CN-122027371-A - False data injection attack identification method based on self-adaptive residual weighting PINN

CN122027371ACN 122027371 ACN122027371 ACN 122027371ACN-122027371-A

Abstract

The invention provides a false data injection attack identification method based on self-adaptive residual weighting PINN, which comprises the steps of constructing a physical equation describing the running state of a system, constructing a self-adaptive residual weighting network, taking sensor measurement data with noise or attack as input and the reconstructed system state as output, training the self-adaptive residual weighting network by using a gradient descent algorithm, mapping the sensor data into the system state meeting the physical consistency in real time by using the self-adaptive residual weighting network, and identifying false data injection attack by using a double criterion mechanism according to the weight coefficient and the physical residual information calculated by the self-adaptive residual weighting network in real time.

Inventors

YU MENG
WEI ZHIHUI

Assignees

南京理工大学

Dates

Publication Date: 20260512
Application Date: 20260415

Claims (8)

1. The false data injection attack identification method based on the adaptive residual weighting PINN is characterized by comprising the following steps of: step 1, determining connection relations and line parameters among nodes according to a network topology file of an industrial control system, and constructing an observation equation describing the running state of the system; Step 2, constructing an adaptive residual weighting network, wherein the adaptive residual weighting network takes sensor measurement data with noise or attack as input and takes a reconstructed system state as output; training the self-adaptive residual weighting network by using a gradient descent algorithm, and completing training when the change amount of the total loss function continuously set for the repeated iteration is smaller than a set convergence threshold or reaches a preset maximum training round number; step 4, mapping the sensor data into a system state meeting physical consistency in real time by utilizing a self-adaptive residual weighting network; and 5, identifying false data injection attack through a double criterion mechanism according to the weight coefficient and physical residual information calculated in real time by the self-adaptive residual weighting network.
2. The false data injection attack identification method based on the adaptive residual weighting PINN according to claim 1, wherein the constructed observation equation describing the operation state of the system is specifically: In the formula, For the theoretical measurement of the vector quantity, Representing a nonlinear mapping function commonly determined by the physical mechanism of the target industrial control system and the spatial topological connection relation of the equipment, Is a state variable vector.
3. The method for identifying false data injection attacks based on adaptive residual weighting PINN according to claim 2, wherein when the industrial control system is an electrical industrial control system, the specific method for constructing an observation equation describing the operation state of the system is as follows: with voltage amplitude at nodes of the electrical power industry control system And phase angle As state variable vector to be estimated Active power collected by a sensor is selected Reactive power As a measurement vector Based on kirchhoff's law and line parameters, an observation equation is constructed The method specifically comprises the following steps: In the formula, Representing nodes Is used for controlling the active power injection amount of the (a), Representing nodes Is used for controlling the reactive power injection amount of the (1), Representing nodes Is used for the voltage amplitude of (a), Representation and node Connected nodes Is used for the voltage amplitude of (a), Representation and node There is a set of all neighbor nodes that are directly electrically connected, Representing connection nodes Sum node Is provided with a circuit-line conductivity of (1), Representing connection nodes Sum node Is used for the line susceptance of (a), Representing nodes And node Voltage phase angle difference between them.
4. The method for identifying false data injection attacks based on adaptive residual weighting PINN of claim 2, wherein when the industrial control system is a fluid pipe network system, the specific method for constructing an observation equation describing the operation state of the system is as follows: fluid pressure of each intersection point in pipe network as state variable vector to be estimated Real-time pressure and flow data acquired by pressure sensors and flow meters deployed on pipe network nodes or pipelines are selected as measurement vectors For any pipe network node At fluid pressure Injecting traffic with nodes for state variable vectors Nonlinear physical equation constructed based on mass conservation law and pipeline resistance characteristics for measuring quantity Expressed as: In the formula, Representation and node A set of directly connected neighboring nodes; 、 Respectively nodes Sum node Is used for estimating the pressure state quantity; For connecting nodes And node Is a physical impedance characteristic coefficient of the pipeline; As a sign function.
5. The false data injection attack identification method based on self-adaptive residual weighting PINN according to claim 1, wherein the self-adaptive residual weighting network includes a fully-connected physical information neural network, a physical constraint module and a self-adaptive weight module, the physical information neural network includes an input layer, a hidden layer and an output layer, actually-collected sensor measurement data is input into the input layer, after linear transformation and nonlinear activation of the hidden layer, a preliminary estimated state vector of the system at the current moment is output at the output layer, the physical constraint module inputs the preliminary estimated state vector output by the physical information neural network into an observation equation describing the running state of the system, obtains a theoretical measurement value, calculates the theoretical measurement value and actually-collected sensor measurement data to obtain a physical residual, and the self-adaptive weight module is used for calculating dynamic weights according to the physical residual.
6. The method for identifying false data injection attacks based on adaptive residual weighting PINN in claim 5, wherein the specific formula for calculating the dynamic weight coefficient by the adaptive weight module according to the physical residual is: In the formula, Represent the first The weighting coefficients of the individual sensor data in the loss function, Represent the first Physical residual values corresponding to the individual sensor data.
7. The method for identifying false data injection attacks based on adaptive residual weighting PINN of claim 1, wherein the total loss function during training of the adaptive residual weighting network is: In the formula, Representing the total loss value at the time of neural network training, Representing the total number of sensors in the system, Represent the first Adaptive weights of individual sensor data, Represent the first The actual measured value of the individual sensor(s), Representing estimating state vectors from a network And the first deduced from the observation equation The theoretical measurements of the individual positions are used, Representing the coefficients of the regularized term, A set of vectors representing all of the sensor weights, Representing the L2 norm.
8. The false data injection attack identification method based on the adaptive residual weighting PINN according to claim 1, wherein the specific formula for identifying the false data injection attack by the dual criterion mechanism according to the weight coefficient and the physical residual information calculated in real time by the adaptive residual weighting network is: In the formula, Represent the first The final attack decision of the number sensor, 1 Represents that the sensor is judged to suffer from false data injection attack; A value of 0 represents that the sensor data is normal or contains only random noise within an allowable range, Represents a theoretical measurement calculated from the system state of the trained adaptive residual weighting network output, Is the first The actual measured value of the individual sensor(s), For the first of the network outputs Adaptive weighting values for individual sensor measurements, Represents a residual decision threshold value, The weight cutoff threshold is indicated.

Description

False data injection attack identification method based on self-adaptive residual weighting PINN Technical Field The invention belongs to the safety technology of industrial control systems, and particularly relates to a false data injection attack identification method based on self-adaptive residual weighting PINN. Background Industrial Control Systems (ICS) are the core "neural hubs" of key infrastructure for electricity, petrochemicals, rail transit, water service, etc. With the deep integration of industrial internet and internet of things, a closed industrial control environment gradually goes to be open, which also makes the network security threat faced by an industrial control system increasingly severe. Among many attack means, false data injection attack is a first threat in the current industrial security field due to strong concealment and high destructive power. An attacker often misdirects the decision making system of the control center through tampering with real-time data (such as voltage, frequency, flow and the like) acquired by the sensor, and the disastrous consequences such as physical damage of equipment, large-area power failure and the like can be caused. The existing defense means mainly depend on the traditional state estimation method (such as a weighted least square method) based on residual detection, but the traditional method is often difficult to identify when facing the carefully constructed intelligent attack vector which accords with the physical topological constraint. Although deep learning technology is introduced in recent years to improve the detection capability, the pure data-driven AI model is highly trained by a large amount of attack samples, and the industrial field often lacks real attack data, and the model lacks interpretability and is difficult to land in actual industrial production. The Physical Information Neural Network (PINN) is used as an emerging method for combining physical mechanism and deep learning, and provides a new idea for solving the problems. However, in practical application, it is found that when the standard PINN model faces large-scale false data injection, the loss function of the model is easy to be "tired" of polluted data, so that the model sacrifices the constraint force of a physical rule for forcibly fitting error data, and normal measurement noise cannot be accurately distinguished from malicious attack data. Disclosure of Invention The invention provides a false data injection attack identification method based on self-adaptive residual weighting PINN. The technical scheme for realizing the purpose of the invention is that the false data injection attack identification method based on the self-adaptive residual weighting PINN comprises the following steps: step 1, determining connection relations and line parameters among nodes according to a network topology file of an industrial control system, and constructing an observation equation describing the running state of the system; Step 2, constructing an adaptive residual weighting network, wherein the adaptive residual weighting network takes sensor measurement data with noise or attack as input and takes a reconstructed system state as output; training the self-adaptive residual weighting network by using a gradient descent algorithm, and completing training when the change amount of the total loss function continuously set for the repeated iteration is smaller than a set convergence threshold or reaches a preset maximum training round number; step 4, mapping the sensor data into a system state meeting physical consistency in real time by utilizing a self-adaptive residual weighting network; and 5, identifying false data injection attack through a double criterion mechanism according to the weight coefficient and physical residual information calculated in real time by the self-adaptive residual weighting network. Compared with the prior art, the method has the remarkable advantages that firstly, the method can dynamically adjust the weight of each measurement data according to the physical residual error size by introducing the self-adaptive residual error weighting mechanism into the loss function of the physical information neural network, greatly enhance the robustness of the system, and can still maintain the accuracy of state estimation and avoid being 'biased' by attack data even if a plurality of sensors are simultaneously subjected to large-scale malicious tampering. Secondly, the invention constructs the physical topological structure and the operation rule of the industrial control system into a mathematical mechanism model, and uses the mathematical mechanism model as a constraint reference of the neural network, thereby reducing the data threshold, not relying on massive historical attack samples, and realizing the defense of novel unknown attacks only by normal operation data and the physical rule. The invention provides good interpretability, and an operation