A method for detecting attacks on quadrotor UAVs based on neural networks

Abstract

The present invention discloses a method for detecting attacks on a quadrotor UAV based on a neural network. First, the operation data of the quadrotor UAV under normal and covert attack conditions are acquired. Secondly, a feedforward neural network model with a specific number of layers is trained using the normal data, and this model is used to detect covert attacks to obtain the UAV channel information that is not affected by the attacks. Then, for the above UAV channels that are not affected by the attacks, an autoencoder model is trained using the normal UAV data, and the autoencoder is used to detect these channels. Finally, the final detection result is obtained by fusing the first-level feedforward neural network detector and the second-level autoencoder detector. This method realizes the detection of covert attacks on the quadrotor UAV. Compared with the traditional model-based attack detection methods, the present invention can not only improve the recognition ability of covert attacks, but also detect new attack patterns online.

Description

Technical Field

[0001] The present invention belongs to the technical field of aircraft safety, and particularly relates to a method for detecting attacks on a quadrotor UAV based on a neural network. Background Art

[0002] The UAV subtracts the manned function on the basis of a manned aircraft. With its simplified system, no need to be equipped with a life support system, and not restricted by the physiological conditions of the pilot, etc., it is more suitable for performing high-risk tasks with high cost performance. Although the UAV is widely used in the military field due to its higher cost performance compared with manned aircraft, the changing battlefield environment and the emerging attack means of the enemy make it increasingly difficult for the UAV to perform tasks safely and reliably on the battlefield. Existing anomaly detection methods cannot detect small-scale unknown attacks and covert attacks in a timely manner, resulting in the situation that the UAV is extremely likely to get out of control after being attacked, affecting the safe and reliable operation of the UAV and the completion rate of the task. Therefore, the problem of attack detection for UAVs in the case of small-scale unknown attacks and covert attacks urgently needs to be solved. Currently, the widely used attack detection method is the cumulative sum, but there are still the following problems that need to be solved urgently affecting its application: how to obtain the UAV model after being attacked. The traditional method is to try different attack types, which is both time-consuming and extremely dependent on the experience of the detection algorithm designer; how to obtain a better detector. In the past, it was to minimize the detection delay caused by false alarms. Since the maximum detection delay is a worst-case scenario, new detection algorithms can improve in terms of detection delay; how to apply the obtained detector to future unknown situations. In the past, it was to repeat the calculation of the previous detector with future data. Due to the rapidity of the UAV system characteristics, new detection algorithms can reduce the detection time delay. The present invention is generated under this background. Summary of the Invention

[0003] Object of the Invention: To fuse a feedforward neural network and an autoencoder for detection, improve the attack detection method, and enhance the detection ability of the UAV for covert attacks.

[0004] Technical Solution:

[0005] A method for detecting covert attacks on a quadrotor UAV based on a neural network includes the following steps:

[0006] Step 1: Collect the input and output data of the quadrotor UAV under normal flight conditions and under covert attack conditions.

[0007] Step 2: Use the input and output data under normal flight conditions to train a neural network with a specific number of layers to obtain a neural network model that can represent the input and output data of the quadrotor UAV under normal flight conditions.

[0008] Step 3: Use the trained neural network model to detect the covert attacks on the quadrotor UAV, so as to obtain the normal state quantity channels and abnormal state quantity channels of the quadrotor UAV, and thus complete the primary detection.

[0009] Step 4: For the normal state quantity channels under the primary detection, use the input and output data of the quadrotor UAV under normal flight conditions to train an autoencoder model that can represent the input and output data of these normal state quantity channels.

[0010] Step 5: Use the trained autoencoder to detect the normal state quantity channels under the primary detection. Finally, fuse the results of the primary feedforward neural network detector and the secondary autoencoder detector to obtain the final detection result.

[0011] Furthermore, the state space expression of the quadrotor UAV under covert attack in Step 1

[0012] x k = Ax k-1 + Bu k-1

[0013]

[0014] where x k = [x y z φ θ ψ u v w p q r] T . A is the state transition matrix, B is the control input matrix, C is the observation matrix, u k is the control input, x k is the state quantity, y k is the output quantity. a k is the covert attack signal. Collect the input data u k of the quadrotor UAV under normal flight conditions, the output data y k , the input data u k of the quadrotor UAV under covert attack, and the output data

[0015] Furthermore, in Step 2, use the input and output data under normal flight conditions to train a neural network with a specific number of layers to obtain a neural network model that can represent the input and output data of the quadrotor UAV under normal flight conditions. In the training stage, the mathematical expression of the neural network for forward value transmission is as follows:

[0016]

[0017] where is a 5-layer neural network, Relu(·) is the neural network activation function, and i = u k-1, i represents the input data of the neural network, d represents the weight values of each layer of the neural network, and its superscript represents the neural network of the corresponding layer. b represents the bias value of each layer of the neural network, and its subscript represents the neural network of the corresponding layer. The value function of the first-step iterative operation of the above neural network and the output of the neural network are as follows:

[0018]

[0019]

[0020] Among them represents the output data of the first step predicted by the neural network for the quadrotor UAV, and y0 represents the output data of the first step of the quadrotor UAV without being covertly attacked. ||·||2 represents the Euclidean norm. The value function of the next-step iterative operation of the neural network and the output of the neural network are as follows:

[0021]

[0022]

[0023] As the neural network iterates successively, we can obtain:

[0024]

[0025]

[0026]

[0027] Among them, J k is the selected value function, and λ k represents the regularization term, which is used to balance the weight ratio of the value functions at different times.

[0028] The neural network uses the Adam algorithm for backpropagation of the reverse gradient. The value function of the neural network in the specific gradient transfer process can be expressed as:

[0029]

[0030] Since only J k in J' k has d for k steps, so

[0031]

[0032] First, take the derivative of J k with respect to d k , and then update the gradient successively according to the Adam algorithm to achieve gradient descent. This method selects a five-layer neural network, and the gradient update of the fifth-layer neural network is as follows:

[0033]

[0034]

[0035]

[0036]

[0037]

[0038]

[0039] wherein represents the gradient of the fifth-layer neural network at the k-th step, represents the momentum-related term of the gradient of the fifth-layer neural network at the k-th step, represents the mean square error-related term of the gradient of the fifth-layer neural network at the k-th step, respectively represent the correction terms of, ε represents an extremely small random number to avoid division by zero during the gradient update process. The calculation method of is the same as above. The derivatives of the gradient-related terms of the value function with respect to Relu and the derivatives of Relu with respect to their respective gradients are calculated in turn. As the gradient is continuously updated, the value function gradually converges, and at the same time, the neural network model can also simulate the input and output data under normal flight conditions of the quadrotor UAV.

[0040] Furthermore, the trained neural network model is used in step 3 to detect the covert attacks suffered by the quadrotor UAV, so as to obtain the normal state quantity channel and the abnormal state quantity channel of the quadrotor UAV, thereby completing the primary detection. During the detection phase, the mathematical expression of the neural network for forward value transfer is as follows:

[0041]

[0042] wherein is a 5-layer neural network, Relu(·) is the neural network activation function, i = u k-1 , i represents the neural network input data, d represents the weight values of each layer of the neural network, the superscript represents the neural network of the corresponding layer, b represents the bias values of each layer of the neural network, and the subscript represents the neural network of the corresponding layer. The value function and the output of the neural network for the first-step iterative operation are as follows:

[0043]

[0044]

[0045] wherein the output data of the normal quadrotor UAV is replaced with the output data of the quadrotor UAV suffering from covert attacks. Represents the first-step output data predicted by the neural network for the quadrotor UAV, Represents the first-step output data when the quadrotor UAV is under a stealth attack. ||·||2 represents the two-norm. The value function for the next iteration operation of the neural network and the output of the neural network are as follows:

[0046]

[0047]

[0048] As the neural network iterates successively, we can obtain:

[0049]

[0050]

[0051]

[0052] where J k is the selected value function, and λ k represents the regularization term, which is used to balance the weight ratio of the value functions at different times. Using the output data y of the normal quadrotor UAV k The trained neural network model cannot accurately predict the output data of the quadrotor UAV under a stealth attack Therefore, the value function of the neural network will not converge, and in this way, a stealth attack is detected, obtaining the normal state channel and the attacked state channel of the quadrotor UAV under the first-level detection.

[0053] Furthermore, for the normal state quantity channels under the first-level detection in step 4, an autoencoder model that can represent the input-output data of these normal state quantity channels is trained using the input-output data of the quadrotor UAV under normal flight. In the training stage, the mathematical expression of the neural network used for the forward value transfer of the autoencoder is as follows:

[0054]

[0055] where is a two-layer neural network, Relu(·) is the activation function of the autoencoder, υ = [y k , u k-1 , υ represents the input data of the autoencoder, which is composed of the input-output data of the normal state channel under the first-level detection, d represents the weight values of each layer of the feature layer, the superscript represents the feature layer of the corresponding layer, and b represents the bias value of each feature layer, and the subscript represents the corresponding feature layer. The value function for the first iteration operation of the above autoencoder and the output of the autoencoder are as follows:

[0056]

[0057]

[0058] where represents the input and output data of the first step of the quadrotor UAV predicted by the neural network, and u0 and y0 represent the input and output data of the first step of the quadrotor UAV without being covertly attacked. ||·||2 represents the two-norm. The value function for the next iteration of the autoencoder and the output of the autoencoder are as follows:

[0059]

[0060]

[0061] As the autoencoding iterates successively, we can obtain:

[0062]

[0063]

[0064]

[0065] where J k is the selected value function, and λ k represents the regularization term, which is used to balance the weight ratio of the value functions at different times.

[0066] The Adam algorithm is used for the backpropagation of the gradient of the autoencoder. The value function of the autoencoder in the specific gradient transfer process can be expressed as:

[0067]

[0068] Since only J k in J' k has k steps of d, so

[0069]

[0070] First, take the derivative of J k with respect to d k , and then update the gradient successively according to the Adam algorithm to achieve gradient descent.

[0071] Furthermore, step 5 uses the trained autoencoder to detect the normal state quantity channels under the first-level detection. Finally, the results of the first-level feedforward neural network detector and the second-level autoencoder detector are fused to obtain the final detection result. In the detection stage, the mathematical expression of the neural network used for the forward value transfer of the autoencoder is as follows:

[0072]

[0073] where It is a two - layer neural network, and Relu(·) is the activation function of the auto - encoder. υ represents the input data of the auto - encoder, d represents the weight values of each layer's feature layer, the superscript represents the feature layer of the corresponding layer number, b represents the bias value of each feature layer, and the subscript represents the corresponding feature layer. The value function of the first - step iterative operation of the above - mentioned auto - encoder and the output of the auto - encoder are as follows:

[0074]

[0075]

[0076] Among them represents the input - output data of the first step predicted by the neural network for the quad - rotor UAV, respectively represent the input - output data of the first step when the quad - rotor UAV is under a stealth attack. ||·||2 represents the two - norm. The value function of the next - step iterative operation of the auto - encoder and the output of the auto - encoder are as follows:

[0077]

[0078]

[0079] With successive iterations of the auto - encoding, we can obtain:

[0080]

[0081]

[0082]

[0083] Among them, J k is the selected value function, and λ k represents the regularization term, which is used to balance the weight ratio of the value functions at different times. The auto - encoder model trained with the input - output data of a normal quad - rotor UAV cannot accurately predict the input - output data of a quad - rotor UAV under a stealth attack. Therefore, the value function of the auto - encoder will not converge, and thus the stealth attack can be detected, obtaining the normal - state channel and the attacked - state channel of the quad - rotor UAV under secondary detection. The results of the primary and secondary detections are fused to obtain the state quantity of the quad - rotor UAV under a stealth attack.

[0084] Beneficial effects:

[0085] A method for detecting attacks on a quadrotor UAV based on a neural network according to the present invention solves the insensitivity to covert attacks during the flight of the UAV. Compared with general model-based attack detection methods, this method can detect covert attacks more timely and does not require the model information of the quadrotor UAV. General attack detection methods obtain the UAV model after being attacked by trial and error of attack types, which is both time-consuming and extremely dependent on the experience of the detector algorithm designer; for improving the detection ability of the detector, in the past, the detection delay caused by false alarms was minimized. However, the maximum detection delay is a worst-case scenario, so traditional detection algorithms can be improved in terms of detection delay. This method is based on the fitting ability and generalization ability of the neural network for non-linear functions. In the training stage, the input and output data of the quadrotor UAV under normal conditions are used to train a neural network model that can represent the input and output of the quadrotor UAV. In the detection stage, the trained neural network model is used for subsequent detection. Compared with the cumulative sum test and the chi-square test, the detection method that combines the feedforward neural network and the autoencoder has stronger detection ability for covert attacks, which will help reduce the impact of attack signals on the flight performance of the UAV. Description of the Drawings

[0086] Figure 1 It shows a block diagram of the attack detection of a quadrotor UAV based on a neural network;

[0087] Figure 2 It shows the training stage of the feedforward neural network;

[0088] Figure 3 It shows the detection stage of the feedforward neural network;

[0089] Figure 4 It shows the training stage of the autoencoder;

[0090] Figure 5 It shows the detection stage of the autoencoder;

[0091] Figure 6 It shows a block diagram of the secondary detection for covert attacks;

[0092] Figures 7 - 18 It shows the primary detection result of this detection method under covert attacks;

[0093] Figure 19 It shows the secondary detection result of this detection method under covert attacks;

[0094] Figure 20 It shows the prediction error of the secondary detection autoencoder in this detection method under covert attacks;

[0095] Figure 21 It shows the chi-square test result for covert attacks. Detailed Implementation Manner

[0096] The embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention and should not be construed as limiting the present invention.

[0097] A method for detecting attacks on a quadrotor UAV based on a neural network is as follows

[0098] Step 1: Collect the input and output data of the quadrotor UAV under normal flight conditions and under covert attacks. In this method, the mass m of the quadrotor UAV is 1.4 kg, and the moments of inertia of the quadrotor UAV are I x = 0.0211 kg·m 2 I y = 0.0219 kg·m 2 I z = 0.0366 kg·m 2 .

[0099] The state-space expression of the UAV

[0100] x k = Ax k-1 + Bu k-1

[0101] y k = Cx k ,

[0102] where x k = [x y z φ θ ψ u v w p q r] T . A is the state transition matrix, B is the control input matrix, C is the observation matrix, u k is the control input, x k is the state variable, and y k is the output variable. The quadrotor UAV adopts LQR control, and the corresponding cost function of LQR is:

[0103]

[0104] The control input is obtained by minimizing the cost function

[0105] u k = -(R + B T PB) -1 B T PAx k

[0106] where Q and R are the state penalty matrix and control penalty matrix of the LQR controller respectively. P is the solution of the following algebraic Riccati equation: A T PA - P + Q - A TPB(R + B T PB) -1 B T PA = 0。

[0107] In step 2, the input and output data under normal flight conditions are used to train a neural network with a specific number of layers to obtain a neural network model that can represent the input and output data of the quadrotor UAV under normal flight conditions. In the training stage, the mathematical expression of the neural network for forward value transfer is as follows:

[0108]

[0109] where is a 5-layer neural network, Relu(·) is the neural network activation function, i = u k-1 , i represents the neural network input data, d represents the weight values of each layer of the neural network, its superscript represents the neural network of the corresponding layer, b represents the bias value of each layer of the neural network, and its subscript represents the neural network of the corresponding layer. The value function of the first-step iterative operation of the above neural network and the output of the neural network are as follows:

[0110]

[0111]

[0112] where represents the first-step output data of the quadrotor UAV predicted by the neural network, and y0 represents the first-step output data of the quadrotor UAV without being covertly attacked. ||·||2 represents the two-norm. The value function of the next-step iterative operation of the neural network and the output of the neural network are as follows:

[0113]

[0114]

[0115] As the neural network iterates successively, we can obtain:

[0116]

[0117]

[0118]

[0119] where J k is the selected value function, and λ k represents the regularization term, which is used to balance the weight ratio of the value functions at different times.

[0120] The neural network uses the Adam algorithm for backpropagation of gradients. The value function of the neural network in the specific gradient transfer process can be expressed as:

[0121]

[0122] Since J' k only has J k with k steps of d, so

[0123]

[0124] First, take the derivative of J k with respect to d k , and then update the gradient successively according to the Adam algorithm to achieve gradient descent. This method selects a five-layer neural network, and the gradient update of the fifth-layer neural network is as follows:

[0125]

[0126]

[0127]

[0128]

[0129]

[0130]

[0131] where represents the gradient of the fifth-layer neural network at the k-th step, represents the momentum-related term of the gradient of the fifth-layer neural network at the k-th step, where β1 = 0.9, represents the mean square error-related term of the gradient of the fifth-layer neural network at the k-th step, where β2 = 0.99, respectively represent correction terms, where β = 0.25, ε represents an extremely small random number, ε = 10 -5 to avoid division by zero during the gradient update process. The calculation method of

[0132] is the same as above. Successively take the derivative of the gradient-related terms of the value function with respect to Relu, and take the derivative of Relu with respect to their respective gradients. As the gradient is continuously updated, the value function gradually tends to converge, and at the same time, the neural network model can also simulate the input and output data under normal flight conditions of the quadrotor UAV.

[0133]

[0134] Among them is a 5-layer neural network, Relu(·) is the activation function of the neural network, i = u k-1 , i represents the input data of the neural network, d represents the weight value of each layer of the neural network, its superscript represents the neural network of the corresponding layer, b represents the bias value of each layer of the neural network, and its subscript represents the neural network of the corresponding layer. The value function of the first-step iterative operation of the above neural network and the output of the neural network are as follows:

[0135]

[0136]

[0137] Among them, the output data of the normal quadrotor UAV is replaced by the output data of the quadrotor UAV under a stealth attack. represents the first-step output data of the quadrotor UAV predicted by the neural network, represents the first-step output data of the quadrotor UAV when it is under a stealth attack. ||·||2 represents the two-norm. The value function of the next-step iterative operation of the neural network and the output of the neural network are as follows:

[0138]

[0139]

[0140] As the neural network iterates successively, we can obtain:

[0141]

[0142]

[0143]

[0144] Among them, J k is the selected value function, λ k represents the regularization term, which is used to balance the weight ratio of the value functions at different times. Using the output data y of the normal quadrotor UAV k The trained neural network model cannot accurately predict the output data of the quadrotor UAV under a stealth attack Therefore, the value function of the neural network will not converge, and in this way, a stealth attack is detected, and the normal state channel and the attacked state channel of the quadrotor UAV under the first-level detection are obtained.

[0145] In step 4, for the normal state quantity channels under the first-level detection, the input and output data of a quadrotor UAV in normal flight are used to train an autoencoder model that can represent the input and output data of these normal state quantity channels. In the training stage, the mathematical expression of the neural network used for the forward value transfer of the autoencoder is as follows:

[0146]

[0147] where is a two-layer neural network, Relu(·) is the activation function of the autoencoder, υ = [y k , u k-1 , υ represents the input data of the autoencoder, which is composed of the input and output data of the normal state channels under the first-level detection, d represents the weight values of each layer of the feature layer, the superscript represents the feature layer corresponding to the layer number, b represents the bias value of each feature layer, and the subscript represents the corresponding feature layer. The value function of the first-step iterative operation of the above autoencoder and the output of the autoencoder are as follows:

[0148]

[0149]

[0150] where represents the first-step input and output data of the quadrotor UAV predicted by the neural network, u 0-1 , y0 represents the first-step input and output data of the quadrotor UAV when it is not under a stealth attack. ||·||2 represents the two-norm. The value function of the next-step iterative operation of the autoencoder and the output of the autoencoder are as follows:

[0151]

[0152]

[0153] As the successive iterations of the autoencoding are carried out, we can obtain:

[0154]

[0155]

[0156]

[0157] where J k is the selected value function, λ k represents the regularization term, which is used to balance the weight proportion of the value functions at different times.

[0158] The Adam algorithm is used for the reverse gradient propagation of the autoencoder. The value function of the autoencoder in the specific gradient transfer process can be expressed as:

[0159]

[0160] Since J' k only has J in it k with k steps of d, so

[0161]

[0162] First, take the derivative of J k with respect to d k and then update the gradient successively according to the Adam algorithm to achieve gradient descent.

[0163] In step 5, use the trained autoencoder to detect the normal state quantity channels under the first-level detection. Finally, fuse the results of the first-level feedforward neural network detector and the second-level autoencoder detector to obtain the final detection result. In the detection stage, the mathematical expression of the neural network used for the forward value transfer of the autoencoder is as follows:

[0164]

[0165] where is a two-layer neural network, Relu(·) is the activation function of the autoencoder, υ represents the input data of the autoencoder, d represents the weight values of each layer of feature layers, the superscript represents the feature layer of the corresponding layer, b represents the bias value of each feature layer, and the subscript represents the corresponding feature layer. The value function of the first-step iterative operation of the above autoencoder and the output of the autoencoder are as follows:

[0166]

[0167]

[0168] where represents the first-step input and output data predicted by the neural network for the quadrotor UAV, represents the first-step input and output data of the quadrotor UAV when it is under a stealth attack. ||·||2 represents the two-norm. The value function of the next-step iterative operation of the autoencoder and the output of the autoencoder are as follows:

[0169]

[0170]

[0171] As the successive iteration of the autoencoding proceeds, we can obtain:

[0172]

[0173]

[0174]

[0175] where J k is the selected value function, and λ k represents the regularization term, which is used to balance the weight ratio of the value functions at different times. The autoencoder model trained with the input and output data of a normal quadrotor UAV cannot accurately predict the input and output data of a quadrotor UAV under a stealth attack. Therefore, the value function of the autoencoder will not converge, and thus the stealth attack can be detected, obtaining the normal state channel and the attacked state channel of the quadrotor UAV under the secondary detection. The results of the primary and secondary detections are fused to obtain the state quantity under the stealth attack.

[0176] Figure 1 represents the block diagram of the quadrotor UAV attack detection. Figure 2 represents the training stage of the feedforward neural network of the primary detector. When the value function of the feedforward neural network converges, the feedforward neural network can simulate the input and output situation of a normal LQR-controlled quadrotor UAV. Figure 3 represents the detection stage of the feedforward neural network of the primary detector. Using the neural network equivalent model obtained in the training stage, the data of the attacked quadrotor UAV is detected. The non-convergence phenomenon of the value function is obtained by using transfer learning, and then the attacked state channel under the primary detection is obtained. Figure 4 represents the training stage of the secondary detection autoencoder. The normal input and output data of the normal state channel under the primary detection are selected to train the autoencoder. When the value function of the autoencoder converges, the autoencoder can extract the features of the normal input and output data. Figure 5 represents the detection stage of the secondary detection autoencoder. Using the autoencoder obtained in the training stage, the attacked input and output data of the normal state channel under the primary detection are detected. The prediction result and the prediction error are obtained by using transfer learning, and then the attacked state channel under the secondary detection is obtained. Figure 6 represents the designed secondary detection block diagram. Figures 7 - 18 represents the detection results of each state under the primary detection. The value functions of all 12 state channels of the quadrotor UAV, except for the roll angular velocity channel, diverge after the stealth attack is added, and then the primary detection results are obtained. Figure 19 represents the detection results of the secondary detection autoencoder. Figure 20 represents the prediction error of the secondary detection autoencoder. The prediction of the roll angular velocity by the autoencoder shows a large deviation after the stealth attack, and then the detection result is obtained, making up for the deficiency of the detection ability of the primary detector. Figure 21 represents that the chi-square test cannot detect the stealth attack. Compared with this detection method, this detection method has stronger detection ability. Compared with the traditional model-based detection method, this detection method directly obtains the detection result using the trained detector, without the need for UAV model parameters and without estimating the state value using the Kalman filter, effectively reducing the detection delay.

[0177] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A method for detecting attacks on a quadrotor UAV based on a neural network, characterized in that: The steps are as follows: 1) Collect the input and output data of the quadrotor UAV under normal flight conditions, and collect the input and output data of the quadrotor UAV under hidden attack conditions; 2) Use the input and output data of the quadrotor UAV under normal flight conditions to train a neural network model; 3) Conduct primary detection: Detect the hidden attack on the quadrotor UAV according to the neural network model to obtain the normal state channel and the hidden attack state channel of the quadrotor UAV; 4) For the normal state quantity channel of the quadrotor UAV under primary detection, use the input and output data of the quadrotor UAV under normal flight conditions to train an autoencoder model; 5) Conduct secondary detection: Detect the normal state quantity channel under primary detection according to the autoencoder model, and obtain the normal state quantity channel and the abnormal state quantity channel of the quadrotor UAV under secondary detection according to the characteristics of the input and output data of the previous step and the next step learned by the autoencoder model; Finally, fuse the neural network model in step 3) and the detection results of the secondary autoencoder model in step 5) to obtain the final detection result; Collect the input data u of the quadrotor UAV under normal flight conditions in step 1) k and the output data y k as well as the input data of the quadrotor UAV under stealth attack and the output data Where: x k = Ax k-1 + Bu k-1 Among them, the state quantity x at step k k = [x y z φ θ ψ u v w p q r] T , a k is the stealth attack signal, x is the position of the quadrotor UAV in the x-axis direction, y is the position of the quadrotor UAV in the y-axis direction, z is the position of the quadrotor UAV in the z-axis direction, θ is the pitch angle of the quadrotor UAV, φ is the roll angle of the quadrotor UAV, ψ is the yaw angle of the quadrotor UAV, u is the speed of the quadrotor UAV in the x-axis direction, v is the speed of the quadrotor UAV in the y-axis direction, w is the speed of the quadrotor UAV in the z-axis direction, p is the roll angular velocity, q is the pitch angular velocity, r is the yaw angular velocity, A is the state transition matrix, B is the control input matrix, C is the observation matrix, and x k-1 with subscript k - 1 is the state quantity at step k - 1.

2. The method for detecting attacks on a quadrotor UAV based on a neural network according to claim 1, wherein, The mathematical expression of the neural network model in step 2) is as follows: Among them is a 5-layer neural network model, Relu(·) is the activation function of the neural network model, i represents the input data of the neural network model, and i = u k-1 , d represents the weight values of each layer of the neural network, its superscript represents the neural network of the corresponding layer, b represents the bias value of each layer of the neural network, and its subscript represents the neural network of the corresponding layer Obtained with successive iterations of the neural network: Among them represents the normal output data predicted by the feedforward neural network in the neural network model during the training process, y k represents the actual output data of the quadrotor UAV, J k is the selected value function, λ k represents the regularization term; Update the weights of the neural network using the Adam algorithm, and the neural network value function J' k is expressed as: Since the neural network value function J' k only has the selected value function J k with d for k steps, so: Perform gradient update on the neural network model. The gradient update of the fifth-layer neural network is as follows: Among them represents the gradient of the fifth-layer neural network at the k-th step, represents the momentum-related term of the gradient of the fifth-layer neural network at the k-th step, represents the mean square error-related term of the gradient of the fifth-layer neural network at the k-th step, respectively represent the correction terms of, ε represents an extremely small random number to avoid division by zero during the gradient update process, β1 represents the learning rate of the momentum gradient-related term, β2 represents the learning rate of the mean square error gradient-related term, and β represents the Adam gradient learning rate, The calculation method of is the same as above. As the gradient is continuously updated, the value function J' k gradually tends to converge, and at the same time, the neural network model simulates the input and output data under the normal flight conditions of the quadrotor UAV.

3. The method for detecting attacks on a quadrotor UAV based on a neural network according to claim 2, characterized in that , in step 3), according to the neural network model: Replace the output data y of a normal quadrotor UAV k with the output data of the quadrotor UAV under a stealth attack As the neural network iterates successively, we can obtain: Using the output data y of a normal quadrotor UAV k The trained neural network model cannot accurately predict the output data of a quadrotor UAV under a stealth attack Therefore, the value function of the neural network will not converge, thereby detecting the stealth attack received, and obtaining the normal state channel and the stealth attack state channel of the quadrotor UAV under the first-level detection.

4. The method for detecting attacks on a quadrotor UAV based on a neural network according to claim 3, wherein , in step 4), the mathematical expression of the autoencoder model is as follows: Among them is a two-layer neural network, Relu(·) is the activation function of the autoencoder, υ = [y k , u k-1 , υ represents the input data of the autoencoder. The autoencoder model iterates successively, and updates the weights of the autoencoder through the Adam algorithm, and then updates the gradient successively according to the Adam algorithm to achieve the descent of the gradient.

5. The method for detecting attacks on a quadrotor UAV based on a neural network according to claim 4, characterized in that, , in step 5), according to the autoencoder model: Replace the output data y of a normal quadrotor UAV k with the output data of a quadrotor UAV under a stealth attack As the autoencoder model iterates successively, the autoencoder model trained with the input and output data of a normal quadrotor UAV cannot accurately predict the input and output data of a quadrotor UAV under a stealth attack. Therefore, the value function of the autoencoder will not converge, detecting the stealth attack in this way and obtaining the normal state channel and the stealth attack state channel of the quadrotor UAV under the secondary detection.

Images