Diversified attack behavior identification method and system based on fusion identification model

Abstract

The invention provides a diversified attack behavior recognition method and system based on a fusion recognition model, and the method comprises the steps: constructing an original training data set based on an attack received by a local end, and determining the type of missing training data based on the number of all types of labels in the original training data set; on the basis of a pre-trained data generation model, supplementing the training data of the missing training data category, and updating a local training data set; performing local training on the attack recognition model based on the updated local training data set, uploading model parameters of the trained local model to a cloud, and determining a model weight by the cloud based on the number of training data in the uploaded original training data set of the local model; and carrying out weighted calculation on the uploaded model parameters of each local model based on the model weights, issuing the updated model parameters to each local end to construct an aggregation model, and carrying out attack identification on data received by the local ends based on the aggregation model.

Description

Technical Field

[0001] The present invention relates to the field of information processing technology, and in particular to a method and system for identifying diversified attack behaviors based on a fusion recognition model. Background Art

[0002] As a key national infrastructure, the stability and security of the power grid are directly related to national energy security, economic development and social stability. With the improvement of the informationization and intelligence level of the power grid, the power grid system is becoming increasingly complex, and the risk of cyber attacks is also increasing. By performing attack identification, potential cyber security threats can be discovered and responded to in a timely manner, thereby effectively ensuring the stable operation of the power grid.

[0003] Federated learning is a machine learning setting whose core idea is that multiple participants collaborate to train models without direct access to the original training data. This learning method mainly addresses the issues of data privacy, data security, and data access rights, and is particularly suitable for areas where data is sensitive or subject to strict privacy regulations. In the field of distribution networks, entities such as different distribution network companies do not communicate data or logs with each other, so the ability to identify attacks is poor. Summary of the invention

[0004] In view of this, an embodiment of the present invention provides a method and system for identifying diversified attack behaviors based on a fusion recognition model to eliminate or improve one or more defects in the prior art.

[0005] One aspect of the present invention provides a method for identifying diversified attack behaviors based on a fusion recognition model. The method is based on a distributed architecture, and the distributed architecture includes a cloud and various local terminals. The steps of the method include:

[0006] Construct an original training data set based on the attacks received by the local end, and determine the missing training data categories based on the number of each type of labels in the original training data set;

[0007] Based on the pre-trained data generation model, the training data of the missing training data categories is supplemented to update the local training data set;

[0008] Locally train the attack recognition model based on the updated local training data set, and upload the model parameters of the trained local model to the cloud;

[0009] The cloud aggregates the model parameters of each local model based on the amount of training data in the original training data set of each uploaded local model, obtains the aggregated parameters corresponding to each model parameter, and constructs an aggregated model, and performs attack identification on the data received by the local end based on the aggregated model.

[0010] By adopting the above scheme, this scheme can first analyze the attacks received locally to determine the less commonly received attack types among the currently received attack types, and supplement the less commonly received attack types through pre-trained data generation models, which can improve the averageness of local training and ensure the model's recognition sensitivity to various types of attacks; this scheme further determines the model weights of the attacks suffered by each local model in the cloud. Since each local terminal is subjected to different types of attacks, some local terminals may be attacked that other local terminals have not been subjected to. Based on the model weights, the model parameters of each uploaded local model are weightedly calculated to ensure the comprehensiveness of the obtained model. In the process of attack identification, it can be ensured that each local terminal can obtain a more comprehensive model and ensure recognition accuracy without the need for the local terminals to communicate data with each other.

[0011] In some embodiments of the present invention, in the step of determining the missing training data category based on the number of each type of label in the original training data set, the label categories in the original training data set are counted, and the ratio of the number of training data items in the label category to the total number of training data items is calculated based on the number of training data items under the label category, and the missing training data category is determined based on the ratio of the number of training data items in the label category to the total number of training data items.

[0012] In some embodiments of the present invention, in the step of determining the missing training data category based on the ratio of the number of training data items of the label category to the total number of training data items, the label category with the largest number of training data items is used as the non-missing training data category, and the remaining label categories are used as the missing training data categories.

[0013] In some embodiments of the present invention, in the step of supplementing the training data of the missing training data category based on the pre-trained data generation model, the missing training data category is encoded and input into the pre-trained data generation model, and the data generation model outputs the training data corresponding to the missing training data category.

[0014] In some embodiments of the present invention, in the step of supplementing the training data of the missing training data category based on the pre-trained data generation model and updating the local training data set, the number of training data items in the missing training data category is supplemented to be equal to the number of data items in the non-missing training data category.

[0015] In some embodiments of the present invention, the data generation model is a conditional generative adversarial network model.

[0016] In some embodiments of the present invention, the attack identification model includes a sequentially connected bidirectional long short-term memory network architecture and a Transformer architecture.

[0017] In some embodiments of the present invention, based on the amount of training data in the uploaded original training data set of each local model, the model parameters of each local model are aggregated to obtain the aggregated parameter corresponding to each model parameter, including:

[0018] The model parameters of each local model are aggregated based on the number of training data in the uploaded original training data sets of each local model and the total number of training data in all original training data sets.

[0019] In some embodiments of the present invention, in the step of calculating based on the number of training data in the original training data sets of each uploaded local model and the total number of training data in all original training data sets, the aggregation parameter corresponding to each model parameter is calculated based on the following formula:

[0020]

[0021] Among them, wglobal represents the aggregation parameter, w i Represents the model parameters of each local model, n i It represents the number of training data in the original training data set of each local model, and K represents the number of local models.

[0022] The second aspect of the present invention also provides a diversified attack behavior identification system based on a fusion recognition model, the system includes a computer device, the computer device includes a processor and a memory, the memory stores computer instructions, the processor is used to execute the computer instructions stored in the memory, when the computer instructions are executed by the processor, the system implements the steps implemented by the method described above.

[0023] The third aspect of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps implemented by the aforementioned diversified attack behavior identification method based on a fusion recognition model.

[0024] Additional advantages, purposes, and features of the present invention will be described in part in the following description, and will become apparent to those skilled in the art after studying the following, or may be learned from the practice of the present invention. The purposes and other advantages of the present invention can be specifically pointed out and obtained in the specification and the accompanying drawings.

[0025] Those skilled in the art will appreciate that the objectives and advantages that can be achieved with the present invention are not limited to the above specific description, and the above and other objectives that can be achieved by the present invention will be more clearly understood from the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The drawings described herein are used to provide a further understanding of the present invention, constitute a part of this application, and do not constitute a limitation of the present invention.

[0027] Figure 1 This is a schematic diagram of an implementation method of the diversified attack behavior identification method based on the fusion identification model of this scheme;

[0028] Figure 2 This is a schematic diagram of the architecture of this solution. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments and the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0030] It should also be noted that, in order to avoid obscuring the present invention due to unnecessary details, only structures and / or processing steps closely related to the solutions according to the present invention are shown in the accompanying drawings, while other details that are not closely related to the present invention are omitted.

[0031] In the prior art, there are currently many methods for identifying attack behaviors. Among them, the most widely used in the field of network security is the use of signature analysis, especially in intrusion detection systems (IDS). IDS maintains a database containing known attack features (or "signatures"). For XSS attacks, these signatures may include specific script fragments, malicious HTML tags, or other executable code. IDS monitors data packets entering and leaving the network and checks whether these data packets contain content that matches known XSS attack signatures. Once it is detected that the content of the data packet matches an entry in the signature database, the system will immediately issue an alarm and take defensive measures according to the configuration, such as blocking the traffic or isolating it. However, it cannot understand and utilize the contextual information in the log files well, and does not play a very good role in complex attack patterns.

[0032] Another method is to use natural language processing tools to analyze communication content or extract key information from log files, which can be relatively easy to maintain and adjust by updating dictionaries and rules. However, it is not suitable for processing large-scale data or effectively processing time series data.

[0033] In the current distribution network system, the common method of attack behavior identification mainly relies on matching known attack signatures or feature codes. Although this method is highly efficient in identifying known types of attacks, it has the following significant problems when facing new or variant forms of attacks, especially in scenarios such as distribution networks that have extremely high requirements for data privacy and security:

[0034] In the distribution network, data is usually distributed in different geographical locations, involving multiple data management nodes. This distributed data storage structure requires the security system to be able to perform effective attack detection without centralizing the data, so as to reduce the risk of data leakage during transmission.

[0035] Existing technologies rely on signature libraries of known attacks, which often makes them powerless when faced with new or unknown attack patterns. Distribution grid systems especially need to be able to foresee and identify attack behaviors that have not yet appeared in any signature database.

[0036] The technical effects of this solution include:

[0037] 1. Enhanced data privacy protection: Through federated learning, the model is allowed to learn independently at each node, and only the model update is shared instead of the data itself, thereby protecting sensitive data from being leaked.

[0038] 2. Adapt to distributed data storage: The model design will fully adapt to the distributed data structure of the distribution network to ensure effective real-time monitoring and attack identification in a distributed environment.

[0039] 3. Improve the ability to detect new and unknown attacks: Combining the time series data processing advantages of the bidirectional long short-term memory network (BiLSTM) and the efficient context analysis capabilities of Transformer, the model will be able to identify complex attack patterns, including those that are difficult to identify with traditional methods.

[0040] Through these innovations, this solution aims to provide a privacy-preserving and efficient cybersecurity solution for distribution networks, enhancing their ability to resist various cyber threats, especially in the face of increasingly complex cyber attack environments.

[0041] like Figure 1 and 2 As shown, the present invention proposes a diversified attack behavior identification method and system based on a fusion identification model. The method is based on a distributed architecture, which includes a cloud and various local terminals. The steps of the method include:

[0042] Step S100, constructing an original training data set based on the attack received by the local end, and determining the missing training data category based on the number of each type of label in the original training data set;

[0043] In the specific implementation process, during the online application process, the local end collects data on attacks received over a period of time, and constructs an original training dataset based on the actual categories of attacks received and the actual response of the local model.

[0044] During the specific implementation process, each local terminal continuously collects log data related to attack behaviors. Each local terminal performs operations such as denoising and normalization on the collected data to ensure that the input format of the data is consistent.

[0045] Step S200, supplementing the training data of the missing training data category based on the pre-trained data generation model to update the local training data set;

[0046] In the specific implementation process, each local end is prone to insufficient data or low frequency of attack behaviors, which will make it difficult to effectively identify the characteristics of attack behaviors during local model training. Therefore, a data generation model is used to simulate and expand attack behavior data, enhance the generalization ability of the model, and reduce the impact of non-independent and identically distributed data on model training. In the data generation process, a differential privacy mechanism is introduced to ensure that the generated simulated data will not leak the key data characteristics of the subsidiary by adjusting the model parameters, thereby protecting data privacy.

[0047] Step S300, locally training the attack recognition model based on the updated local training data set, and uploading the model parameters of the trained local model to the cloud;

[0048] In the specific implementation process, after data preprocessing and generated data supplementation, the local end uses the existing real data and generated data to train the local model and output the parameters of the local model training (model weights, biases, etc.). This information will be used for the next step of global aggregation.

[0049] After completing local model training, each local end uploads the model parameters to the central aggregation server in the cloud. To ensure data privacy, the model parameters will be encrypted or noise will be added before uploading. After the central server receives the model parameters uploaded by each local end, it uses the federated learning model aggregation algorithm to perform a weighted average of the models of each subsidiary. The algorithm dynamically adjusts the weights based on the amount of data provided by each local end and the model performance.

[0050] In step S400, the cloud aggregates the model parameters of each local model based on the amount of training data in the original training data set of each local model uploaded, obtains the aggregated parameters corresponding to each model parameter, and constructs an aggregated model, and performs attack identification on the data received by the local end based on the aggregated model.

[0051] In the specific implementation process, the cloud sends the updated global model to each local terminal. After receiving the global model, each local terminal continues to fine-tune its local data to ensure that the model can adapt to the subsidiary's own special data distribution. This process continuously optimizes the performance of the global model through multiple rounds of iterations.

[0052] By adopting the above scheme, this scheme can first analyze the attacks received locally to determine the less commonly received attack types among the currently received attack types, and supplement the less commonly received attack types through pre-trained data generation models, which can improve the averageness of local training and ensure the model's recognition sensitivity to various types of attacks; this scheme further determines the model weights of the attacks suffered by each local model in the cloud. Since each local terminal is subjected to different types of attacks, some local terminals may be attacked that other local terminals have not been subjected to. Based on the model weights, the model parameters of each uploaded local model are weightedly calculated to ensure the comprehensiveness of the obtained model. In the process of attack identification, it can be ensured that each local terminal can obtain a more comprehensive model and ensure recognition accuracy without the need for the local terminals to communicate data with each other.

[0053] In some embodiments of the present invention, in the step of determining the missing training data category based on the number of each type of label in the original training data set, the label categories in the original training data set are counted, and the ratio of the number of training data items in the label category to the total number of training data items is calculated based on the number of training data items under the label category, and the missing training data category is determined based on the ratio of the number of training data items in the label category to the total number of training data items.

[0054] In some embodiments of the present invention, in the step of determining the missing training data category based on the ratio of the number of training data items of the label category to the total number of training data items, the label category with the largest number of training data items is used as the non-missing training data category, and the remaining label categories are used as the missing training data categories.

[0055] In some embodiments of the present invention, in the step of supplementing the training data of the missing training data category based on the pre-trained data generation model, the missing training data category is encoded and input into the pre-trained data generation model, and the data generation model outputs the training data corresponding to the missing training data category.

[0056] In some embodiments of the present invention, in the step of supplementing the training data of the missing training data category based on the pre-trained data generation model and updating the local training data set, the number of training data items in the missing training data category is supplemented to be equal to the number of data items in the non-missing training data category.

[0057] In some embodiments of the present invention, the data generation model is a conditional generative adversarial network model.

[0058] In some embodiments of the present invention, the attack identification model includes a sequentially connected bidirectional long short-term memory network architecture and a Transformer architecture.

[0059] In the specific implementation process, this solution is based on the bidirectional long short-term memory network (BiLSTM) and federated learning attack behavior recognition method technology, and at the same time uses Transformer to optimize the network structure and reduce the impact of non-independent and identically distributed training data on the model training accuracy. The front-end data generation model.

[0060] This paper proposes a composite model based on deep learning, combining BiLSTM and Transformer technology, for processing and analyzing time series data. The system adopts a federated learning framework, which can achieve collaborative training across multiple data nodes while protecting privacy.

[0061] In the specific implementation process, since the log data of each subsystem of the distribution network is heterogeneous and unevenly distributed, and the data privacy requirements between subsidiaries are high, the data generation model of this solution adopts the conditional generative adversarial network based on differential privacy as the data generation model. Its main purpose is to protect the privacy of each subsidiary during the data generation process, while ensuring that the generated data has good quality and practicality. The differential privacy is shown in the following formula. Where A represents the generator or discriminator; D and D ′ are two similar data sets; ∈ represents the privacy budget, which is used to quantify the privacy protection degree of the algorithm. The larger it is, the smaller the added noise is, the higher the accuracy of the model is, but the privacy protection is weaker; δ represents the failure probability, which is used to indicate the probability that differential privacy protection may not hold.

[0062] P[A(D)∈S]≤e ∈P [A(D ′ )∈S++δ

[0063] The function of the generator is to generate realistic data to deceive the discriminator. In this paper, noise protection is added through the differential privacy mechanism, as shown in formula (2). z is the random noise extracted from the prior distribution p(z); D is the discriminator, G(z) is the data generated by the generator, and DP is the noise term used to protect privacy.

[0064] L G =E z～p(z) [log(1-D(G(z)))++DP

[0065] The role of the discriminator is to distinguish between real data and generated data, and also introduce differential privacy noise, as shown in formula (3). G(z) and DP have the same meaning as above. data Represents data sampled from the real data distribution, and DP is the differential privacy noise added during gradient update.

[0066]

[0067] In this paper, differential privacy noise is generated through Gaussian distribution, as shown in formula (4), where σ is the standard deviation of the noise and δ is the failure probability. When the discriminator performs gradient updates, the added noise ensures that the original data will not be leaked.

[0068]

[0069] Attack behavior data includes network traffic with timestamps and its corresponding attack type. Attack behavior evolves over time, and the input data is arranged in chronological order to capture the characteristics of attacks at different time points. These data are input into the training model of this solution in the format of time series.

[0070] BiLSTM focuses on how past attack behavior characteristics affect subsequent attack behaviors, and processes time series data from the forward direction in an attempt to capture the development trend of network attacks. In the reverse direction, the suspicious characteristics of previous time series data are inferred through subsequent attack behaviors, and the attack characteristics that were difficult to detect in the early stage are identified through the later attack results. By merging bidirectional hidden states (hidden states are used to store information about the current moment of the network and pass it to the next step), the training model can simultaneously capture the dependencies from the past to the future and from the future to the past in the time series. For example, a DDos attack only causes a small disturbance in network activity at the beginning of the time series, but in the process of reverse analysis, the training model can identify the connection between the later traffic explosion and the previous network fluctuations.

[0071]

[0072] Cyber ​​attack behaviors usually have global patterns across time, which makes it impossible to detect attack features in local time series. The self-attention mechanism introduced by Transformer can solve this problem by calculating the correlation between different time steps, as shown in the following formula:

[0073]

[0074] Q stands for the query matrix, which is used to find other relevant time steps or input features at each time step. The query vector of each time step will be compared with all other time steps to find the most relevant part. K is the key matrix, which is used to calculate the similarity with the query. The similarity between the key and the query determines the weight of the attention allocation. V stands for the value matrix, which represents the feature information required in the final output. Unlike BiLSTM, which focuses on the direct order of time series, it can also capture complex relationships between different time points, such as abnormal behavior of a node attacking other network nodes at other time points. The global dependency information extracted by the Transformer layer is sent to the fully connected layer for mapping to specific attack behavior classifications to complete the identification of attack behaviors.

[0075] By combining the local dependency of BiLSTM with the global dependency of Transformer, the model can not only capture the neighboring dependency of each time step, but also capture the long-range dependency in the entire time series from a more macro perspective. In this way, the model can more accurately focus on the details and global information in the time series at the same time.

[0076] In actual data, time series may contain some discontinuous points or noise data. The memory and gate mechanism of BiLSTM can filter noise to a certain extent. The self-attention mechanism of Transformer is based on global weight allocation, so noise data is sometimes also assigned weights. Therefore, after combining BiLSTM, Transformer can assign weights to filtered data information, thereby reducing the impact of noise data on the training process.

[0077] The computational overhead in the training process is also a non-negligible part. BiLSTM has lower computational overhead than Transformer. The complexity of the self-attention mechanism is O(n2), which is proportional to the square of the length of the sequence to be processed. Therefore, when the amount of data is large, the computational and storage requirements of Transformer will increase significantly. In the early stage of training the model, BiLSTM is first used to capture local dependencies. In this way, the time series information input to Transformer has already extracted local dependency information. Transformer only needs to further capture global dependencies, which effectively reduces the amount of computation.

[0078] In this paper, in order to meet the high data privacy requirements and difficult data sharing requirements among the distribution network subsidiaries, federated learning provides a suitable solution that enables companies to achieve unified attack behavior identification through local model training and global model aggregation. In order to further enhance privacy protection, this technical solution combines differential privacy technology and CGAN. Differential privacy noise is added to the model parameters uploaded to the server to ensure that even if the model parameters are intercepted, local sensitive data cannot be inferred. At the same time, CGAN is used to simulate non-independent and identically distributed data generation to further enhance the model's generalization ability for different data distribution scenarios.

[0079] In order to make the global performance more stable under non-IID, the proximal term constraints are added during local training. The optimization goal is as follows: the client should not only minimize the local loss during training, but also control the model parameters and the global model parameters w global This proximal term limits the parameters of the client model from deviating too far from the global model. i (w) represents the loss of local data of client i, w i is the client local model parameter, w global represents the global model parameter. μ is the proximal term coefficient. A larger μ value will make the local model closer to the global model, while a smaller μ value will allow more local deviations.

[0080]

[0081] In some embodiments of the present invention, based on the amount of training data in the uploaded original training data set of each local model, the model parameters of each local model are aggregated to obtain the aggregated parameter corresponding to each model parameter, including:

[0082] The model parameters of each local model are aggregated based on the number of training data in the uploaded original training data sets of each local model and the total number of training data in all original training data sets.

[0083] In some embodiments of the present invention, in the step of calculating based on the number of training data in the original training data sets of each uploaded local model and the total number of training data in all original training data sets, the aggregation parameter corresponding to each model parameter is calculated based on the following formula:

[0084]

[0085] Among them, wglobal represents the aggregation parameter, w1, w2 and w n Respectively represent the same location model parameters of each local model, n1, n2 and n nRespectively represent the number of training data in the original training data set of each local model.

[0086] Adopting the above scheme, this scheme applies the federated training mechanism. Federated learning is a machine learning setting whose core idea is that multiple participants collaborate to train the model without directly accessing the original training data. This learning method mainly solves the problems of data privacy, data security and data access rights, and is especially suitable for areas where data is sensitive or needs to comply with strict privacy regulations. In the field of distribution network, various entities (such as different distribution network companies) may not communicate data or logs for the privacy and security of data. Therefore, the use of the federated learning mechanism in the training process can ensure that multiple entities can jointly train data without directly exchanging data.

[0087] Under such a training mechanism, data analysis and model training steps can be performed without leaking detailed user electricity consumption data and other sensitive information. Secondly, as a critical infrastructure, it is crucial to protect the distribution network system from cyber attacks. Through federated learning, model updates can be shared among multiple nodes instead of raw data. Therefore, different distribution network operators can collaborate to improve the accuracy and reliability of the model without worrying about the leakage of sensitive data.

[0088] BiLSTM is a network structure in deep learning. This structure combines two LSTM layers into one unit. One LSTM layer is responsible for the information flow of the forward time step, and the other is responsible for the information flow of the reverse time step. This captures the context information in the time series data. This structure is particularly suitable for sequence data analysis.

[0089] In summary, in multiple rounds of federated learning iterations, this solution can dynamically adapt to the latest attack behavior patterns. Through periodic local model training iterations and central model aggregation, this training mechanism can capture new attack behavior features in a timely manner and improve the ability to detect attack behaviors.

[0090] The training mechanism in this paper combines federated learning with attack behavior identification in the distribution network environment. Considering that data cannot be directly shared between local terminals and that the log data of each system is heterogeneous and unevenly distributed, a pre-data generation model and a federated training framework are adopted to maintain data privacy and improve the generalization ability of the global attack identification model through a federated aggregation strategy.

[0091] In order to improve the resolution of attack data, a network architecture combining BiLSTM and Transformer is used to effectively capture local and global dependency information in time series. It can not only improve the recognition ability of instantaneous and latent attack behaviors by combining previous and subsequent information, but also identify complex attack patterns based on the self-attention mechanism, which is especially effective for large-scale distributed attacks. The simulation experiment results show that the F1 score difference is smaller than that of the centralized training mechanism, but it significantly improves the data privacy security and recognition generalization ability of each local end.

[0092] An embodiment of the present invention also provides a diversified attack behavior identification system based on a fusion recognition model, the system includes a computer device, the computer device includes a processor and a memory, the memory stores computer instructions, the processor is used to execute the computer instructions stored in the memory, when the computer instructions are executed by the processor, the system implements the steps implemented by the method described above.

[0093] The embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps implemented by the aforementioned diversified attack behavior identification method based on the fusion identification model are implemented. The computer-readable storage medium can be a tangible storage medium, such as a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a floppy disk, a hard disk, a removable storage disk, a CD-ROM, or any other form of storage medium known in the technical field.

[0094] It should be understood by those skilled in the art that the exemplary components, systems and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software or a combination of the two. Whether it is performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of the present invention are programs or code segments used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link via a data signal carried in a carrier.

[0095] It should be clear that the present invention is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present invention.

[0096] In the present invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with features of other embodiments or replace features of other embodiments.

[0097] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the embodiments of the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A diversified attack behavior recognition method based on a fusion recognition model, characterized in that: The method is based on a distributed architecture, which includes a cloud and various local terminals. The steps of the method include: Construct an original training data set based on the attacks received by the local end, and determine the missing training data categories based on the number of each type of labels in the original training data set; Based on the pre-trained data generation model, the training data of the missing training data categories is supplemented to update the local training data set; Locally train the attack recognition model based on the updated local training data set, and upload the model parameters of the trained local model to the cloud; The cloud aggregates the model parameters of each local model based on the amount of training data in the original training data set of each uploaded local model, obtains the aggregated parameters corresponding to each model parameter, and constructs an aggregated model, and performs attack identification on the data received by the local end based on the aggregated model.

2. The diversified attack behavior recognition method based on the fusion recognition model according to claim 1 is characterized in that: In the step of determining the missing training data category based on the number of each type of label in the original training data set, the label categories in the original training data set are counted, and the ratio of the number of training data items in the label category to the total number of training data items is calculated based on the number of training data items under the label category, and the missing training data category is determined based on the ratio of the number of training data items in the label category to the total number of training data items.

3. The diversified attack behavior recognition method based on the fusion recognition model according to claim 2 is characterized in that: In the step of determining the missing training data category based on the ratio of the number of training data items of the label category to the total number of training data items, the label category with the largest number of training data items is used as the non-missing training data category, and the remaining label categories are used as the missing training data categories.

4. The diversified attack behavior recognition method based on the fusion recognition model according to claim 3 is characterized in that: In the step of supplementing the training data of the missing training data category based on the pre-trained data generation model, the missing training data category is encoded and input into the pre-trained data generation model, and the data generation model outputs the training data corresponding to the missing training data category.

5. The diversified attack behavior recognition method based on the fusion recognition model according to claim 4 is characterized in that: In the step of supplementing the training data of the missing training data category based on the pre-trained data generation model and updating the local training data set, the number of training data items in the missing training data category is supplemented to be equal to the number of data items in the non-missing training data category.

6. The diversified attack behavior recognition method based on the fusion recognition model according to claim 1 is characterized in that: The data generation model is a conditional generative adversarial network model.

7. According to claim 1, the diversified attack behavior recognition method based on the fusion recognition model is characterized in that: The attack identification model includes a sequentially connected bidirectional long short-term memory network architecture and a Transformer architecture.

8. The diversified attack behavior recognition method based on a fusion recognition model according to any one of claims 1 to 7, characterized in that: The steps of aggregating the model parameters of each local model based on the amount of training data in the uploaded original training data set of each local model to obtain the aggregated parameter corresponding to each model parameter include: The model parameters of each local model are aggregated based on the number of training data in the uploaded original training data set of each local model and the total number of training data in all the original training data sets.

9. The diversified attack behavior recognition method based on fusion recognition model according to claim 8 is characterized in that: In the step of calculating based on the amount of training data in the original training data sets of each uploaded local model and the total amount of training data in all the original training data sets, the aggregation parameter corresponding to each model parameter is calculated based on the following formula: Among them, wglobal represents the aggregation parameter, w i Represents the model parameters of each local model, n i It represents the number of training data in the original training data set of each local model, and K represents the number of local models.

10. A diversified attack behavior recognition system based on a fusion recognition model, characterized in that: The system includes a computer device, which includes a processor and a memory. The memory stores computer instructions. The processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps implemented by the method as described in any one of claims 1 to 9.

Images