An artificial intelligence information security protection system and method

Abstract

The application provides an artificial intelligence information protection method and system, which is applied to a computing architecture including a center server and multiple edge nodes. First, in the current iteration learning period of the edge node, local system running data flow is obtained, and a local security evaluation model is input to obtain a local security evaluation value and a prediction confidence. According to the confidence level classification mark, certain data and pending data are marked and temporarily stored in the corresponding set. Then, the data features and evaluation values of the certain set and the pending set are uploaded to the center server, and an updated global security evaluation model is received and issued by the center server. Finally, in the next iteration period, the updated model is used to recalculate the evaluation value of the pending data, generate and execute a localized security protection strategy, and upload the strategy to the center server for global coordination. The above scheme can realize the iteration evaluation and privacy constraint of low confidence data in artificial intelligence information protection, thereby improving the overall security robustness of the system.

Description

Technical Field

[0001] This application relates to the field of information protection technology, and more specifically, to an information security protection system and method for artificial intelligence. Background Technology

[0002] In the field of AI-driven information protection, edge computing architecture, as a key framework for distributed data processing, is of great significance for achieving real-time security assessment, privacy protection, and system robustness optimization. Traditional security assessment methods often rely on centralized data aggregation models, which are prone to sensitive information leakage and single point of failure risks, and are difficult to support dynamic collaboration of large-scale edge nodes. With the surge in the number of edge devices and the increasing complexity of data flows, refined local assessment and global model collaboration capabilities are gradually becoming core requirements for AI-driven information protection systems.

[0003] Current mainstream AI-based information protection methods generally suffer from privacy leaks in their model update mechanisms, making it difficult to adapt to the variable uncertainty of data within edge nodes. This is particularly true in low-confidence data processing or cross-node collaboration scenarios, where the non-linear correlation between local evaluation values ​​and global threat intelligence is difficult to extract effectively, resulting in insufficient model responsiveness to unknown risks. The root cause of this deficiency lies in the traditional methods' failure to introduce an iterative learning mechanism that dynamically distinguishes between locally confident data and pending data. They also lack methods for temporarily storing low-confidence data and privacy enhancement strategies for the global model, thus limiting the system's ability to analyze information leakage and its protection stability. Therefore, how to achieve iterative evaluation and privacy constraints on low-confidence data in AI-based information protection, thereby improving the overall security robustness of the system, has become a challenging problem for the industry. Summary of the Invention

[0004] This invention provides a method and system for protecting information in artificial intelligence, which can realize iterative evaluation and privacy constraints on low-confidence data in artificial intelligence information protection, and improve the overall security and robustness of the system.

[0005] In a first aspect, the present invention provides an information protection method for artificial intelligence, applied in a computing architecture comprising a central server and multiple edge nodes, the method comprising:

[0006] Within the current iterative learning cycle of the edge node, the local system operation data stream to be analyzed is acquired, and the system operation data stream is input into the local security assessment model deployed on the edge node to obtain the local security assessment value and prediction confidence corresponding to each data.

[0007] If the confidence level of the local security assessment model's prediction for the data is lower than a preset threshold, it is marked as pending data and temporarily stored in the local pending set; otherwise, it is marked as certain data and stored in the local certain set.

[0008] Upload the data features and corresponding local security assessment values ​​of the local confidence set and local undetermined set within the current iterative learning cycle to the central server.

[0009] Receive the updated global security assessment model issued by the central server;

[0010] After the next iteration learning cycle begins, the updated global security assessment model is used to recalculate the security assessment value of the pending data in the local pending set, and a security protection strategy for the edge node is generated based on the recalculated security assessment value. The security protection strategy is then uploaded to the central server for global security strategy coordination.

[0011] In some embodiments, acquiring the local system operation data stream to be analyzed and inputting the system operation data stream into the local security assessment model deployed on the edge node to obtain the local security assessment value and prediction confidence corresponding to each data point specifically includes:

[0012] Perform data cleaning operations on the acquired system operation data stream to remove redundant fields and noisy data;

[0013] Feature extraction is performed on the cleaned data, and then the extracted features are normalized to obtain a structured data feature vector.

[0014] The structured data feature vectors are input into the local security assessment model to calculate the local security assessment value and prediction confidence for each data point.

[0015] In some embodiments, if the prediction confidence of the local security assessment model for the data is lower than a preset threshold, it is marked as pending data and temporarily stored in the local pending set; otherwise, it is marked as certain data and stored in the local certain set. Specifically, this includes:

[0016] Compare the predicted confidence level with a preset threshold;

[0017] If the prediction confidence level is higher than a preset threshold, the data is marked as confident data and stored in the local confidence set. At the same time, a preliminary local security protection strategy is generated based on the local security assessment value and executed at the edge node.

[0018] The confirmed data is input into the incremental training process of the local security assessment model to calculate the loss function and generate parameter updates;

[0019] If the prediction confidence is lower than a preset threshold, the data is marked as pending data, anonymized to generate a feature summary, and the anonymized feature summary is temporarily stored in the local pending set.

[0020] In some embodiments, uploading the data features and corresponding local security assessment values ​​of the local confidence set and the local undetermined set within the current iterative learning cycle to the central server specifically includes:

[0021] Extract the true labels and local security assessment values ​​of the confidence data from the local confidence set, calculate the mean squared error loss function, and generate the model gradient vector as the parameter update amount;

[0022] Extract anonymized feature summaries and corresponding ground labels from the local undetermined set of data;

[0023] Gaussian noise is added to the model gradient vector to perturb it, and the perturbed model gradient vector, the anonymized feature summary, and the corresponding real label are encrypted using a homomorphic encryption algorithm.

[0024] The encrypted model gradient vectors, encrypted anonymized feature summaries, and encrypted corresponding real labels are uploaded to the central server.

[0025] In some embodiments, receiving the updated global security assessment model issued by the central server specifically includes:

[0026] The central server receives encrypted model gradient vectors from each edge node, decrypts them to obtain decrypted gradient vectors, and aggregates all decrypted gradient vectors to generate a global gradient update.

[0027] The central server optimizes the global gradient update based on the global threat intelligence database and calculates the global model parameter adjustment values;

[0028] The central server generates a real label verification set and applies the global model parameter adjustment value to the current global security assessment model to obtain the updated global security assessment model.

[0029] The updated global security assessment model and the real label verification set are distributed to each edge node.

[0030] The edge node receives the updated global security assessment model and the real label verification set.

[0031] In some embodiments, after the start of the next iteration learning cycle, the updated global security assessment model is used to recalculate the security assessment value of the pending data in the local pending set, and a security protection policy for the edge node is generated based on the recalculated security assessment value. The security protection policy is then uploaded to the central server for global security policy coordination. Specifically, this includes:

[0032] At the end of the current iteration cycle, load the updated global security assessment model;

[0033] Anonymized feature digests of pending data are extracted from the local pending set, and the anonymized feature digests are input into the updated global security assessment model to recalculate the security assessment value of the pending data.

[0034] Calculate the final security assessment value based on the recalculated security assessment value and the true label of the pending data;

[0035] Based on the final security assessment value and preset local scenario adaptation rules, a localized security protection strategy is generated and executed at the edge node.

[0036] Extract the summary of the localized security protection policy, upload the summary to the central server, and after receiving the summary data packet, the central server decodes it, aggregates the summary data of all edge nodes, generates a global security situation analysis report, and performs cross-node policy coordination.

[0037] In some embodiments, the method further includes:

[0038] At the end of each iterative learning cycle, the central server collects the encrypted parameter updates and real label feedback uploaded by all edge nodes;

[0039] The central server decrypts and aggregates the updated encrypted parameters, performs global model optimization, generates a global security situation report based on objective verification data, and distributes the global security situation report to each edge node.

[0040] Edge nodes receive the global security posture report and use the data in the report to further fine-tune the parameters of the local security assessment model to form a closed-loop iterative mechanism;

[0041] Edge nodes periodically scan their local pending sets and delete processed data to free up storage resources.

[0042] Secondly, the present invention provides an information protection system for artificial intelligence, applied in a computing architecture comprising a central server and multiple edge nodes, the system comprising:

[0043] The acquisition module is used to acquire the local system operation data stream to be analyzed within the current iterative learning cycle of the edge node, and input the system operation data stream into the local security assessment model deployed on the edge node to obtain the local security assessment value and prediction confidence corresponding to each data.

[0044] The processing module is used to mark the data as pending data and temporarily store it in the local pending set if the prediction confidence of the local security assessment model is lower than a preset threshold; otherwise, it is marked as certain data and stored in the local certain set.

[0045] The processing module is also used to upload the data features and corresponding local security assessment values ​​of the local confidence set and the local undetermined set in the current iterative learning cycle to the central server.

[0046] The processing module is also used to receive the updated global security assessment model issued by the central server;

[0047] The execution module is used to recalculate the security assessment value of the pending data in the local pending set using the updated global security assessment model after the start of the next iteration learning cycle, generate a security protection policy for the edge node based on the recalculated security assessment value, and upload the security protection policy to the central server for global security policy coordination.

[0048] Thirdly, the present invention provides a computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the aforementioned artificial intelligence information protection method.

[0049] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned information protection method for artificial intelligence.

[0050] The technical solutions provided by the embodiments disclosed in this invention have the following beneficial effects:

[0051] In this embodiment, within the current iterative learning cycle of the edge node, the local system operation data stream to be analyzed is acquired, and the system operation data stream is input into the local security assessment model deployed on the edge node to obtain the local security assessment value and prediction confidence level corresponding to each data. If the prediction confidence level of the local security assessment model for the data is lower than a preset threshold, it is marked as pending data and temporarily stored in the local pending set; otherwise, it is marked as certain data and stored in the local certain set. The data features and corresponding local security assessment values ​​of the local certain set and the local pending set within the current iterative learning cycle are uploaded to the central server. The updated global security assessment model issued by the central server is received. After the start of the next iterative learning cycle, the security assessment value of the pending data in the local pending set is recalculated using the updated global security assessment model, and a security protection strategy for the edge node is generated based on the recalculated security assessment value. The security protection strategy is then uploaded to the central server for global security strategy coordination.

[0052] Therefore, this application, by introducing a confidence threshold mechanism at edge nodes to classify local data into certain and pending categories, can highlight the potential impact of low-confidence data on the overall security assessment while protecting privacy. This establishes an assessment expression mechanism with uncertainty discrimination capabilities. The classification process extracts the coupling features between local assessment values ​​and confidence levels, achieving high-dimensional privacy representation of the data. Based on this, a global model update path is constructed, enhancing the system's ability to perceive differences in edge data from the source. Secondly, by uploading the anonymized features of the certain and pending sets to the central server, the optimization intensity of the global model can be dynamically adjusted, effectively suppressing the interference of low-confidence data on the model and highlighting the contribution features of high-confidence data, thus realizing model injection. The controllable adjustment of intention provides a clearer semantic basis for subsequent iterative evaluations. Then, based on the received updated global model, the security assessment value of the pending data is recalculated, and a local protection strategy is generated and uploaded for coordination, thereby outputting a global strategy with privacy tolerance capabilities. This significantly improves the adaptability of the information protection system to complex risk behaviors. Finally, when entering the next cycle, the re-evaluation process of low-confidence data is activated through an iterative mechanism, enabling the system to further strengthen privacy constraints in a dynamic environment and minimize the misleading effect of data leakage on system judgments. This mechanism, through dynamic privacy constraints, constructs an adaptive adjustment path for information protection in a distributed scenario, effectively overcoming the problem of insufficient robustness caused by fixed evaluation patterns in traditional methods. In summary, the proposed solution can realize iterative evaluation and privacy constraints of low-confidence data in artificial intelligence information protection, thereby improving the overall security robustness of the system. Attached Figure Description

[0053] Figure 1 This is an exemplary flowchart of an information protection method for artificial intelligence according to some embodiments of the present invention;

[0054] Figure 2 This is an exemplary flowchart illustrating the uploading of data to a central server according to some embodiments of the present invention;

[0055] Figure 3 This is a schematic diagram of the structure of an artificial intelligence information protection system according to some embodiments of the present invention;

[0056] Figure 4 This is a schematic diagram of the structure of a computer device for implementing an information protection method for artificial intelligence according to some embodiments of the present invention. Detailed Implementation

[0057] To better understand the above technical solutions, a detailed description of the solutions will be provided below in conjunction with the accompanying drawings and specific implementation methods. (Reference) Figure 1The figure is an exemplary flowchart of an artificial intelligence information protection method according to some embodiments of the present invention. The method mainly includes the following steps:

[0058] In step 101, within the current iterative learning cycle of the edge node, the local system operation data stream to be analyzed is acquired, and the system operation data stream is input into the local security assessment model deployed on the edge node to obtain the local security assessment value and prediction confidence corresponding to each data.

[0059] The local security assessment model is a model used at edge nodes to perform security analysis on locally processed system operation data. Specifically, it is constructed as follows: First, security features that the model can parse are built. Second, through training with security-labeled samples, a mapping rule between security features and risk states is established. Once training is complete, the local security assessment model is obtained. It should be noted that the local security assessment value is a quantitative result of the mapping to the risk state corresponding to the current input feature, while the prediction confidence level measures the reliability of the local security assessment value. Essentially, it is a quantitative judgment of the local security assessment model's fit between the current input feature and the established mapping between security features and risk states. That is, the mapping between security features and risk states is the core logical foundation of the local security model; the local security assessment value is the risk quantification result of the mapping; and the prediction confidence level is the reliability verification result of the mapping result. These three together constitute the model's complete judgment of the local data security state. Through training, the error between the prediction result and the actual situation is continuously calculated, ultimately yielding the final judgment logic for the local security assessment value and the prediction confidence level.

[0060] In addition, training can be performed using supervised learning algorithms, support vector machine algorithms, or other supervised learning algorithms; no specific limitations are made here.

[0061] In this invention, system operation data stream refers to network traffic, device logs, or business interaction data generated locally by edge nodes. This invention requires data cleaning, feature extraction, and normalization to ensure the standardization of data input to the model and avoid model failure caused by directly inputting raw data. Specifically, the system operation data stream can be captured from local system logs and network traffic; data cleaning operations are performed on the system operation data stream to remove redundant fields and noisy data; features are extracted from the cleaned data, such as calculating source IP entropy, port access frequency, data transmission volume fluctuation indicators, packet header checksum statistics, and transmission error rate; the extracted features are normalized to obtain a structured data feature vector; the structured data feature vector is input into the local security assessment model to calculate the local security assessment value and prediction confidence for each data point. The local security assessment model uses a supervised learning framework and performs local inference through forward propagation. (That is, edge nodes utilize locally deployed supervised learning models to calculate structured data feature vectors in a unidirectional process of "input → intermediate calculation → output" (i.e., forward propagation), quickly generating security assessment values ​​and confidence levels locally, achieving real-time, privacy-preserving security analysis independent of a central authority.) During the inference process, a pre-set local threat intelligence database is used to match data features to generate initial true labels, and these initial true labels are compared with the model output to assist in evaluation. It should be noted that labels are identifiers used to mark the security status attributes of data (such as edge node system operation data), while true labels are benchmark identifiers used to clearly mark the actual security status of system operation data in edge nodes. This process conforms to the distributed model iteration logic, performing inference and evaluation only locally, avoiding central reliance on prediction results; at the same time, generating initial labels through a local threat intelligence database ensures that there is reliable evidence to support subsequent iterations, rather than relying on pure prediction values.

[0062] In step 102, if the confidence level of the local security assessment model in predicting the data is lower than a preset threshold, it is marked as pending data and temporarily stored in the local pending set; otherwise, it is marked as certain data and stored in the local certain set.

[0063] The "confidence data" refers to high-reliability data selected by the edge node through a local security assessment model. Specifically, the predicted confidence level is compared with a preset threshold. If the predicted confidence level is higher than the preset threshold, the data is marked as confidence data and stored in a local confidence set. Simultaneously, a preliminary local security protection strategy is generated based on the local security assessment value. This preliminary strategy includes setting access control rules or traffic blocking instructions and immediately executing the strategy on the edge node. The confidence data is then input into the incremental training process of the local security assessment model to calculate the loss function and generate parameter updates. If the predicted confidence level is lower than the preset threshold, the data is marked as pending data, anonymized to generate a feature summary, and temporarily stored in a local pending set. Within the current iteration cycle, the true label of the pending data is obtained by sending cross-validation requests to adjacent edge nodes or by calling a local manual annotation interface. This step avoids long-term temporary storage of complete data, storing only anonymized feature summaries, which is in line with the limited resources of edge computing and prevents storage overflow; at the same time, by generating and executing preliminary protection strategies locally in real time, it ensures the real-time requirements of information security, rather than delaying across cycles.

[0064] It should be noted that the purpose of the loss function is to compare the model's prediction of the certain data (i.e., the local security assessment value) with the true label of the certain data and calculate the error between the two (for example, by using mean squared error or cross-entropy to calculate the error between the two), which will not be elaborated here.

[0065] In step 103, the data features and corresponding local security assessment values ​​of the local confidence set and the local undetermined set within the current iterative learning cycle are uploaded to the central server.

[0066] refer to Figure 2 This diagram is an exemplary flowchart illustrating the uploading of data to a central server according to some embodiments of the present invention, specifically including:

[0067] In step 1031, the true labels and local security assessment values ​​of the confidence data are extracted from the local confidence set, the mean squared error loss function is calculated, and the model gradient vector is generated as the parameter update amount. The model gradient vector is a quantitative indicator of the direction and magnitude of the adjustment of the internal parameters (such as feature weights and biases) of the local security analysis model.

[0068] In step 1032, anonymized feature summaries and corresponding real labels of the pending data are extracted from the local pending set;

[0069] In step 1033, a differential privacy mechanism is applied to the model gradient vector by adding Gaussian noise for perturbation. That is, Gaussian noise is added to the model gradient vector in the plaintext domain based on the distribution characteristics (e.g., mean, variance) of the model gradient vector to ensure that the noise matches the statistical distribution of the plaintext gradient (i.e., the original model gradient vector that has not yet undergone homomorphic encryption), and to avoid the deviation of noise addition due to the disorder of the ciphertext.

[0070] In step 1034, the perturbed model gradient vector, the anonymized feature summary, and the corresponding real label are encrypted. The encryption can be performed using a homomorphic encryption algorithm or other partially homomorphic encryption algorithms to ensure that the privacy of the noisy sensitive data is not compromised during transmission.

[0071] In step 1035, the encrypted model gradient vector, encrypted anonymized feature summary, and encrypted corresponding ground truth labels are uploaded to the central server. This process conforms to the core principle of distributed model iteration, uploading only the model parameter update (gradient vector) and ground truth label feedback, rather than the prediction results, supporting the central server in optimizing the global model through backpropagation. Simultaneously, differential privacy and homomorphic encryption ensure privacy protection, avoiding leakage caused by directly uploading data features, thus meeting information protection objectives.

[0072] In step 104, the updated global security assessment model is received from the central server.

[0073] In practice, the central server receives encrypted model gradient vectors, encrypted anonymized feature summaries, and encrypted corresponding real labels from each edge node. It then decrypts and aggregates all gradient vectors, anonymized feature summaries, and real labels to generate a global gradient update. The central server optimizes the aggregated gradient update by combining the global threat intelligence database and the aggregated real labels, calculating global model parameter adjustment values. The central server generates a real label verification set, which includes known security event samples extracted from the threat intelligence database and aggregated real labels. The global model parameter adjustment values ​​are applied to the current global security assessment model to obtain an updated global security assessment model. The updated global security assessment model and the real label verification set are then distributed to each edge node. Each edge node receives the updated global security assessment model and the real label verification set and uses the real label verification set to calculate the model accuracy locally to verify generalization performance. This step forms a closed-loop iteration through real label feedback, avoiding label bias and using the verification set to prevent data leakage, thus ensuring objective model evaluation.

[0074] In step 105, after the start of the next iteration learning cycle, the updated global security assessment model is used to recalculate the security assessment value of the pending data in the local pending set, and a security protection strategy for the edge node is generated based on the recalculated security assessment value. The security protection strategy is then uploaded to the central server for global security strategy coordination.

[0075] In specific implementation, at the end of the current iteration cycle, the updated global security assessment model is loaded; anonymized feature summaries of the pending data are extracted from the local pending set, and the anonymized feature summaries are input into the updated global security assessment model to recalculate the security assessment value; the final security assessment value is calculated by combining the recalculated security assessment value and the acquired real labels; based on the final security assessment value and preset local scenario adaptation rules, a localized security protection strategy is generated, wherein the localized security protection strategy includes updating firewall rules, activating intrusion detection modules, or adjusting data encryption levels, which are not specifically limited here; the localized security protection strategy is immediately executed at the edge node; and the data is then submitted to the relevant authorities. Extract a summary of the localized security protection policy (specifically, the type, triggering conditions, execution actions, and scope of impact of the localized security protection policy can be extracted; this is only an example and not a limitation), perform structured encoding on the extracted summary to obtain a summary data packet; upload the summary data packet to the central server through an encrypted channel, wherein the upload includes sending the metadata and verification value of the summary data packet; after receiving the summary data packet, the central server decodes and aggregates the summary data of all edge nodes to generate a global security situation analysis report and perform cross-node policy coordination, wherein cross-node policy coordination includes comparing the policy summaries of different nodes to identify common threats and issuing coordination instructions.

[0076] It should be noted that this step ensures the real-time generation and execution of protection strategies, avoids cross-cycle delays, and meets the real-time requirements of information security. Localized strategy generation retains scenario adaptability, and the center only coordinates common threats, rather than making unified adjustments, reducing efficiency redundancy and maintaining privacy protection through summary uploading.

[0077] In some embodiments, this application may further include the following steps: at the end of each iterative learning cycle, the central server collects the encrypted parameter update volume and real label feedback uploaded by all edge nodes; the central server decrypts and aggregates the encrypted parameter update volume, performs global model optimization, which may include applying the stochastic gradient descent algorithm to update the model weights, without specific limitations; the central server generates a global security situation report based on objective verification data, wherein the objective verification data includes confirmed security events extracted from the global threat intelligence database (the global security situation report may include threat distribution statistics and risk trend analysis); the global security situation report is distributed to each edge node; the edge node receives the global security situation report and uses the data in the report to further fine-tune the parameters of the local security assessment model, forming a closed-loop iterative mechanism; in addition, the edge node periodically scans the local undetermined set, deletes the processed data to release storage resources, thereby avoiding the exhaustion of edge resources through periodic cleanup and strengthening the closed-loop iterative logic.

[0078] Furthermore, in another aspect of the present invention, in some embodiments, the present invention provides an artificial intelligence-based information protection system, see reference. Figure 3 The figure is a schematic diagram of the structure of an artificial intelligence information protection system 300 according to some embodiments of the present invention. The artificial intelligence information protection system 300 includes: an acquisition module 301, a processing module 302, and an execution module 303, which are described below:

[0079] The acquisition module 301 in this invention is mainly used to acquire the local system operation data stream to be analyzed within the current iterative learning cycle of the edge node, and input the system operation data stream into the local security assessment model deployed on the edge node to obtain the local security assessment value corresponding to each data. If the prediction confidence of the local security assessment model for the data is lower than a preset threshold, it is marked as pending data and temporarily stored in the local pending set; otherwise, it is marked as certain data and stored in the local certain set.

[0080] Processing module 302, in this invention, is mainly used to upload the data features and corresponding local security assessment values ​​of the local confidence set and the local undetermined set within the current iterative learning cycle to the central server; the processing module 302 is also used to receive the updated global security assessment model issued by the central server;

[0081] The execution module 303 in this invention is mainly used to recalculate the security assessment value of the pending data in the local pending set using the updated global security assessment model after the start of the next iteration learning cycle, generate a security protection strategy for the edge node based on the recalculated security assessment value, and upload the security protection strategy to the central server for global security strategy coordination.

[0082] In addition, the present invention also provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described artificial intelligence information protection method.

[0083] In some embodiments, reference Figure 4 This figure is a schematic diagram of the structure of a computer device for implementing an information protection method for artificial intelligence according to some embodiments of the present invention. The method in the above embodiments can be achieved through... Figure 4 The computer device shown is used to implement this, and the computer device 400 includes at least one processor 401, a communication bus 402, a memory 403, and at least one communication interface 404.

[0084] The processor 401 may be a general-purpose central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more for controlling the execution of the information protection method for artificial intelligence in this invention.

[0085] The communication bus 402 may include a path for transmitting information between the aforementioned components.

[0086] The memory 403 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or it may be an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CDROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), a magnetic disk or other magnetic storage device, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto. The memory 403 may exist independently and be connected to the processor 401 via a communication bus 402. The memory 403 may also be integrated with the processor 401.

[0087] The memory 403 stores program code for executing the present invention, and its execution is controlled by the processor 401. The processor 401 executes the program code stored in the memory 403. The program code may include one or more software modules.

[0088] Communication interface 404 uses any transceiver-like device for communicating with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area network (WLAN), etc.

[0089] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single CPU) processor or a multi-core (multi CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0090] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This embodiment of the invention does not limit the type of computer device.

[0091] In addition, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described artificial intelligence information protection method.

[0092] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0093] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An information protection method for artificial intelligence, applied in a computing architecture comprising a central server and multiple edge nodes, characterized in that, The method includes: Within the current iterative learning cycle of the edge node, the local system operation data stream to be analyzed is acquired, and the system operation data stream is input into the local security assessment model deployed on the edge node to obtain the local security assessment value and prediction confidence corresponding to each data. If the confidence level of the local security assessment model's prediction for the data is lower than a preset threshold, it is marked as pending data and temporarily stored in the local pending set; otherwise, it is marked as certain data and stored in the local certain set. Upload the data features and corresponding local security assessment values ​​of the local confidence set and local undetermined set within the current iterative learning cycle to the central server. Receive the updated global security assessment model issued by the central server; After the next iteration learning cycle begins, the updated global security assessment model is used to recalculate the security assessment value of the pending data in the local pending set, and a security protection strategy for the edge node is generated based on the recalculated security assessment value. The security protection strategy is then uploaded to the central server for global security strategy coordination. Specifically, if the local security assessment model's prediction confidence level for the data is lower than a preset threshold, it is marked as pending data and temporarily stored in the local pending set; otherwise, it is marked as certain data and stored in the local certain set. Compare the predicted confidence level with a preset threshold; If the prediction confidence level is higher than a preset threshold, the data is marked as confident data and stored in the local confidence set. At the same time, a preliminary local security protection strategy is generated based on the local security assessment value and executed at the edge node. The confirmed data is input into the incremental training process of the local security assessment model to calculate the loss function and generate parameter updates; If the prediction confidence is lower than a preset threshold, the data is marked as pending data, anonymized to generate a feature summary, and the anonymized feature summary is temporarily stored in the local pending set. Additionally, within the current iteration cycle, the true labels of the pending data are obtained by sending cross-validation requests to adjacent edge nodes or by calling the local manual annotation interface.

2. The method according to claim 1, characterized in that, Acquire the local system operation data stream to be analyzed, and input the system operation data stream into the local security assessment model deployed on the edge node to obtain the local security assessment value and prediction confidence corresponding to each data point. Specifically, this includes: Perform data cleaning operations on the acquired system operation data stream to remove redundant fields and noisy data; Feature extraction is performed on the cleaned data, and then the extracted features are normalized to obtain a structured data feature vector. The structured data feature vectors are input into the local security assessment model to calculate the local security assessment value and prediction confidence for each data point.

3. The method according to claim 1, characterized in that, Uploading the data features and corresponding local security assessment values ​​of the local confidence set and local undetermined set within the current iterative learning cycle to the central server specifically includes: Extract the true labels and local security assessment values ​​of the confidence data from the local confidence set, calculate the mean squared error loss function, and generate the model gradient vector as the parameter update amount; Extract anonymized feature summaries and corresponding ground labels from the local undetermined set of data; Gaussian noise is added to the model gradient vector to perturb it, and the perturbed model gradient vector, the anonymized feature summary, and the corresponding real label are encrypted using a homomorphic encryption algorithm. The encrypted model gradient vectors, encrypted anonymized feature summaries, and encrypted corresponding real labels are uploaded to the central server.

4. The method according to claim 1, characterized in that, Receiving the updated global security assessment model issued by the central server specifically includes: The central server receives encrypted model gradient vectors from each edge node, decrypts them to obtain decrypted gradient vectors, and aggregates all decrypted gradient vectors to generate a global gradient update. The central server optimizes the global gradient update based on the global threat intelligence database and calculates the global model parameter adjustment values; The central server generates a real label verification set and applies the global model parameter adjustment value to the current global security assessment model to obtain the updated global security assessment model. The updated global security assessment model and the real label verification set are distributed to each edge node. The edge node receives the updated global security assessment model and the real label verification set.

5. The method according to claim 1, characterized in that, After the next iteration of the learning cycle begins, the updated global security assessment model is used to recalculate the security assessment values ​​of the pending data in the local pending set. Based on the recalculated security assessment values, a security protection strategy for the edge node is generated, and the security protection strategy is uploaded to the central server for global security strategy coordination. Specifically, this includes: At the end of the current iteration cycle, load the updated global security assessment model; Anonymized feature digests of pending data are extracted from the local pending set, and the anonymized feature digests are input into the updated global security assessment model to recalculate the security assessment value of the pending data. Calculate the final security assessment value based on the recalculated security assessment value and the true label of the pending data; Based on the final security assessment value and preset local scenario adaptation rules, a localized security protection strategy is generated and executed at the edge node. Extract the summary of the localized security protection policy, upload the summary to the central server, and after receiving the summary data packet, the central server decodes it, aggregates the summary data of all edge nodes, generates a global security situation analysis report, and performs cross-node policy coordination.

6. The method according to claim 1, characterized in that, The method further includes: At the end of each iterative learning cycle, the central server collects the encrypted parameter updates and real label feedback uploaded by all edge nodes; The central server decrypts and aggregates the updated encrypted parameters, performs global model optimization, generates a global security situation report based on objective verification data, and distributes the global security situation report to each edge node. Edge nodes receive the global security posture report and use the data in the report to further fine-tune the parameters of the local security assessment model to form a closed-loop iterative mechanism; Edge nodes periodically scan their local pending sets and delete processed data to free up storage resources.

7. An artificial intelligence information protection system, applied in a computing architecture comprising a central server and multiple edge nodes, wherein the information protection is performed using the method described in any one of claims 1 to 6, characterized in that, The system includes: The acquisition module is used to acquire the local system operation data stream to be analyzed within the current iterative learning cycle of the edge node, and input the system operation data stream into the local security assessment model deployed on the edge node to obtain the local security assessment value and prediction confidence corresponding to each data. The processing module is used to mark the data as pending data and temporarily store it in the local pending set if the prediction confidence of the local security assessment model is lower than a preset threshold; otherwise, it is marked as certain data and stored in the local certain set. The processing module is also used to upload the data features and corresponding local security assessment values ​​of the local confidence set and the local undetermined set in the current iterative learning cycle to the central server. The processing module is also used to receive the updated global security assessment model issued by the central server; The execution module is used to recalculate the security assessment value of the pending data in the local pending set using the updated global security assessment model after the start of the next iteration learning cycle, generate a security protection policy for the edge node based on the recalculated security assessment value, and upload the security protection policy to the central server for global security policy coordination.

8. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing code, and the processor being configured to retrieve the code and execute the information protection method for artificial intelligence as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the information protection method for artificial intelligence as described in any one of claims 1 to 6.

Images