Industrial control service model updating iteration method and system for resisting confrontation sample attack

Abstract

The invention relates to the field of industrial control, and provides an industrial control service model updating iteration method and system for resisting confrontation sample attacks, and the method comprises the steps: obtaining a service mirror image model of an online service model; obtaining a basic use case data set of the production business, and performing performance and efficiency test on the business mirror image model by using the basic use case data set to generate a stability reference index; a new data stream of the production service is discriminated by using the service mirror image model, and an adversarial sample is identified and an incremental data set is obtained through similarity matching between a discrimination mode and an attack feature; the incremental data set is fused into the basic use case data set in stages, and iterative fine tuning is performed on the business mirror image model according to deviation evaluation of the stability reference index; and updating the parameters of the online business model and the basic use case data set based on the fine-tuned business mirror image model and the incremental data set. According to the method, an adversarial sample detection mechanism is introduced in the business model updating process, so that wrong guidance of adversarial sample attacks to the business model can be effectively avoided.

Description

Technical Field

[0001] This invention relates to the field of industrial control technology, and in particular to a method and system for updating and iterating industrial control business models to resist adversarial sample attacks. Background Technology

[0002] With the increasing intelligence of industrial control systems, machine learning algorithms are widely used in building key business models such as equipment fault diagnosis, anomaly detection, and production optimization. These models rely on historical data for training and inference, but the industrial environment is dynamic. Therefore, these business models need to use regularly updated real-time data to adjust model parameters to adapt to new system operating states. However, in the actual process of model updating and adjustment, the diversity and complexity of data sources make it difficult to guarantee the reliability and accuracy of datasets. Low-quality data may lead to a decline in model performance or even introduce potential biases. More seriously, attackers may interfere with the model training process by injecting carefully crafted adversarial examples, causing the model to learn incorrect features during parameter fine-tuning, thereby disrupting its decision-making logic.

[0003] Therefore, there is an urgent need for a robust model update and iteration method that can effectively identify and filter adversarial examples during the data input stage, while dynamically monitoring model performance during parameter optimization to ensure that the updated model can both adapt to new data and resist malicious interference. Summary of the Invention

[0004] The present invention aims to solve at least one of the problems existing in the prior art, and provides a method and system for updating and iterating industrial control business models to resist adversarial sample attacks.

[0005] One aspect of the present invention provides a method for updating and iterating an industrial control business model to resist adversarial sample attacks, comprising the following steps:

[0006] By cloning the runtime environment and reconstructing the model configuration, a business mirror model of the online business model is obtained;

[0007] Obtain the basic use case dataset for production operations, and use the basic use case dataset to perform performance and efficiency tests on the business image model to generate stability benchmark indicators.

[0008] The business mirroring model is used to identify new data streams in the production business. By matching the similarity between the identification pattern and the attack features, adversarial examples are identified and incremental datasets are obtained.

[0009] The incremental dataset is integrated into the basic use case dataset in stages, and the parameters of the business image model are iteratively fine-tuned based on the deviation assessment of the stability benchmark index.

[0010] Based on the fine-tuned business mirror model and the incremental dataset, the parameters of the online business model and the basic use case dataset are updated synchronously.

[0011] Optionally, obtaining the business mirror model of the online business model through runtime environment cloning and model configuration reconstruction includes:

[0012] Clone the runtime environment of the online business system, including the operating system version, software dependency package version, and environment variable configuration, and use container technology to build an offline testing system;

[0013] Export the configuration file of the online business model, and based on the configuration file of the online business model, establish the same model structure as the online business model in the offline testing system to obtain the business mirror model;

[0014] The structural consistency between the business mirror model and the online business model is verified using a computational graph.

[0015] Optionally, the step of obtaining the basic use case dataset for production operations, and using the basic use case dataset to perform performance and efficiency tests on the business image model, and generating stability benchmark indicators, includes:

[0016] Configure the access control policy from the offline testing system to the data center, and open the access interface according to the principle of least privilege;

[0017] Within a specified time window, the offline testing system initiates a basic test case data request to the data center and obtains the basic test case dataset issued by the data center based on the basic test case data request through an encrypted transmission channel.

[0018] The offline testing system performs integrity verification on the basic test case dataset. If the basic test case dataset has not been tampered with, it loads the basic test case dataset into the business image model to perform benchmark tests on performance and efficiency.

[0019] The performance and efficiency benchmark test results are used as the stability benchmark indicators for the business image model.

[0020] Optionally, the step of using the business mirroring model to discriminate new data streams from the production business, identifying adversarial examples and obtaining incremental datasets by matching the similarity between the discriminative patterns and attack features, includes:

[0021] The new data stream is sequentially input into the business mirror model to obtain the classification confidence probability corresponding to each data sample in the new data stream, thereby generating the corresponding discrimination sequence.

[0022] A service supervision model with a two-layer LSTM network structure is established, and the discrimination sequence is input into the service supervision model to enable the service supervision model to learn the discrimination pattern by itself.

[0023] The discrimination pattern features of the discrimination sequence are extracted using a sliding window, and the similarity between the discrimination pattern features and the pre-constructed adversarial attack features is calculated.

[0024] Based on a similarity threshold, adversarial samples in the new data stream are identified, and the new data stream after removing the adversarial samples is integrated into the incremental dataset.

[0025] Optionally, the step of integrating the incremental dataset into the basic use case dataset in stages, and iteratively fine-tuning the parameters of the business mirror model based on the deviation assessment of the stability benchmark metric, includes:

[0026] The incremental dataset is randomly split into multiple data subsets, and each data subset is injected into the basic use case dataset for fusion in stages;

[0027] Several parameters in the business mirror model, sorted by weight value, are frozen. After each injection of the data subset, the business mirror model is iteratively trained to fine-tune the unfrozen parameters in the business mirror model.

[0028] After each round of parameter fine-tuning, the performance and efficiency metrics of the business mirror model are evaluated, and the performance and efficiency deviations are calculated based on the stability benchmark metrics.

[0029] When the performance and efficiency deviation exceeds a preset deviation threshold, the iterative training of the business image model is stopped, and a rollback operation is performed on the business image model after the iteration stops.

[0030] Optionally, the step of synchronously updating the parameters of the online business model and the basic use case dataset based on the fine-tuned business mirror model and the incremental dataset includes:

[0031] The offline testing system generates a configuration file for the business image model, sends a model update request to the online business system, and after receiving confirmation of the request from the online business system, sends the configuration file for the business image model to the online business system.

[0032] Within the allowed business interruption time window, the online business system replaces the configuration file of the online business model on its local machine with the configuration file of the business image model sent by the offline test system, thereby completing the parameter update and reconstruction of the online business model.

[0033] The offline testing system sends a data update request to the data center, and after receiving confirmation of the request from the data center, uploads the incremental dataset to the data center.

[0034] The data center removes a portion of the data from the basic use case dataset in chronological order based on the total number of samples in the incremental dataset, and then stores the incremental dataset into the basic use case dataset after removing the data, so as to ensure that the total number of samples included in the basic use case dataset remains unchanged.

[0035] Another aspect of the present invention provides an industrial control system business model update and iteration system that resists adversarial example attacks, the industrial control system business model update and iteration system that resists adversarial example attacks includes:

[0036] The image building module is used to obtain a business image model of the online business model through runtime environment cloning and model configuration reconstruction;

[0037] The performance testing module is used to acquire the basic use case dataset of the production business, use the basic use case dataset to perform performance and efficiency tests on the business image model, and generate stability benchmark indicators.

[0038] The incremental generation module is used to use the business mirror model to identify new data streams of the production business, and to identify adversarial examples and obtain incremental datasets by matching the similarity between the identification pattern and the attack features.

[0039] The iterative update module is used to integrate the incremental dataset into the basic use case dataset in stages, and to iteratively fine-tune the parameters of the business image model based on the deviation evaluation of the stability benchmark index.

[0040] The data update module is used to synchronously update the parameters of the online business model and the basic use case dataset based on the fine-tuned business mirror model and the incremental dataset.

[0041] Another aspect of the present invention provides an electronic device comprising:

[0042] At least one processor; and,

[0043] A memory that is communicatively connected to at least one processor; wherein,

[0044] The memory stores instructions that can be executed by at least one processor, which enables the at least one processor to execute the aforementioned method for updating and iterating the industrial control business model to resist adversarial sample attacks.

[0045] In another aspect, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned method for updating and iterating an industrial control business model to resist adversarial sample attacks.

[0046] In another aspect, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method for updating and iterating an industrial control business model to resist adversarial sample attacks.

[0047] Compared with existing technologies, this invention introduces an adversarial example detection mechanism during the business model update process, which can effectively avoid the misguidance of the business model caused by adversarial example attacks. At the same time, by using an iterative optimization mechanism to gradually fine-tune the model parameters, it can better monitor the model performance deviation caused by insufficient dataset quality. Attached Figure Description

[0048] One or more embodiments are illustrated by way of example with the corresponding pictures in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0049] Figure 1 The flowchart illustrates a method for updating and iterating an industrial control business model to resist adversarial sample attacks, as provided in one embodiment of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details are presented in the various embodiments of the present invention to facilitate a better understanding of the invention. However, the technical solutions claimed in the present invention can be implemented even without these technical details and with various variations and modifications based on the following embodiments. The division of the various embodiments below is for ease of description and should not constitute any limitation on the specific implementation of the present invention. The various embodiments can be combined with and referenced by each other without contradiction.

[0051] One embodiment of the present invention relates to a method for updating and iterating industrial control business models to resist adversarial sample attacks, the process of which is as follows: Figure 1 As shown, the process includes steps S1 to S5.

[0052] Step S1: Obtain the business mirror model of the online business model by cloning the runtime environment and reconstructing the model configuration.

[0053] For example, step S1 includes: cloning the runtime environment of the online business system, including the operating system version, software dependency package version, and environment variable configuration; building an offline test system using container technology; exporting the configuration file of the online business model; establishing a model structure identical to the online business model in the offline test system based on the configuration file of the online business model to obtain a business image model; and using a computation graph to verify the structural consistency between the business image model and the online business model.

[0054] Specifically, assuming the online business system runs on an Ubuntu operating system, and the online business model is built using the Tensorflow machine learning framework in Python, i.e., the online business model is a Tensorflow model. Step S1 first clones the runtime environment of the online business system. Linux commands are executed on the server of the online business system to extract the operating system's distribution information, including an exported list of software dependencies and system environment variable configurations. Based on the Ubuntu version number, the corresponding system image is selected, and an offline test system is created using Docker technology. All dependent software is then installed in the offline test system environment. Next, Step S1 exports the Tensorflow model's configuration file. The serialized structure of this configuration file contains information such as the connection methods of neural network layers, the number and weights of neurons in each layer, and the shapes of input and output tensors. The same algorithm library is called in the offline test system to reconstruct the online business model structure, i.e., the Tensorflow model structure, to obtain the business image model. Finally, Step S1 generates a computation graph of the online business model and the business image model, and uses a graph isomorphism algorithm to analyze the consistency between the upstream and downstream connection orders of the two models at each node.

[0055] Step S2: Obtain the basic use case dataset for production operations, use the basic use case dataset to perform performance and efficiency tests on the business image model, and generate stability benchmark indicators.

[0056] For example, step S2 includes: configuring the access control policy from the offline testing system to the data center, and opening the access interface according to the principle of least privilege; within a specified time window, using the offline testing system to initiate a basic test case data request to the data center, and obtaining the basic test case dataset issued by the data center based on the basic test case data request through an encrypted transmission channel; the offline testing system performs integrity verification on the basic test case dataset, and if the basic test case dataset has not been tampered with, loads the basic test case dataset into the business image model to perform performance and efficiency benchmark tests; and uses the performance and efficiency benchmark test results as the stability benchmark indicators of the business image model.

[0057] Specifically, in this embodiment, a time-based Access Control List (ACL) rule is configured on the data center's egress firewall, specifying the IP address of the offline testing system and the database port number. The offline testing system establishes a TLS / SSL secure tunnel to the firewall via a client and initiates a basic test case data request to the data center within a specified time window. The data center, based on database access account permissions, compresses the data within the read restriction range and sends it to the offline testing system as a basic test case dataset, along with the corresponding MD5 checksum of the compressed file. The offline testing system uses the MD5 checksum of the compressed file to verify the integrity of the basic test case dataset. After confirming that the data has not been tampered with, the basic test case dataset is loaded into the business image model to perform performance and efficiency benchmark tests. Precision, recall, and other identification accuracy metrics, as well as computational efficiency metrics such as sample processing speed, are statistically analyzed on the basic test case dataset. These metrics are used as stability benchmarks for the business image model.

[0058] Step S3: Use the business mirroring model to identify new data streams in production operations, and identify adversarial examples and obtain incremental datasets by matching the similarity between the identification patterns and attack features.

[0059] For example, step S3 includes: sequentially inputting the new data stream into the business mirror model to obtain the classification confidence probability corresponding to each data sample in the new data stream, thereby generating the corresponding discrimination sequence; establishing a business supervision model with a two-layer LSTM network structure, inputting the discrimination sequence into the business supervision model to enable the business supervision model to learn the discrimination pattern; extracting the discrimination pattern features of the discrimination sequence using a sliding window, and calculating the similarity between the discrimination pattern features and the pre-constructed adversarial attack features; identifying adversarial samples in the new data stream according to the similarity threshold, and integrating the new data stream after removing adversarial samples into an incremental dataset.

[0060] Specifically, in this embodiment, the new data stream is preprocessed according to the format requirements of the basic use case data to obtain the input tensor of the business mirror model. This input tensor is then loaded into the business mirror model for classification and recognition. For each sample, the confidence difference between the top-two categories (i.e., the confidence difference between the first-ranked category and the second-ranked category) is retained. A discriminant sequence is generated according to the temporal order of the samples. A business supervision model with a two-layer LSTM (Long Short-Term Memory) network structure is established using the Tensorflow framework. The first-layer LSTM network structure is used to learn the short-term fluctuation pattern of the discriminant sequence, and the second-layer LSTM network structure is used to capture the long-term prediction pattern of the decision sequence. The input sequence is reconstructed through self-supervised learning under the Seq2Seq (Sequence-to-Sequence) structure. The discriminative pattern features of the discriminative sequence are extracted using a sliding window, and the similarity between the discriminative pattern features and the pre-constructed adversarial attack features is calculated. Based on the similarity threshold, it is determined whether the new data stream within the sliding window conforms to the adversarial sample. If the data within the sliding window conforms to the adversarial sample, the data is removed. The remaining data in the new data stream after the removal of the data are integrated to obtain the incremental dataset.

[0061] Step S4: Integrate the incremental dataset into the basic use case dataset in stages, and iteratively fine-tune the parameters of the business image model based on the deviation assessment of the stability benchmark indicators.

[0062] For example, step S4 includes: randomly splitting the incremental dataset into multiple data subsets, and injecting each data subset into the basic use case dataset for fusion in stages; freezing several parameters in the business mirror model sorted by weight value, and iteratively training the business mirror model after each data subset injection to fine-tune the unfrozen parameters in the business mirror model; evaluating the performance and efficiency indicators of the business mirror model after each round of parameter fine-tuning, and calculating the performance and efficiency deviation based on the stability benchmark indicator; stopping the iterative training of the business mirror model when the performance and efficiency deviation is greater than a preset deviation threshold, and performing a rollback operation on the business mirror model after the iteration stops.

[0063] Specifically, in this embodiment, a stratified sampling method is used to randomly divide the incremental dataset into 10 data subsets, ensuring that the category distribution of each data subset is approximately the same as the overall dataset. These data subsets are then integrated into the basic use case dataset in stages. After each data subset injection, the business mirror model is fine-tuned. The top 30% of parameters in the business mirror model, sorted by weight value, are frozen as key parameters, allowing updates only to the remaining 70% of parameters. A low learning rate and mini-batch mode are used for training to avoid overfitting to the mixed dataset. After each round of parameter fine-tuning, the performance and efficiency metrics of the business mirror model are evaluated. A deviation fluctuation threshold is set based on a stability benchmark, and the relative deviation between the evaluated current performance and efficiency metrics and the stability benchmark is calculated. If the relative deviation is less than the set deviation fluctuation threshold, the injection of incremental dataset data subsets and parameter fine-tuning of the business mirror model continue. If the relative deviation is greater than the set deviation fluctuation threshold, the iteration process is stopped, and a rollback operation is performed on the business mirror model after the iteration stops.

[0064] Step S5: Based on the fine-tuned business mirror model and incremental dataset, synchronize and update the parameters of the online business model and the basic use case dataset.

[0065] For example, step S5 includes: generating a configuration file for the business image model using the offline testing system, sending a model update request to the online business system, and sending the configuration file for the business image model to the online business system after receiving confirmation of the request from the online business system; within the allowed business interruption time window, the online business system replaces the configuration file of the online business model locally with the configuration file of the business image model sent by the offline testing system, completing the parameter update and reconstruction of the online business model; the offline testing system sends a data update request to the data center, and uploads the incremental dataset to the data center after receiving confirmation of the request from the data center; the data center removes some data from the basic use case dataset in chronological order according to the total number of samples in the incremental dataset, and stores the incremental dataset in the basic use case dataset after removing some data, so as to ensure that the total number of samples included in the basic use case dataset remains unchanged.

[0066] Specifically, in this embodiment, after generating the configuration file of the business image model, the offline testing system sends a model update request to the online business system. Upon receiving confirmation of the request from the online business system, it sends the configuration file of the business image model to the online business system. Based on the business operation status, an offline time is set. Within the business interruption time window, the model configuration file of the online business system is updated and replaced, replacing the local configuration file of the online business model with the configuration file of the business image model sent by the offline testing system. Then, the business model is reconstructed and business operation is restored. Simultaneously, the offline testing system sends a data update request to the data center. Upon receiving confirmation of the request from the data center, it uploads the incremental dataset to the data center. The data center updates the basic test case dataset using a differential rotation method. First, based on the total number of samples in the incremental dataset, the earliest time-series data is removed from the basic test case dataset. Then, the incremental dataset is merged and stored in the data pool of the basic test case dataset, ensuring that the total number of samples in the basic test case dataset remains unchanged.

[0067] The industrial control business model update and iteration method for resisting adversarial example attacks provided by the embodiments of the present invention introduces an adversarial example detection mechanism during the business model update process, which can effectively avoid the misguidance of the business model caused by adversarial example attacks. At the same time, by using an iterative optimization mechanism to gradually fine-tune the model parameters, it can better monitor the model performance deviation caused by insufficient dataset quality.

[0068] Another embodiment of the present invention relates to an industrial control business model update and iteration system that resists adversarial sample attacks, including an image building module, a performance testing module, an incremental generation module, an iterative update module, and a data update module.

[0069] The image building module is used to obtain a business image model of the online business model by cloning the runtime environment and reconstructing the model configuration.

[0070] The performance testing module is used to obtain the basic use case dataset of production business, and to perform performance and efficiency tests on the business image model using the basic use case dataset, generating stability benchmark indicators.

[0071] The incremental generation module is used to identify new data streams from production operations using a business mirror model. By matching the similarity between the identification pattern and attack features, it identifies adversarial examples and obtains incremental datasets.

[0072] The iterative update module is used to integrate the incremental dataset into the basic use case dataset in stages, and to iteratively fine-tune the parameters of the business image model based on the deviation assessment of the stability benchmark indicators.

[0073] The data update module is used to synchronously update the parameters of the online business model and the basic use case dataset based on the fine-tuned business mirror model and the incremental dataset.

[0074] The specific implementation method of the industrial control business model update and iteration system for resisting adversarial sample attacks provided in the embodiments of the present invention can be found in the method for updating and iterating the industrial control business model for resisting adversarial sample attacks provided in the embodiments of the present invention, and will not be repeated here.

[0075] The industrial control business model update and iteration system for resisting adversarial example attacks provided by the embodiments of the present invention introduces an adversarial example detection mechanism during the business model update process, which can effectively avoid the misguidance of the business model caused by adversarial example attacks. At the same time, by using an iterative optimization mechanism to gradually fine-tune the model parameters, it can better monitor the model performance deviation caused by insufficient dataset quality.

[0076] Another embodiment of the present invention relates to an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the industrial control business model update and iteration method for resisting adversarial sample attacks described in the above embodiment.

[0077] The memory and processor are connected via a bus, which can include any number of interconnecting buses and bridges, connecting various circuits of one or more processors and memories. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.

[0078] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.

[0079] Another embodiment of the present invention relates to a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the industrial control business model update and iteration method for resisting adversarial sample attacks described in the above embodiments.

[0080] That is, those skilled in the art will understand that all or part of the steps in the methods described in the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0081] Another embodiment of the present invention relates to a computer program product, including a computer program that, when executed by a processor, implements the industrial control business model update and iteration method for resisting adversarial sample attacks described in the above embodiments.

[0082] Those skilled in the art will understand that the above embodiments are specific implementations of the present invention, and in practical applications, various changes can be made in form and detail without departing from the spirit and scope of the present invention.

Claims

1. A method for updating and iterating an industrial control business model to resist adversarial sample attacks, characterized in that, Includes the following steps: By cloning the runtime environment and reconstructing the model configuration, a business mirror model of the online business model is obtained; Obtain the basic use case dataset for production operations, and use the basic use case dataset to perform performance and efficiency tests on the business image model to generate stability benchmark indicators. The business mirroring model is used to identify new data streams in the production business. By matching the similarity between the identification pattern and the attack features, adversarial examples are identified and incremental datasets are obtained. The incremental dataset is integrated into the basic use case dataset in stages, and the parameters of the business image model are iteratively fine-tuned based on the deviation assessment of the stability benchmark index. Based on the fine-tuned business mirror model and the incremental dataset, the parameters of the online business model and the basic use case dataset are updated synchronously.

2. The method for updating and iterating industrial control business models to resist adversarial sample attacks according to claim 1, characterized in that, The process of obtaining a business mirror model of the online business model through runtime environment cloning and model configuration reconstruction includes: Clone the runtime environment of the online business system, including the operating system version, software dependency package version, and environment variable configuration, and use container technology to build an offline testing system; Export the configuration file of the online business model, and based on the configuration file of the online business model, establish the same model structure as the online business model in the offline testing system to obtain the business mirror model; The structural consistency between the business mirror model and the online business model is verified using a computational graph.

3. The method for updating and iterating industrial control business models to resist adversarial sample attacks according to claim 2, characterized in that, The process of acquiring the basic use case dataset for production operations, using the basic use case dataset to perform performance and efficiency tests on the business image model, and generating stability benchmark metrics includes: Configure the access control policy from the offline testing system to the data center, and open the access interface according to the principle of least privilege; Within a specified time window, the offline testing system initiates a basic test case data request to the data center and obtains the basic test case dataset issued by the data center based on the basic test case data request through an encrypted transmission channel. The offline testing system performs integrity verification on the basic test case dataset. If the basic test case dataset has not been tampered with, it loads the basic test case dataset into the business image model to perform benchmark tests on performance and efficiency. The benchmark test results of performance and efficiency are used as the stability benchmark indicators of the business image model.

4. The method for updating and iterating industrial control business models to resist adversarial sample attacks according to claim 1, characterized in that, The process of using the business mirroring model to discriminate new data streams from the production business, identifying adversarial examples and obtaining incremental datasets by matching the similarity between the discriminant patterns and attack features, includes: The new data stream is sequentially input into the business mirror model to obtain the classification confidence probability corresponding to each data sample in the new data stream, thereby generating the corresponding discrimination sequence. A service supervision model with a two-layer LSTM network structure is established, and the discrimination sequence is input into the service supervision model to enable the service supervision model to learn the discrimination pattern by itself. The discrimination pattern features of the discrimination sequence are extracted using a sliding window, and the similarity between the discrimination pattern features and the pre-constructed adversarial attack features is calculated. Based on a similarity threshold, adversarial samples in the new data stream are identified, and the new data stream after removing the adversarial samples is integrated into the incremental dataset.

5. The method for updating and iterating industrial control business models to resist adversarial sample attacks according to claim 1, characterized in that, The process of integrating the incremental dataset into the basic use case dataset in stages, and iteratively fine-tuning the parameters of the business mirror model based on the deviation assessment of the stability benchmark index, includes: The incremental dataset is randomly split into multiple data subsets, and each data subset is injected into the basic use case dataset for fusion in stages; Several parameters in the business mirror model, sorted by weight value, are frozen. After each injection of the data subset, the business mirror model is iteratively trained to fine-tune the unfrozen parameters in the business mirror model. After each round of parameter fine-tuning, the performance and efficiency metrics of the business mirror model are evaluated, and the performance and efficiency deviations are calculated based on the stability benchmark metrics. When the performance and efficiency deviation exceeds a preset deviation threshold, the iterative training of the business image model is stopped, and a rollback operation is performed on the business image model after the iteration stops.

6. The method for updating and iterating industrial control business models to resist adversarial sample attacks according to claim 2, characterized in that, The process of synchronously updating the parameters of the online business model and the basic use case dataset based on the fine-tuned business mirror model and the incremental dataset includes: The offline testing system generates a configuration file for the business image model, sends a model update request to the online business system, and after receiving confirmation of the request from the online business system, sends the configuration file for the business image model to the online business system. Within the allowed business interruption time window, the online business system replaces the configuration file of the online business model on its local machine with the configuration file of the business image model sent by the offline test system, thereby completing the parameter update and reconstruction of the online business model. The offline testing system sends a data update request to the data center, and after receiving confirmation of the request from the data center, uploads the incremental dataset to the data center. The data center removes a portion of the data from the basic use case dataset in chronological order based on the total number of samples in the incremental dataset, and then stores the incremental dataset into the basic use case dataset after removing the data, so as to ensure that the total number of samples included in the basic use case dataset remains unchanged.

7. An industrial control business model update and iteration system that resists adversarial sample attacks, characterized in that, include: The image building module is used to obtain a business image model of the online business model through runtime environment cloning and model configuration reconstruction; The performance testing module is used to acquire the basic use case dataset of the production business, use the basic use case dataset to perform performance and efficiency tests on the business image model, and generate stability benchmark indicators. The incremental generation module is used to use the business mirror model to identify new data streams of the production business, and to identify adversarial examples and obtain incremental datasets by matching the similarity between the identification pattern and the attack features. The iterative update module is used to integrate the incremental dataset into the basic use case dataset in stages, and to iteratively fine-tune the parameters of the business image model based on the deviation evaluation of the stability benchmark index. The data update module is used to synchronously update the parameters of the online business model and the basic use case dataset based on the fine-tuned business mirror model and the incremental dataset.

8. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the industrial control business model update and iteration method for resisting adversarial sample attacks 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 the processor, it implements the industrial control business model update and iteration method for resisting adversarial sample attacks as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the industrial control business model update and iteration method for resisting adversarial sample attacks as described in any one of claims 1 to 6.

Images