals; when the model performance of a second edge node's sub-model decreases by more than a preset threshold, the retraining task is triggered.

[0132] The preset time interval, the difference threshold, and the preset threshold can be optimal values ​​selected based on the experiment.

[0133] Through the above embodiments, regular performance evaluation and drift detection can be achieved, timely detection of performance degradation caused by changes in data distribution can be made, and model failure can be avoided; automatic model retraining can also ensure that the model is always adapted to the latest business data, thereby improving overall performance.

[0134] As can be seen from the above technical solutions, this invention can preprocess the local data collected by each edge node based on a lightweight semantic indexing mechanism, achieving cross-domain data logical alignment rather than physical migration, thus eliminating ETL dependence; it employs a multi-layer privacy protection mechanism to encrypt the local model parameters and gradients of each edge node, effectively improving data security; it adds random perturbations to the gradient ciphertext that fails sensitivity detection, achieving verifiable privacy constraints; based on a dynamic node weight optimization algorithm and multi-dimensional state data, it implements targeted configuration of aggregation strategies according to different node attributes; it uses a multi-layer semantic alignment aggregation algorithm for model aggregation, improving the adaptability of the global model on edge nodes; it employs a cross-domain knowledge distillation mechanism for cross-domain knowledge fusion, strengthening the generalization ability of the global model; and it fine-tunes the parameters of the initial model of each edge node based on the data to be processed at each edge node, meeting the personalized business needs of the node's locality, thereby assisting in the execution of intelligent inference locally at each edge node.

[0135] like Figure 3 The diagram shown is a structural schematic of a computer device that implements a preferred embodiment of the cross-domain reasoning method based on federated learning according to the present invention.

[0136] The computer device 1 may include a memory 12, a processor 13, and a bus (the arrow in the figure represents the bus), and may also include a computer program stored in the memory 12 and executable on the processor 13, such as a cross-domain inference program based on federated learning.

[0137] Those skilled in the art will understand that the schematic diagram is merely an example of computer device 1 and does not constitute a limitation on computer device 1. Computer device 1 can be either a bus topology or a star topology. Computer device 1 may also include more or fewer other hardware or software than shown in the diagram, or different component arrangements. For example, computer device 1 may also include input / output devices, network access devices, etc.

[0138] It should be noted that the computer device 1 described is merely an example. Other existing or future electronic products that are adaptable to this invention should also be included within the scope of protection of this invention and are incorporated herein by reference.

[0139] The memory 12 includes at least one type of readable storage medium, such as flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 12 can be an internal storage unit of the computer device 1, such as a portable hard drive of the computer device 1. In other embodiments, the memory 12 can be an external storage device of the computer device 1, such as a plug-in portable hard drive, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the computer device 1. Furthermore, the memory 12 can include both internal and external storage units of the computer device 1. The memory 12 can be used not only to store application software and various types of data installed on the computer device 1, such as the code of a cross-domain inference program based on federated learning, but also to temporarily store data that has been output or will be output.

[0140] In some embodiments, the processor 13 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits packaged with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 13 is the control unit of the computer device 1, connecting various components of the computer device 1 via various interfaces and lines. It executes programs or modules stored in the memory 12 (e.g., executing cross-domain inference programs based on federated learning) and calls data stored in the memory 12 to perform various functions of the computer device 1 and process data.

[0141] The processor 13 executes the operating system of the computer device 1 and various installed applications. The processor 13 executes these applications to implement the steps described in the various embodiments of the federated learning-based cross-domain inference method, for example... Figure 1 The steps are shown.

[0142] For example, the computer program may be divided into one or more modules / units, which are stored in the memory 12 and executed by the processor 13 to complete the present invention. The one or more modules / units may be a series of computer-readable instruction segments capable of performing specific functions, which describe the execution process of the computer program in the computer device 1. For example, the computer program may be divided into a preprocessing unit 110, a training unit 111, an encryption unit 112, an update unit 113, a generation unit 114, an aggregation unit 115, a fusion unit 116, a fine-tuning unit 117, and an inference unit 118.

[0143] The integrated unit implemented as a software functional module described above can be stored in a computer-readable storage medium. This software functional module, stored in a storage medium, includes several instructions to cause a computer device (which may be a personal computer, computer equipment, or network device, etc.) or processor to execute portions of the federated learning-based cross-domain inference method described in the various embodiments of this invention.

[0144] If the modules / units integrated in the computer device 1 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware devices. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above.

[0145] The computer program includes computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory, etc.

[0146] Furthermore, the computer-readable storage medium may primarily include a stored program area and a stored data area, wherein the stored program area may store the operating system, an application program required for at least one function, etc.; and the stored data area may store data created based on the use of blockchain nodes, etc.

[0147] The blockchain referred to in this invention is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked together using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include an underlying blockchain platform, a platform product service layer, and an application service layer.

[0148] The bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This bus can be divided into address bus, data bus, control bus, etc. For ease of representation, in... Figure 3 The bus is represented by only one straight line, but this does not mean that there is only one bus or one type of bus. The bus is configured to enable communication between the memory 12 and at least one processor 13, etc.

[0149] Although not shown, the computer device 1 may also include a power supply (such as a battery) to power various components. Preferably, the power supply can be logically connected to the at least one processor 13 through a power management device, thereby enabling functions such as charging management, discharging management, and power consumption management. The power supply may also include one or more DC or AC power supplies, recharging devices, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components. The computer device 1 may also include various sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be described in detail here.

[0150] Furthermore, the computer device 1 may also include a network interface. Optionally, the network interface may include a wired interface and / or a wireless interface (such as a Wi-Fi interface, a Bluetooth interface, etc.), which is typically used to establish a communication connection between the computer device 1 and other computer devices.

[0151] Optionally, the computer device 1 may further include a user interface, which may be a display, an input unit (such as a keyboard), and optionally, a standard wired interface or a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen, etc. The display may also be appropriately referred to as a screen or display unit, used to display information processed in the computer device 1 and to display a visual user interface.

[0152] It should be understood that the embodiments described are for illustrative purposes only and are not limited to this structure in the scope of the patent application.

[0153] It will be understood by those skilled in the art that Figure 3 The structure shown does not constitute a limitation on the computer device 1, and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0154] Combination Figure 1 The memory 12 in the computer device 1 stores multiple instructions to implement a cross-domain inference method based on federated learning, and the processor 13 can execute the multiple instructions to achieve: In response to cross-domain inference commands triggered by different business domains corresponding to each edge node, business data and system operation data are collected at each edge node as local data, and the local data collected at each edge node is preprocessed based on a lightweight semantic indexing mechanism to obtain the data to be processed at each edge node. The model is trained at each edge node using the data to be processed at each edge node, and the local model, local model parameters and local model gradient of each edge node are obtained. A multi-layered privacy protection mechanism is used to encrypt the local model parameters and local model gradients of each edge node, resulting in ciphertext of the parameters and gradients of each edge node. Sensitivity detection is performed on the gradient ciphertext of each edge node, and random perturbation is added to the gradient ciphertext that fails the sensitivity detection to update the gradient ciphertext; Collect multidimensional state data of each edge node, and generate an aggregation strategy based on the dynamic node weight optimization algorithm and the multidimensional state data; In the cloud, a multi-layer semantic alignment aggregation algorithm is used to aggregate models based on the aggregation strategy, the parameter ciphertext, and the updated gradient ciphertext to obtain an initial global model. At each edge node, a cross-domain knowledge distillation mechanism is used to perform cross-domain knowledge fusion based on the initial global model to obtain the initial model of each edge node; In the cloud, the parameters of the initial model of each edge node are fine-tuned based on the data to be processed at each edge node to obtain the sub-model of each edge node. The sub-models of each edge node are used to execute inference tasks locally on each edge node, and the inference results are sent to the trigger of the corresponding inference task.

[0155] Specifically, the processor 13's implementation method for the above instructions can be found in [reference needed]. Figure 1 The descriptions of the relevant steps in the corresponding embodiments are not repeated here.

[0156] It should be noted that all data involved in this case was legally obtained. Software tools or components not belonging to this company that appear in the embodiments of this application are merely illustrative examples and do not represent actual use.

[0157] In the several embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0158] This invention can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0159] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0160] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0161] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0162] Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within the invention. No appended diagram markings in the claims should be construed as limiting the scope of the claims.

[0163] Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices described in this invention can also be implemented by a single unit or device through software or hardware. Terms such as "first," "second," etc., are used to indicate names and do not indicate any specific order.

[0164] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A cross-domain reasoning method based on federated learning, characterized in that, The federated learning-based cross-domain reasoning method includes: In response to cross-domain inference commands triggered by different business domains corresponding to each edge node, business data and system operation data are collected at each edge node as local data, and the local data collected at each edge node is preprocessed based on a lightweight semantic indexing mechanism to obtain the data to be processed at each edge node. The model is trained at each edge node using the data to be processed at each edge node, and the local model, local model parameters and local model gradient of each edge node are obtained. A multi-layered privacy protection mechanism is used to encrypt the local model parameters and local model gradients of each edge node, resulting in ciphertext of the parameters and gradients of each edge node. Sensitivity detection is performed on the gradient ciphertext of each edge node, and random perturbation is added to the gradient ciphertext that fails the sensitivity detection to update the gradient ciphertext; Collect multidimensional state data of each edge node, and generate an aggregation strategy based on the dynamic node weight optimization algorithm and the multidimensional state data; In the cloud, a multi-layer semantic alignment aggregation algorithm is used to aggregate models based on the aggregation strategy, the parameter ciphertext, and the updated gradient ciphertext to obtain an initial global model. At each edge node, a cross-domain knowledge distillation mechanism is used to perform cross-domain knowledge fusion based on the initial global model to obtain the initial model of each edge node; In the cloud, the parameters of the initial model of each edge node are fine-tuned based on the data to be processed at each edge node to obtain the sub-model of each edge node. The sub-models of each edge node are used to execute inference tasks locally on each edge node, and the inference results are sent to the trigger of the corresponding inference task.

2. The cross-domain reasoning method based on federated learning as described in claim 1, characterized in that, The preprocessing of local data collected from each edge node based on the lightweight semantic indexing mechanism yields the following data to be processed for each edge node: For each edge node, obtain the structured data and unstructured data of the edge node; The structured data is converted into a business-related first feature vector using the lightweight feature extractor built into the edge node; The unstructured data is transformed into a second feature vector of a specified dimension using a pre-trained lightweight model; Define a semantic dimension and construct a semantic mapping dictionary based on the semantic dimension; wherein the semantic mapping dictionary is used to reflect the mapping relationship between the semantic dimension and heterogeneous features; The first feature vector and the second feature vector are mapped to a unified semantic space according to the semantic mapping dictionary to obtain the data to be processed for the edge node.

3. The cross-domain reasoning method based on federated learning as described in claim 2, characterized in that, The method employs a multi-layered privacy protection mechanism to encrypt the local model parameters and gradients of each edge node, resulting in the following encrypted parameters and gradients for each edge node: By adding controllable random noise to the local model gradient of each edge node using the differential privacy algorithm, the desensitized model gradient of each edge node is obtained. The desensitized model gradient and local model parameters of each edge node are homomorphically encrypted using a homomorphic encryption algorithm to obtain the encrypted gradient and parameter ciphertext of each edge node. A secure multi-party computation algorithm is used to split the encryption gradient of each edge node into multiple fragments to obtain the gradient ciphertext of each edge node.

4. The cross-domain reasoning method based on federated learning as described in claim 1, characterized in that, The step of performing sensitivity detection on the gradient ciphertext of each edge node and adding random perturbations to the gradient ciphertext that fails the sensitivity detection to update the gradient ciphertext includes: The sensitive feature recognition rule base is invoked; wherein, the sensitive feature recognition rule base is used to store the correspondence between sensitive features, deviation thresholds, and disturbance noise information; The gradient ciphertext of each edge node is matched against the sensitive feature recognition rule base. The matched gradient ciphertext is identified as a gradient ciphertext that fails the sensitivity detection, and random perturbation is added to the gradient ciphertext that fails the sensitivity detection based on the matched perturbation noise information to update the gradient ciphertext.

5. The cross-domain reasoning method based on federated learning as described in claim 1, characterized in that, The process of collecting multidimensional state data from each edge node and generating an aggregation strategy based on a dynamic node weight optimization algorithm and the multidimensional state data includes: Collect multi-dimensional status data of each edge node according to network dimension, data dimension, computing dimension, historical dimension, and task dimension; Identify strong and weak nodes in each edge node based on the multidimensional state data of each edge node; Aggregate weights are assigned based on the strong and weak nodes among the edge nodes; wherein the weight of the strong nodes is higher than that of the weak nodes. Identify high-frequency real-time task nodes and high-precision-requirement task nodes in each edge node based on the multi-dimensional state data of each edge node. Configure the high-frequency real-time task node as asynchronous aggregation mode, and configure the high-precision requirement task node as synchronous aggregation mode.

6. The cross-domain reasoning method based on federated learning as described in claim 3, characterized in that, In the cloud, the initial global model is obtained by using a multi-layer semantic alignment aggregation algorithm based on the aggregation strategy, the parameter ciphertext, and the updated gradient ciphertext, including: The multiple fragments are restored using the secure multi-party computation algorithm to obtain the encryption gradient of each edge node; The encryption gradients of each edge node are semantically aligned according to the semantic mapping dictionary. According to the aggregation strategy, the aligned encryption gradients are weighted and aggregated based on the parameter ciphertext to obtain the initial global model.

7. The cross-domain reasoning method based on federated learning as described in claim 1, characterized in that, In the cloud, after using a multi-layer semantic alignment aggregation algorithm to aggregate models based on the aggregation strategy, the parameter ciphertext, and the updated gradient ciphertext to obtain an initial global model, the method further includes: The aggregation strategy is adjusted using a reinforcement learning mechanism.

8. The cross-domain reasoning method based on federated learning as described in claim 1, characterized in that, At each edge node, a cross-domain knowledge distillation mechanism is used to perform cross-domain knowledge fusion based on the initial global model, resulting in the following initial models for each edge node: The initial global model is used as the teacher model, and the knowledge of the teacher model is transformed into an encrypted knowledge graph; Using the local model of each edge node as the student model, and based on the data to be processed at each edge node, each student model is guided to learn the knowledge of the teacher model according to the encrypted knowledge graph, thus obtaining the initial model of each edge node.

9. The cross-domain reasoning method based on federated learning as described in claim 1, characterized in that, In the cloud, the initial model of each edge node is fine-tuned based on the data to be processed at each edge node, resulting in a sub-model for each edge node, including: Extract business requirement data from the data to be processed at each edge node; Based on the business requirements data of each edge node, the parameters of the initial model of each edge node are fine-tuned to obtain the sub-model of each edge node.

10. The cross-domain reasoning method based on federated learning as described in claim 1, characterized in that, After performing inference locally at each edge node using the sub-model of each edge node, the method further includes: A drift detection algorithm is used to detect the feature differences between the data features of each edge node and the feature differences between the data to be processed at each edge node in real time; when the feature difference corresponding to the first edge node is greater than the difference threshold, it is determined that the first edge node has experienced data distribution drift, and a retraining task is triggered; and / or The model performance of each edge node's sub-model is detected at preset time intervals; when the model performance of a second edge node's sub-model decreases by more than a preset threshold, the retraining task is triggered.

Images