Unmanned aerial vehicle network intrusion detection multi-classification method based on knowledge distillation

Abstract

The application provides a UAV network intrusion detection multi-classification method based on knowledge distillation, first, a dataset composed of multiple groups of UAV network traffic data containing normal samples and multiple attack samples is acquired; communication features are extracted from the dataset, a directed multi-attribute graph G based on application layer protocol enhancement is constructed, a training graph and a test graph are obtained; the training graph is used to train a teacher model, forward reasoning is performed on the training graph and the test graph through the trained teacher model, training embedding vectors and test embedding vectors are obtained; the training embedding vectors are used as the input of a student model, knowledge distillation is realized through joint optimization of a hard label loss function and a soft label loss function, and a trained student model is obtained; the trained student model is used for reasoning, the test embedding vectors are used as the input, and prediction labels are output. The method can still obtain effective expression without labeled sample types, and significantly reduces reasoning cost under the condition of ensuring detection performance.

Description

Technical Field

[0001] This application relates to the technical field of drone network intrusion detection methods, specifically to a multi-classification method for drone network intrusion detection based on knowledge distillation. Background Technology

[0002] With the rise of the low-altitude economy, the application of drones in urban logistics, environmental monitoring, emergency rescue, and public safety is expanding. The communication network of drones is the core support for low-altitude airspace operations. Low-altitude economy drone networks typically consist of aerial drones, ground base stations, edge nodes, and a back-end control center. Their communication link structure is complex, and due to the open operating environment, they exhibit high mobility and frequent interactions. In such network environments, communication links and nodes are highly vulnerable to attacks, including denial-of-service attacks, spoofing attacks, traffic hijacking, and other malicious intrusion behaviors. Once an attack occurs, it can lead to drone network service interruptions, mission disruptions, data leaks, aircraft control failures, and even serious threats to public safety. Therefore, intrusion detection for drone networks is not only a necessary means to ensure communication security but also a fundamental technology supporting the healthy development of the low-altitude economy.

[0003] Existing drone network intrusion detection methods mainly include traditional machine learning methods based on feature engineering and detection methods based on deep learning. Traditional machine learning methods based on feature engineering manually extract traffic statistics features or system behavior indicators and combine them with classification models such as support vector machines, random forests, and K-nearest neighbors to achieve anomaly identification. Deep learning-based detection methods use deep learning models such as autoencoders, LSTM, CNN, and GNN, which can automatically extract high-dimensional features and capture nonlinear attack patterns, especially excelling in modeling spatiotemporal dependencies and graph structure information.

[0004] However, existing methods have the following shortcomings:

[0005] (1) Although existing deep learning models can achieve high detection accuracy in conventional network environments, their model parameters are large and their floating-point operations are high, which significantly exceeds the computing power of UAV embedded platforms. These models often have to rely on external GPUs or cloud computing resources to run, making it difficult to deploy them directly on UAV terminals with limited power consumption, storage and load, and failing to meet the technical requirements of UAV networks for lightweight and low-latency inference.

[0006] (2) Most existing methods abstract network traffic or host behavior into independent feature vectors, lacking the ability to model the structural dependencies and association patterns between nodes. They are difficult to effectively reflect the global and local spatial features of UAV networks, including point-to-point interaction, multi-node collaboration, and link dependencies, resulting in limited model accuracy and generalization ability.

[0007] (3) Existing models often rely on large-scale, high-quality labeled data for training. However, in actual UAV networks, attack samples are scarce and difficult to label in a timely manner. Real attack behaviors are covert and diverse, and also have dynamic evolution characteristics. It is difficult to fully cover them with a small number of labels, resulting in existing models being inadequate in dealing with emerging threats and prone to problems such as missed detections and poor adaptability to unknown attacks. Summary of the Invention

[0008] The purpose of this invention is to provide a multi-classification method for UAV network intrusion detection based on knowledge distillation. By using knowledge distillation technology and edge-level unsupervised graph learning, the method reduces the computational complexity of the inference stage while ensuring the accuracy of intrusion detection, thereby improving the efficiency of the model in real-time detection in UAV network environments.

[0009] This invention provides a multi-classification method for UAV network intrusion detection based on knowledge distillation, comprising the following steps:

[0010] Step S1: Obtain a dataset consisting of multiple sets of drone network traffic data, including normal samples and various attack samples;

[0011] Step S2: Extract communication features from the dataset, generate corresponding edge feature vectors for each traffic data, assign auxiliary identifiers and sample type labels, then divide the dataset into training set and test set, and construct a directed multi-attribute graph G based on application layer protocol enhancement to perform graph modeling on the traffic data in the training set and test set, transforming the interaction relationship between communication entities into a graph structure to obtain training graph and test graph.

[0012] Step S3: Build a neural network model, which includes a teacher model and a student model. The teacher model is based on the edge-level unsupervised graph learning method DGI, and uses the graph neural network E-GraphSAGE as the encoder for message passing and feature aggregation. The student model is a multilayer perceptron MLP, which consists of N fully connected layers and N-1 sets of optional normalized nonlinear units cascaded alternately, where N≥1.

[0013] Step S4: Train the teacher model using the training graph, and then use the trained teacher model to perform forward inference on the training graph and the test graph respectively to obtain the training embedding vector and the test embedding vector respectively.

[0014] Step S5, student model training stage: The student model takes the training embedding vector output by the teacher model as input, and obtains a well-trained student model through joint optimization of hard label loss function and soft label loss function.

[0015] Step S6: Use the trained student model for inference. Take the test embedding vector generated by the teacher model as input, and the student model outputs the test prediction label. Compare it with the real label in the test graph and calculate the values ​​of each index to evaluate the performance of drone network intrusion detection.

[0016] Furthermore, step S2 includes the following steps:

[0017] Step S21: First, the fields of communication features in the traffic data of the UAV network are processed into strings to ensure the consistency of the field features when used as node identifiers in the attribute graph; at the same time, infinite values ​​and missing values ​​in the dataset are replaced and filled, replacing infinite values ​​with 0 and filling missing values ​​with 0 to ensure the stability of subsequent feature processing.

[0018] The communication characteristics include source address, destination address, source port, and destination port;

[0019] Step S22: Group the traffic data in the dataset according to the sample type label, and perform random downsampling to reduce computing power requirements to form a training set and a test set; the training set and the test set are divided in a 7:3 ratio;

[0020] Step S23: The field features in the dataset are converted into numerical columns using target encoding, and the numerical columns are scaled using normalization. The multidimensional features of each traffic data are merged into an edge feature vector h and stored in the attribute label to obtain the original edge set.

[0021] The field features include address protocol field, application layer protocol, cumulative TCP flag value, cumulative client TCP flag value, cumulative server TCP flag value, ICMP type and code combination value, ICMP type value, DNS query transaction identifier, DNS query type, and FTP client command return code.

[0022] Step S24: Before constructing the attribute graph, the source address nodes of the edges with the same protocol value are paired up by the application layer protocol field. According to the preset rules, auxiliary edges are added to the potential communication relationship between two source address nodes that do not have the original edge, and the auxiliary edge set is obtained as a structural supplement to the original edge set to enhance the structural integrity and semantic relevance of the UAV network traffic attribute graph.

[0023] Step S25: Assign auxiliary identifiers to the traffic data in the dataset. The value of the auxiliary identifier is 0 or 1. A value of 0 indicates that the traffic data belongs to normal samples, and a value of 1 indicates that the traffic data belongs to attack samples. During the construction of the attribute graph, the sample type label of the attack sample is converted into an integer code by the label encoder and stored as the edge attribute of the attribute graph.

[0024] After merging the original edge set and the auxiliary edge set, the source address and destination address are used as the start and end points of the attribute graph. Edge feature vectors h, sample type labels, and auxiliary identifiers are carried as edge attributes on the edges from the start to the end point of the attribute graph. The attribute graph is modeled using the multi-graph data structure in the NetworkX graph computing library. The multi-graph data structure supports multiple parallel edges between the same pair of nodes, resulting in an undirected attribute graph. The undirected attribute graph is then converted into a directed attribute graph using the to_directed function to reflect the directionality of communication. Finally, the UAV network traffic data is transformed into a graph object that can be processed by the Deep Graph Learning Library (DGL) in graph neural networks, obtaining a directed multi-attribute graph G based on application layer protocol enhancement, thus obtaining the training and test graphs. Subsequently, nodes in the directed multi-attribute graph G based on application layer protocol enhancement are assigned all-one placeholder feature vectors consistent with the edge feature dimension to meet the input requirements of the encoder in the subsequent neural network model.

[0025] Further, in step S3, the structure of the optional normalized nonlinear units in the student model is determined by the parameter norm_type. When norm_type=batch, all optional normalized nonlinear units in the student model consist of batch normalization, ReLU activation function, and regularization operation; when norm_type=layer, all optional normalized nonlinear units in the student model consist of layer normalization, ReLU activation function, and regularization operation; when norm_type=none, all optional normalized nonlinear units in the student model consist of ReLU activation function and regularization operation.

[0026] In the student model, the first fully connected layer performs a linear transformation from the input dimension to the hidden dimension. The middle fully connected layers and normalized nonlinear units perform a constant-dimensional linear transformation from the hidden dimension to the hidden dimension. The last fully connected layer performs a linear transformation from the hidden dimension to the output dimension. Except for the last layer, after each fully connected layer performs a linear transformation, it sequentially performs an optional normalization layer, a ReLU activation function, and a regularization operation in the optional normalized nonlinear unit, forming a normalization-to-nonlinear-to-regularization sequence. The last layer only performs the linear transformation of the fully connected layer. The optional normalization layer is determined by the parameter norm_type, and performs batch normalization, layer normalization, or no normalization operation.

[0027] Further, step S4 includes the following steps:

[0028] Step S41: Using the training graph as input, the training graph is used as a positive sample of the training graph. A corrupted graph is created as a negative sample of the training graph through a corruption function. The positive and negative samples of the training graph are input into the encoder. The encoder outputs the positive embedding of the positive sample of the training graph and the negative embedding of the negative sample of the training graph. The Readout function is used to take the mean of all positive sample embeddings. The Sigmoid activation function is used to compress the range of the mean of the positive sample embeddings to between 0 and 1, thereby generating a global graph summary S containing all the semantics of the graph.

[0029] Step S42: The discriminator D uses a bilinear scoring method to calculate positive and negative sample scores. Specifically, positive sample scores are obtained by calculating the dot product similarity between the positive sample embedding and the global graph summary S, and negative sample scores are obtained by calculating the dot product similarity between the negative sample embedding and the global graph summary S. Then, the obtained positive sample scores and labels with a value of 1, and negative sample scores and labels with a value of 0 are fed into the binary cross-entropy loss function. The binary cross-entropy loss function measures the degree to which positive sample scores tend to 1 and negative sample scores tend to 0, forming the discriminant loss.

[0030] Step S43: Backpropagate the discriminant loss and update the weights of the discriminator D and the parameters of the encoder simultaneously to complete the training and parameter optimization of the teacher model.

[0031] Step S44: Perform a forward inference operation on the training graph and the test graph respectively to obtain the training embedding vector and the test embedding vector respectively. The training embedding vector is transformed by the linear mapping layer to obtain the log odds of the teacher model training, which is used for the training of the student model. The test embedding vector is used for the inference of the student model.

[0032] Further, step S5 includes the following steps:

[0033] The training embedding vectors output by the teacher model are used as input to the student model for training. The log-odds ratio of the student model's training is then output. The hard-label loss is calculated using the log-odds ratio of the student model's training and the ground truth labels on the training graph. Simultaneously, the probability distribution obtained by applying a Softmax activation function to the log-odds ratio of the student model's training is combined with the probability distribution obtained by applying a Softmax activation function to the log-odds ratio of the teacher model's training to calculate the soft-label loss. The soft-label loss measures the difference between the student model's output distribution and the projected soft-label distribution of the teacher model, constraining the student model. The Adam optimizer is then used for backpropagation and parameter updates on the student model. Finally, the student model is jointly optimized using the hard-label loss function and the soft-label loss function to obtain the trained student model.

[0034] Further, step S6 includes the following steps:

[0035] Using the test embedding vector generated by the trained teacher model as input, the pre-trained student model is used for inference, and the log odds of the student model's test are output. The log odds are then converted into a test probability distribution using the Softmax activation function, and the sample type label corresponding to the highest probability is output as the test prediction label. Subsequently, the student model accesses the true label on the test map and compares it with the converted test probability distribution. The multi-class performance of the model is evaluated by precision, recall, and F1 score.

[0036] This method first transforms UAV network traffic data into a communication graph structure within the UAV network. A graph representation learning model is used as the teacher model, and a lightweight multilayer perceptron (MLP) is used as the student model. Embedding vectors of the training and test graphs are extracted from the trained teacher model. Cross-entropy loss and distillation loss are introduced during the training of the student model to ensure the student model fully learns the teacher's knowledge. This approach ensures intrusion detection accuracy while effectively reducing the computational complexity of model inference, adapting to the power consumption, storage, and computing power limitations of terminal devices in the UAV network environment, and exhibiting excellent lightweight deployment characteristics.

[0037] Based on the above-described knowledge distillation-based multi-classification method for UAV network intrusion detection, this invention provides a UAV network intrusion detection device, which includes one or more processors and a memory; the memory is coupled to the one or more processors, and the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to implement the knowledge distillation-based multi-classification method for UAV network intrusion detection.

[0038] In addition, the present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the aforementioned knowledge distillation-based multi-classification method for UAV network intrusion detection.

[0039] The beneficial technical effects of this invention are as follows:

[0040] This invention constructs a three-stage lightweight multi-classification model for UAV network intrusion detection. Unlike existing technologies that mainly focus on grayscale images or use convolutional processing for modeling, this invention innovatively performs directed multi-attribute graph modeling based on application layer protocol enhancement. After dividing the UAV network into multiple protocol domain subsets, auxiliary edge enhancement is performed, which effectively compensates for the topology sparsity or missing parts caused by link jitter, service interruption or session discontinuity in UAV communication links. Furthermore, at the teacher model level, most existing solutions employ convolutional or attention structures. This invention uses an edge-level unsupervised graph learning method, which generates negative samples of the training graph by randomly shuffling edge features to satisfy the graph modeling requirements. The encoder can simultaneously perceive node features, edge attributes, and neighborhood relationships, generating embedding vectors without attack labels. It obtains finer-grained structural features than grayscale images or convolutions at the same parameter level, and can maintain continuous path perception in scenarios with topological changes and discontinuous links. At the student model level, most existing solutions still use or retain the convolutional or attention operations of the teacher model. Lightweight methods are limited to reducing the number of channels or model layers. In contrast, the student model in this invention is a multilayer perceptron, which can achieve rapid inference while reducing memory and latency. At the knowledge distillation level, most existing distillation methods require retaining the complete forward path of the teacher model during the student model training phase, and keeping the entire teacher model in memory to calculate soft label or feature map loss. This invention extracts the embedding vector output by the teacher model as the knowledge carrier, so the student model training phase no longer calls the teacher model, nor does it require temperature scaling or soft label branching. This reduces training memory usage and communication overhead while maintaining stable detection performance. When the terminal is deployed, only a lightweight student model needs to be loaded to achieve real-time inference.

[0041] This invention introduces an edge-level unsupervised graph learning method (Deep Graph Infomax) into the teacher model and combines it with an encoder (E-GraphSAGE) capable of simultaneously perceiving node features, edge attributes, and neighborhood relationships. This enables the model to fully mine the potential correlation information of the graph structure in UAV networks where labels are scarce and data distribution is dynamically changing, significantly improving the structural integrity and semantic distinguishability of node and edge embeddings. Simultaneously, knowledge distillation technology is used to effectively transfer the structural knowledge learned by the highly complex teacher model to the lightweight student model. While ensuring intrusion detection accuracy, this effectively reduces the computational complexity of the inference stage, significantly improving the model's operational efficiency in real-time detection in UAV network environments. Furthermore, due to the small parameter size of the student model, the method of this invention possesses good generalization performance and lightweight deployment characteristics, adapting to the operating conditions of UAV networks where computing power, energy consumption, and payload are limited. It maintains high detection capability against multiple types of attacks, thus overcoming the shortcomings of existing technologies in terms of insufficient detection accuracy, scarce and difficult-to-obtain labels, excessive inference overhead, and deployment difficulties. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart of a multi-classification method for UAV network intrusion detection based on knowledge distillation, as described in an embodiment of the present invention.

[0044] Figure 2 This is a schematic diagram of a directed multi-attribute graph G based on application layer protocol enhancement constructed according to an embodiment of the present invention.

[0045] Figure 3 This is a schematic diagram of the neural network model in an embodiment of the present invention.

[0046] Figure 4 This is the loss function curve during the teacher model training phase in an embodiment of the present invention.

[0047] Figure 5 This is the multi-class classification result of the dataset NF-CSE-CIC-IDS2018-v1 in this embodiment of the invention.

[0048] Figure 6 This is the multi-class classification result of the dataset NF-CSE-CIC-IDS2018-v2 in this embodiment of the invention. Detailed Implementation

[0049] To better understand the above-described objects, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention; however, the invention may be practiced in other ways different from those described herein, and therefore, the invention is not limited to the specific embodiments disclosed below.

[0050] This invention provides a multi-classification method for UAV network intrusion detection based on knowledge distillation. This method innovatively combines knowledge distillation technology with edge-level unsupervised graph learning methods, and transfers the representation of UAV network traffic graph by the teacher model to a lightweight student model. While ensuring detection accuracy, it reduces computational complexity. By using static graph modeling to characterize the spatial dependencies of communication nodes, it improves the generalization of the model and achieves lightweight deployment.

[0051] like Figure 1 As shown, a multi-classification method for UAV network intrusion detection based on knowledge distillation includes the following steps:

[0052] Step S1: Obtain a dataset consisting of multiple sets of drone network traffic data, including normal samples and various attack samples;

[0053] Step S2: Extract communication features from the dataset, generate corresponding edge feature vectors for each traffic data, assign auxiliary labels and sample type labels, then divide the dataset into training and test sets, and construct a directed multi-attribute graph G based on application layer protocol enhancement to perform graph modeling on the traffic data in the training and test sets, transforming the interaction relationships between communication entities into graph structures, and obtaining the training graph (train_g) and the test graph (test_g).

[0054] Step S3: Build a neural network model, which includes a teacher model and a student model. The teacher model is based on the edge-level unsupervised graph learning method DGI, and uses the graph neural network E-GraphSAGE as the encoder for message passing and feature aggregation. The student model is a multilayer perceptron MLP, which consists of N fully connected layers and N-1 sets of optional normalized nonlinear units cascaded alternately, where N≥1.

[0055] Step S4: Train the teacher model using the training graph, and then use the trained teacher model to perform forward inference on the training graph and the test graph respectively to obtain the training embedding vector (Z_train) and the test embedding vector (Z_test).

[0056] Step S5, student model training stage (i.e. knowledge distillation process): The student model takes the training embedding vector output by the teacher model as input, and obtains a well-trained student model through joint optimization of hard label loss function and soft label loss function.

[0057] Step S6: Use the trained student model for inference. Take the test embedding vector generated by the teacher model as input, and the student model outputs the test prediction label. Compare it with the real label in the test graph and calculate the values ​​of each index to evaluate the performance of drone network intrusion detection.

[0058] In this embodiment, the specific method of step S1 includes: obtaining the v1 version dataset NF-CSE-CIC-IDS2018-v1 based on stream format records, generated from the original packet capture files of the dataset CSE-CIC-IDS2018; and obtaining the v2 version dataset NF-CSE-CIC-IDS2018-v2 based on stream format records, generated from the original packet capture files of the dataset CSE-CIC-IDS2018. The dataset source address is: https: / / staff.itee.uq.edu.au / marius / NIDS_datasets / . The dataset CSE-CIC-IDS2018 was jointly released by the Canadian Communications Security Agency (CSE) and the Canadian Cyber ​​Security Institute (CIC) in 2018, simulating attack behaviors in a real enterprise network environment.

[0059] The dataset contains features of drone network traffic data. Specifically, the dataset NF-CSE-CIC-IDS2018-v1 includes: source address (IPV4_SRC_ADDR), source port (L4_SRC_PORT), destination address (IPV4_DST_ADDR), destination port (L4_DST_PORT), address protocol field (PROTOCOL), application layer protocol (L7_PROTO), number of incoming bytes (IN_BYTES), number of outgoing bytes (OUT_BYTES), number of incoming packets (IN_PKTS), number of outgoing packets (OUT_PKTS), cumulative TCP flags (TCP_FLAGS), and flow duration in milliseconds (FLOW_DURATION_MILLISECONDS).

[0060] Key features in the NF-CSE-CIC-IDS2018-v2 dataset include source address (IPV4_SRC_ADDR), source port (L4_SRC_PORT), destination address (IPV4_DST_ADDR), destination port (L4_DST_PORT), address protocol field (PROTOCOL), application layer protocol (L7_PROTO), inbound bytes (IN_BYTES), outbound bytes (OUT_BYTES), inbound packets (IN_PKTS), outbound packets (OUT_PKTS), cumulative TCP flags (TCP_FLAGS), flow duration in milliseconds (FLOW_DURATION_MILLISECONDS), cumulative client TCP flags (CLIENT_TCP_FLAGS), cumulative server TCP flags (SERVER_TCP_FLAGS), ICMP type and code combination value (ICMP_TYPE), and ICMP... Type value (ICMP_IPV4_TYPE), DNS query transaction identifier (DNS_QUERY_ID), DNS query type (DNS_QUERY_TYPE), FTP client command return code (FTP_COMMAND_RET_CODE).

[0061] Specifically, the NF-CSE-CIC-IDS2018-v1 dataset contains a total of 8,392,401 flows, of which 1,019,203 (12.14%) are attack samples and 7,373,198 (87.86%) are benign samples. The NF-CSE-CIC-IDS2018-v2 dataset was generated by increasing the number of network flows on top of the NF-CSE-CIC-IDS2018-v1 dataset; the total number of network flows after the increase is 18,893,708, of which 2,258,141 (11.95%) are attack samples and 16,635,567 (88.05%) are benign samples.

[0062] The two datasets contain the following sample types: Benign traffic, Bot attacks, Brute Force attacks (Web), Cross-Site Scripting (XSS), High-Intensity Request Tool Distributed Denial-of-Service (DDoS attack-HOIC), Low-Orbit Ion Cannon Tool User Datagram Protocol Distributed Denial-of-Service (DDoS attack-LOIC-UDP), Low-Orbit Ion Cannon Tool Hypertext Transfer Protocol Distributed Denial-of-Service (DDoS attacks-LOIC-HTTP), GoldenEye Denial-of-Service (DoS attacks-GoldenEye), Hulk Denial-of-Service (DoS attacks-Hulk), Slow HTTP Test Denial-of-Service (DoS attacks-SlowHTTPTest), Slowloris Denial-of-Service (DoSattacks-Slowloris), File Transfer Protocol Brute Force (FTP-BruteForce), Infilteration, Structured Query Language Injection (SQL Injection), and Secure Shell Protocol Brute Force (SSH-Bruteforce).

[0063] In this embodiment, step S2 specifically includes the following steps:

[0064] Step S21, Data Preprocessing and Feature Encoding: In the process of UAV network intrusion detection, in order to accurately model the interaction relationship between UAV communication entities, firstly, the built-in str(·) function of Python is called to stringify the fields of source address, destination address, source port and destination port in the traffic data of the UAV network to ensure the consistency of field features as attribute graph node identifiers; at the same time, infinite values ​​and missing values ​​in the dataset are replaced and filled, replacing infinite values ​​with 0 and filling missing values ​​with 0 to ensure the stability of subsequent feature processing;

[0065] Step S22: Divide the traffic data samples in the dataset: group the traffic data in the dataset according to the sample type label, and perform random downsampling to reduce computing power requirements to form a training set and a test set; the training set and the test set are divided in a 7:3 ratio so as to construct training graphs and test graphs respectively;

[0066] Step S23, Feature Extraction: Field features in the dataset are converted into numerical columns using target encoding. For the dataset NF-CSE-CIC-IDS2018-v1, target encoding is used to form numerical columns for features such as address protocol, application layer protocol, inbound bytes, outbound bytes, inbound packets, outbound packets, accumulated TCP flags, and stream duration in milliseconds. For the dataset NF-CSE-CIC-IDS2018-v2, the following features are extracted: address protocol, application layer protocol, accumulated TCP flags, accumulated client TCP flags, and server TCP flags. Features such as bit accumulation value, ICMP type and code combination value, ICMP type value, DNS query transaction identifier, DNS query type, and FTP client command return code are encoded into numerical columns using target encoding. The numerical columns are then scaled using a normalization method, and the multidimensional features of each traffic data are merged into an edge feature vector h and stored in the attribute label. In this step, each traffic data record after the above processing will be regarded as an original graph edge with "source address as the starting point and destination address as the ending point" in subsequent steps. The training set and test set after this feature processing constitute the original edge set used in the subsequent graph modeling steps.

[0067] Step S24, Construction of auxiliary edges based on application layer protocol enhancement: Before constructing the attribute graph, the same application layer protocol values ​​in the training set and the test set are divided into multiple protocol domain subsets by the application layer protocol field. Each protocol value corresponds to a protocol domain subset. In each protocol domain subset, all original edges in the domain are traversed, the source address nodes of all original edges are extracted and paired up. If the two source address nodes obtained by pairing are different and there is no edge with the two source address nodes as endpoints in the current protocol domain, auxiliary edges are constructed for the corresponding two source address nodes according to the preset direction determination rules (e.g., based on the ordered relationship of node identifiers), and semantically empty 1 placeholder feature vectors are assigned. All the auxiliary edges generated in this way construct an auxiliary edge set as a structural supplement to the original edge set, which is used to explicitly express the potential communication relationship within the protocol domain and improve the connectivity of the graph structure.

[0068] Step S25, Graph Modeling: Assign auxiliary labels to the traffic data in the dataset. The value of the auxiliary label is 0 or 1. A value of 0 indicates that the traffic data belongs to normal samples (Benign), and a value of 1 indicates that the traffic data belongs to attack samples. During the construction of the attribute graph, the sample type label (Attack) of the attack sample is converted into an integer code by the label encoder (LabelEncoder) and stored as the edge attribute of the attribute graph.

[0069] After merging the original edge set and the auxiliary edge set, the source address and destination address are used as the start and end points of the attribute graph. The edges from the start to the end point of the attribute graph carry the aforementioned edge feature vector h, sample type label, and auxiliary identifier as edge attributes. The attribute graph is modeled using the MultiGraph data structure in the NetworkX graph computing library. The MultiGraph structure supports multiple parallel edges between the same pair of nodes, resulting in an undirected attribute graph. The undirected attribute graph is then converted into a directed attribute graph using the to_directed function to reflect the directionality of communication. Finally, the UAV network traffic data is transformed into a graph object that can be processed by the Deep Graph Learning Library (DGL) in graph neural networks (from_networkx), thus obtaining a directed multi-attribute graph G based on application layer protocol enhancement (e.g., ...). Figure 2 As shown in the figure, the training graph (train_g) and the test graph (test_g) are obtained. Then, the nodes in the directed multi-attribute graph G enhanced by the application layer protocol are assigned a placeholder feature vector with all 1s consistent with the edge feature dimension (that is, the placeholder feature vector is a multi-dimensional array of numerical 1s without semantic information) to meet the input requirements of the encoder in the subsequent neural network model.

[0070] Through this process, the spatial dependencies and traffic interaction patterns of communication links in the UAV network can be accurately modeled, enabling UAVs to effectively perform intrusion detection even in environments with limited computing power and high-interaction communication. It also provides a structured input basis for subsequent representation learning of the teacher model and knowledge distillation of the student model.

[0071] Specifically, in step S3, the teacher model is based on the edge-level unsupervised graph learning method DGI (Deep GraphInfomax), using the graph neural network E-GraphSAGE as the encoder. The student model (i.e., the multilayer perceptron MLP) consists of N fully connected layers (N≥1) and N-1 groups of optional normalized nonlinear units determined by parameters (norm_type), which are alternately cascaded. The optional normalized nonlinear unit structure consists of optional normalization layers (BatchNorm normalization operation). The optional normalized nonlinear units (NLUs) consist of layer normalization (LayerNorm), ReLU activation function, and dropout. When norm_type=batch, all optional normalized NLUs consist of batch normalization, ReLU activation function, and dropout. When norm_type=layer, all optional normalized NLUs consist of layer normalization, ReLU activation function, and dropout. When norm_type=none, all optional normalized NLUs consist of ReLU activation function and dropout. In the student model, the first fully connected layer performs a linear transformation from the input dimension to the hidden dimension, the middle fully connected layers and normalized NLUs perform a constant-dimensional linear transformation from the hidden dimension to the hidden dimension, and the last fully connected layer performs a linear transformation from the hidden dimension to the output dimension. Except for the last layer, after each fully connected layer performs a linear transformation, the optional normalized layer (BatchNorm) in the optional normalized NLUs is executed sequentially. Alternatively, LayerNorm (or Layer normalization operation), ReLU activation function, and Dropout (regularization operation) are used to form a sequence from normalization to nonlinearity and then to regularization. The last layer only retains the linear transformation of the fully connected layer, omitting the optional normalized nonlinear unit (optional normalized layer, ReLU activation function, and regularization operation).

[0072] Specifically, such as Figure 3 As shown, step S4 includes the following steps: During the training phase of the teacher model, the training image is used as input, and the training image is used as a positive sample of the training image (train_g). + A corrupted graph (train_g) is created using a corruption function C to serve as a negative sample for the training graph. - ), and the positive samples of the training image (train_g) + ) and training graph negative samples (train_g - The input is fed into the encoder (Graph Neural Network E-GraphSAGE), and the encoder outputs the positive embeddings (Zi) of the positive samples in the training graph. uv + ) and negative sample embeddings of training graph negative samples (Z uv -), using the Readout function to embed all positive samples (Z uv + The mean is taken, and positive samples are embedded into Z using an activation function (Sigmoid). uv + The mean value is compressed to between 0 and 1, thus generating a global graph summary S containing all the semantics of the graph. A bilinear scoring method is used by the discriminator D to calculate positive and negative sample scores. Specifically, positive sample scores are obtained by calculating the dot product similarity between the positive sample embedding and the global graph summary S, and negative sample scores are obtained by calculating the dot product similarity between the negative sample embedding and the global graph summary S. The obtained positive sample scores, along with labels with a value of 1, and negative sample scores, along with labels with a value of 0, are fed into a binary cross-entropy loss function. This function measures the degree to which positive sample scores tend to 1 and negative sample scores tend to 0, forming the discriminative loss. The discriminative loss is backpropagated to synchronously update the weights of the discriminator D and the parameters of the encoder. The edge-level unsupervised graph learning method DGI maximizes the consistency between the "real edge embedding" and the "global graph embedding," while minimizing the matching degree between the "pseudo-edge embedding" and the graph. After training, a forward inference operation is performed on both the training and test graphs (forward inference refers to feeding the training and test graphs into the already trained teacher model, allowing the model to execute the forward(·) function once, but without updating the parameters, only to obtain the embedding vectors), to obtain the training embedding vector and the test embedding vector respectively. The training embedding vector is transformed by a linear mapping layer to obtain the log odds (T-logits) of the teacher model training, which is used for the training of the student model, and the test embedding vector is used for the inference of the student model.

[0073] in, Figure 4 The loss function curve during the teacher model training phase in this embodiment is shown. As can be seen from the loss function curve, the training loss decreases rapidly within a short iteration period and then gradually stabilizes, demonstrating strong convergence. At the same time, the curve fluctuates very little in the middle and later stages, indicating that the training process maintains good stability and the overall training performance is ideal.

[0074] Step S5 includes the following steps:

[0075] The student model takes the training embedding vectors output by the teacher model as input and passes them to the student model (i.e., a multilayer perceptron, MLP) for training. The student model then outputs its training log-logits (S-Logits). The hard-label loss is calculated using the student model's training log-logits and the ground truth labels on the training graph (train_g.edata['Attack']). Simultaneously, the student model's training log-logits are processed using a softmax activation function to obtain a probability distribution. This probability distribution, along with the teacher model's training log-logits (T-logits) processed using a softmax activation function, is used to calculate the soft-label loss. Specifically, the temperature-scaled KL divergence measures the difference between the student model's output distribution and the teacher model's projected soft-label distribution, constraining the student model. The optimizer (Adam) then performs backpropagation and updates the student model's parameters. Finally, the student model is jointly optimized using the hard-label loss function and the soft-label loss function, resulting in a well-trained student model.

[0076] Specifically, the total loss function (Total_loss) consists of the hard-labeled loss (loss_cls) and the soft-labeled loss (loss_kd). The final definition and calculation of the total loss function (Total_loss) are as follows:

[0077] ,

[0078] In the formula, loss_cls is the hard-label loss of knowledge distillation, which uses the cross-entropy loss function. loss_kd is the soft-label loss of knowledge distillation, which uses the Kullback–Leibler Divergence Loss. α is the weight coefficient of the loss function, and T represents the distillation temperature, used to balance the importance of different loss terms.

[0079] The calculation expression for the cross-entropy loss function is as follows:

[0080] ,

[0081] In the formula, i represents the category of the sample, n is the total number of categories, and Z i Z represents the unnormalized predicted score (logits) of the i-th class sample, indicating the score of the i-th class output by the network, i.e., the raw score before the Softmax activation function; j represents the class of the sample; Z j y represents the unnormalized predicted score (logits) of the j-th class sample; i The hard label indicator variable representing the sample in category i. This represents the logarithm of the probability that the model predicts for class i. Softmax is the activation function used to convert unnormalized predicted scores (logits) into a probability distribution;

[0082] The formula for calculating the KL divergence loss is as follows:

[0083] ,

[0084] In the formula, i represents the category of the sample, n represents the total number of categories, Q is the target distribution of the teacher model, and P is the prediction distribution of the student model; Q(i) represents the prediction probability distribution of the teacher model on category i, and P(i) represents the prediction probability distribution of the student model on category i.

[0085] More specifically, the expression for the predicted probability distribution of the teacher model on category i is:

[0086] ,

[0087] In the formula, i and j represent the categories of the sample, n is the total number of categories, and t i c represents the teacher model's raw score on category i. j Let represent the teacher model's raw score in category j, T represent the distillation temperature, and exp(·) represent an exponential function with base e, where This indicates that the original score t will be... i The unnormalized weights are obtained by scaling by temperature T and then mapping using an exponential function. This indicates that the original score c will be... j Unnormalized weights obtained by scaling with temperature T and then mapping with an exponential function;

[0088] The expression for the predicted probability distribution of the student model on category i is:

[0089] ,

[0090] In the formula, i and j represent the categories of the sample, n is the total number of categories, and s i s represents the student model's raw score on category i. j represents the student model's raw score in category j, and T represents the distillation temperature.

[0091] Step S6 includes the following steps: using the test embedding vector generated by the trained teacher model as input, the trained student model (i.e., multilayer perceptron MLP) is used for inference, and the log odds (S-logits-test) of the student model test is output. The log odds are then converted into a test probability distribution using the Softmax activation function, and the sample type label corresponding to the highest probability is output as the test prediction label. Subsequently, the student model accesses the true label on the test map and compares it with the converted test probability distribution. Three metrics, namely precision, recall, and F1 score, are selected to evaluate the multi-class classification performance of the model.

[0092] Specifically, precision measures the proportion of samples that the model predicts as attacks, but which are actually attacks. High precision means a low number of false positives. The calculation formula is shown below:

[0093] ,

[0094] Recall reflects a model's ability to identify attacks and measures its sensitivity to attack behavior. It reflects the percentage of true instances the model successfully captures. The calculation formula is shown below:

[0095] ,

[0096] The F1 score is the harmonic mean of precision and recall. The best F1 score is 1, and the worst is 0. The F1 score is particularly useful for imbalanced classification problems, helping us understand how well the model balances its performance in handling different classes, especially when the number of positive and negative samples is unequal. It provides a metric for measuring the model's overall ability to distinguish between positive and negative classes. The calculation formula is shown below:

[0097] ,

[0098] In this context, TP represents a true positive, indicating that the model predicts a positive class (positive), and the actual label is also positive. FP represents a false positive, indicating that the model predicts a positive class, but the actual label is negative. FN represents a false negative, indicating that the model predicts a negative class, but the actual label is positive.

[0099] More specifically, the evaluation results of the embodiments of the present invention are as follows: Figure 5 and Figure 6 As shown, Figure 5 The results of the multi-class classification for the dataset NF-CSE-CIC-IDS2018-v1 are as follows: Figure 6This section presents the multi-class classification results for the NF-CSE-CIC-IDS2018-v2 dataset. This drone network intrusion detection test uses precision, recall, and F1 score to reflect the model's detection performance; higher values ​​for these three metrics indicate better detection results. On NF-CSE-CIC-IDS2018-v1, the weighted average of precision, recall, and F1 score exceeded 95%, while on NF-CSE-CIC-IDS2018-v2, the weighted average of precision, recall, and F1 score reached as high as 99%. Figure 5 and Figure 6 As can be seen, except for a few attack categories whose detection performance is generally poor due to sample imbalance in the dataset, the detection performance of botnet attacks, distributed denial-of-service attacks (DDoS attacks-LOIC-HTTP) using the low-orbit ion cannon tool, denial-of-service attacks (DoS attacks-GoldenEye), denial-of-service attacks (DoS attacks-Hulk), denial-of-service attacks (DoS attacks-Slowloris) that slow down web server crashes, and brute-force attacks (SSH) in both datasets is excellent, with low false positive rates and high recall rates.

[0100] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-classification method for UAV network intrusion detection based on knowledge distillation, characterized in that, Includes the following steps: Step S1: Obtain a dataset consisting of multiple sets of drone network traffic data, including normal samples and various attack samples; Step S2: Extract communication features from the dataset, generate corresponding edge feature vectors for each traffic data, assign auxiliary identifiers and sample type labels, then divide the dataset into training set and test set, and construct a directed multi-attribute graph G based on application layer protocol enhancement to perform graph modeling on the traffic data in the training set and test set, transforming the interaction relationship between communication entities into a graph structure to obtain training graph and test graph. Step S3: Build a neural network model, which includes a teacher model and a student model. The teacher model is based on the edge-level unsupervised graph learning method DGI, and uses the graph neural network E-GraphSAGE as the encoder for message passing and feature aggregation. The student model is a multilayer perceptron (MLP) consisting of N fully connected layers and N-1 sets of optional normalized nonlinear units cascaded alternately, where N ≥ 1; Step S4: Train the teacher model using the training graph, and then use the trained teacher model to perform forward inference on the training graph and the test graph respectively to obtain the training embedding vector and the test embedding vector respectively. Step S4 includes the following steps: Step S41: Using the training graph as input, the training graph is used as a positive sample of the training graph. A corrupted graph is created as a negative sample of the training graph through a corruption function. The positive and negative samples of the training graph are input into the encoder. The encoder outputs the positive embedding of the positive sample of the training graph and the negative embedding of the negative sample of the training graph. The Readout function is used to take the mean of all positive sample embeddings. The Sigmoid activation function is used to compress the range of the mean of the positive sample embeddings to between 0 and 1, thereby generating a global graph summary S containing all the semantics of the graph. Step S42: The discriminator D uses a bilinear scoring method to calculate positive and negative sample scores. Specifically, positive sample scores are obtained by calculating the dot product similarity between the positive sample embedding and the global graph summary S, and negative sample scores are obtained by calculating the dot product similarity between the negative sample embedding and the global graph summary S. Then, the obtained positive sample scores and labels with a value of 1, and negative sample scores and labels with a value of 0 are fed into the binary cross-entropy loss function. The binary cross-entropy loss function measures the degree to which positive sample scores tend to 1 and negative sample scores tend to 0, forming the discriminant loss. Step S43: Backpropagate the discriminant loss and update the weights of the discriminator D and the parameters of the encoder simultaneously to complete the training and parameter optimization of the teacher model. Step S44: Perform a forward inference operation on the training graph and the test graph respectively to obtain the training embedding vector and the test embedding vector respectively. The training embedding vector is transformed by the linear mapping layer to obtain the log odds of the teacher model training, which is used for the training of the student model. The test embedding vector is used for the inference of the student model. Step S5, student model training stage: The student model takes the training embedding vector output by the teacher model as input, and obtains a well-trained student model through joint optimization of hard label loss function and soft label loss function. Step S6: Use the trained student model for inference. Input the test embedding vector generated by the teacher model into the trained student model. The student model outputs the test prediction label and compares it with the real label in the test graph. Calculate the values ​​of each index to evaluate the performance of drone network intrusion detection.

2. The method according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: First, the fields of communication features in the traffic data of the UAV network are processed into strings to ensure the consistency of the field features when used as node identifiers in the attribute graph; at the same time, infinite values ​​and missing values ​​in the dataset are replaced and filled, replacing infinite values ​​with 0 and filling missing values ​​with 0 to ensure the stability of subsequent feature processing. The communication characteristics include source address, destination address, source port, and destination port; Step S22: Group the traffic data in the dataset according to the sample type label, and perform random downsampling to reduce computing power requirements to form a training set and a test set; the training set and the test set are divided in a 7:3 ratio; Step S23: The field features in the dataset are converted into numerical columns using target encoding, and the numerical columns are scaled using normalization. The multidimensional features of each traffic data are merged into an edge feature vector h and stored in the attribute label to obtain the original edge set. The field features include address protocol field, application layer protocol, cumulative TCP flag value, cumulative client TCP flag value, cumulative server TCP flag value, ICMP type and code combination value, ICMP type value, DNS query transaction identifier, DNS query type, and FTP client command return code. Step S24: Before constructing the attribute graph, the source address nodes of the original edges with the same protocol value are paired up by the application layer protocol field. According to the preset rules, auxiliary edges are added to the missing potential communication relationship between two source address nodes that do not have original edges, and the auxiliary edge set is obtained as a structural supplement to the original edge set to enhance the structural integrity and semantic relevance of the UAV network traffic attribute graph. Step S25: Assign auxiliary identifiers to the traffic data in the dataset. The value of the auxiliary identifier is 0 or 1. A value of 0 indicates that the traffic data belongs to normal samples, and a value of 1 indicates that the traffic data belongs to attack samples. During the construction of the attribute graph, the sample type label of the attack sample is converted into an integer code by the label encoder and stored as the edge attribute of the attribute graph. After merging the original edge set and the auxiliary edge set, the source address and destination address are used as the start and end points of the attribute graph. Edge feature vectors h, sample type labels, and auxiliary identifiers are carried as edge attributes on the edges from the start to the end point of the attribute graph. The attribute graph is modeled using the multi-graph data structure in the NetworkX graph computing library. The multi-graph data structure supports multiple parallel edges between the same pair of nodes, resulting in an undirected attribute graph. The undirected attribute graph is then converted into a directed attribute graph using the to_directed function to reflect the directionality of communication. Finally, the UAV network traffic data is transformed into a graph object that can be processed by the Deep Graph Learning Library (DGL) in graph neural networks, obtaining a directed multi-attribute graph G based on application layer protocol enhancement, thus obtaining the training and test graphs. Subsequently, nodes in the directed multi-attribute graph G based on application layer protocol enhancement are assigned all-one placeholder feature vectors consistent with the edge feature dimension to meet the input requirements of the encoder in the subsequent neural network model.

3. The method according to claim 2, characterized in that, In step S3, the structure of the optional normalized nonlinear units in the student model is determined by the parameter norm_type. When norm_type=batch, all optional normalized nonlinear units in the student model consist of batch normalization, ReLU activation function, and regularization operation; when norm_type=layer, all optional normalized nonlinear units in the student model consist of layer normalization, ReLU activation function, and regularization operation; when norm_type=none, all optional normalized nonlinear units in the student model consist of ReLU activation function and regularization operation. In the student model, the first fully connected layer performs a linear transformation from the input dimension to the hidden dimension. The middle fully connected layers and normalized nonlinear units perform a constant-dimensional linear transformation from the hidden dimension to the hidden dimension. The last fully connected layer performs a linear transformation from the hidden dimension to the output dimension. Except for the last layer, after each fully connected layer performs a linear transformation, it sequentially performs an optional normalization layer, a ReLU activation function, and a regularization operation in the optional normalized nonlinear unit, forming a normalization-to-nonlinear-to-regularization sequence. The last layer only performs the linear transformation of the fully connected layer. The optional normalization layer is determined by the parameter norm_type, and performs batch normalization, layer normalization, or no normalization operation.

4. The method according to claim 3, characterized in that, Step S5 includes the following steps: The training embedding vectors output by the teacher model are used as input to the student model for training. The log-odds ratio of the student model's training is then output. The hard-label loss is calculated using the log-odds ratio of the student model's training and the ground truth labels on the training graph. Simultaneously, the probability distribution obtained by applying a Softmax activation function to the log-odds ratio of the student model's training is combined with the probability distribution obtained by applying a Softmax activation function to the log-odds ratio of the teacher model's training to calculate the soft-label loss. The soft-label loss measures the difference between the student model's output distribution and the projected soft-label distribution of the teacher model, constraining the student model. The Adam optimizer is then used for backpropagation and parameter updates on the student model. Finally, the student model is jointly optimized using the hard-label loss function and the soft-label loss function to obtain the trained student model.

5. The method according to claim 4, characterized in that, Step S6 includes the following steps: Using the test embedding vector generated by the trained teacher model as input, the pre-trained student model is used for inference, and the log odds of the student model's test are output. The log odds are then converted into a test probability distribution using the Softmax activation function, and the sample type label corresponding to the highest probability is output as the test prediction label. Subsequently, the student model accesses the true label on the test map and compares it with the converted test probability distribution. The multi-class performance of the model is evaluated by precision, recall, and F1 score.

6. The method according to claim 4, characterized in that, When optimizing the student model using both hard-labeled and soft-labeled loss functions, the formula for calculating the overall loss function Total_loss is as follows: ； In the formula, loss_cls is the hard label loss of knowledge distillation, which uses the cross-entropy loss function; loss_kd is the soft label loss of knowledge distillation, which uses the KL divergence loss; α is the weight coefficient of the loss function; and T represents the distillation temperature, which is used to balance the importance of different loss terms. The calculation expression for the cross-entropy loss function is as follows: ； In the formula, i represents the category of the sample, n is the total number of categories, and Z i Z represents the unnormalized predicted score for the i-th class sample, indicating the score of the i-th class output by the network, i.e., the raw score before the Softmax activation function; j represents the class of the sample; Z j y represents the unnormalized predicted score of the j-th class sample; i The hard label indicator variable representing the sample in category i. This represents the logarithm of the probability that the model predicts for class i. Softmax is the activation function used to convert unnormalized predicted scores into a probability distribution; The formula for calculating the KL divergence loss is as follows: ； In the formula, i represents the category of the sample, n represents the total number of categories, Q is the target distribution of the teacher model, and P is the prediction distribution of the student model; Q(i) represents the prediction probability distribution of the teacher model on category i, and P(i) represents the prediction probability distribution of the student model on category i. The expression for the predicted probability distribution of the teacher model on category i is: ； In the formula, i and j represent the categories of the sample, n is the total number of categories, and t i c represents the teacher model's raw score on category i. j represents the teacher model's raw score on category j, and T represents the distillation temperature; Let represent an exponential function with base e, where This indicates that the original score t will be... i The unnormalized weights are obtained by scaling by temperature T and then mapping using an exponential function. This indicates that the original score c will be... j Unnormalized weights obtained by scaling with temperature T and then mapping with an exponential function; The expression for the predicted probability distribution of the student model on category i is: ； In the formula, i and j represent the categories of the sample, n is the total number of categories, and s i s represents the student model's raw score on category i. j represents the student model's raw score in category j, and T represents the distillation temperature.

7. A drone network intrusion detection device, characterized in that, It includes one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to implement the knowledge distillation-based multi-classification method for UAV network intrusion detection as described in any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, The system stores instructions that, when executed by a processor, implement the knowledge distillation-based multi-classification method for UAV network intrusion detection as described in any one of claims 1-6.

Images