SAR adversarial image detection system and method based on multi-objective neural network architecture optimization

Abstract

The application discloses a SAR (Synthetic Aperture Radar) image detection system and method based on multi-target neural network architecture optimization. Firstly, image data sets from a SAR historical database are acquired, and data analysis and normalization are performed on the image data sets, which are taken as input data sets of a SAR image detection offline training module. A neural network architecture optimization platform based on a non-dominated sorting genetic algorithm is used to obtain a lightweight and highly robust neural network optimal architecture model for a FUSAR-ship data set. The model is convenient to deploy and can be used for online identification of images generated by real-time SAR monitoring of the ground. The application can be used for automatic generation and optimization design of a SAR image detection neural network architecture, improves the anti-robustness and accuracy of a complex SAR image detection system, reduces the complexity of a SAR identification model, and greatly reduces the required resources in deployment of the identification model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent remote sensing, in particular to a SAR (Synthetic Aperture Radar) adversarial image detection system and method based on multi-objective neural network architecture optimization. BACKGROUND

[0002] SAR is different from other infrared detection, visible light imaging and other optical detection means, and is an active microwave imaging sensor that is not affected by factors such as clouds, fog and light. It can observe target areas at any time, which makes it widely used in civilian and military fields, and SAR image detection has become increasingly important. However, the increasingly widespread application of SAR and its importance in various fields make it an easy target for adversaries, making it vulnerable to various attacks. Therefore, there is an urgent need for the development of a lightweight SAR image detection system with high robustness and high accuracy in the civilian and military fields.

[0003] The generation of SAR images is easily affected by the configuration conditions of the observed target. When the morphology, obstructions and imaging parameters of the same observed target have large differences, the SAR image will also change greatly, which increases the difficulty of SAR image detection. With the continuous superiority of deep learning in computer vision, many scholars have applied it to SAR image detection and designed various deep neural network models with high detection performance. However, when faced with various adversarial attacks (such as FGSM, PGD, CW, GN, DeepFool, SSP, etc.), the detection performance of classic deep neural network models such as VGG16, Densenet201 and ResNet101 will decrease sharply, indicating that the existing deep neural network models have poor adversarial robustness and pose a serious security risk. At the same time, the deep neural network models designed in the current SAR image detection field often have complex network structures, and automatic optimization design of deep neural network models has not been realized. A large amount of professional knowledge is required to construct them, which is time-consuming and labor-intensive, and difficult to meet the multi-performance requirements of SAR detection systems in real-time, accuracy, adversarial robustness and lightweight. Therefore, automatic multi-objective optimization design of SAR adversarial image detection neural network models is a technical problem that needs to be solved in the field of SAR image detection. SUMMARY

[0004] The present application aims to overcome the shortcomings of the prior art and provide a lightweight SAR adversarial image detection system and method with high robustness based on multi-objective neural network architecture optimization.

[0005] The application aims to realize the above technical scheme, and the application discloses a lightweight high-robustness SAR adversarial image detection system based on multi-objective neural network architecture optimization.

[0006] The data acquisition module obtains a to-be-detected target center of a high-resolution image of a SAR historical database, takes the center of a minimum circumscribed eccentric circle of the target center as a midpoint, expands 256 pixel points outward, obtains a SAR image slice dataset with a size of 512 pixels*512 pixels, performs normalization processing on the image slice dataset and classifies the image slice dataset, divides the image slice dataset into a training set and a verification set, obtains an adversarial verification set after performing adversarial attack on the verification set, and takes the training set and the adversarial verification set as inputs of an offline optimization training module;

[0007] The offline optimization training module is based on a multi-objective neural network architecture optimization solver, obtains optimal neural network model architecture information of lightweight high robustness, and transmits the optimal neural network model architecture information to an online detection module;

[0008] The online recognition module performs online deployment on the lightweight neural network optimal model, performs online detection on a real-time acquired SAR image, and timely outputs class information of SAR image classification and performance indicators of evaluating image classification.

[0009] A lightweight high-robustness SAR adversarial image detection method based on multi-objective neural network architecture optimization and applied to the system comprises the following steps.

[0010] (1) Data set acquisition and preprocessing. The data acquisition module of the SAR image detection obtains a source data set X by performing to-be-detected target center acquisition, image slicing and class labeling on a high-resolution image generated by a Gaofen-3 satellite, scales each image to 256 pixels*256 pixels, and performs data normalization processing according to formula (1) to obtain an offline training data set X o , inputs an offline optimization training module based on multi-objective neural network architecture optimization, divides X o according to a 4:1 ratio to obtain a training set X o-train and a verification set X o-valid .

[0011]

[0012] X oj represents the jth dimension feature of the offline training data set X o , and X jdenotes the j-th dimension of the source data set X, d denotes the maximum value of the dimension of X o ;

[0013] (2) Set the parameter values of the offline optimization training module based on multi-objective neural network architecture optimization, including population size N, the number of cells C constituting the neural network architecture, the number of chromosomes l of each cell, the number of iteration optimization I of architecture optimization, the training round Epoch of neural network offline training, the size of the adversarial disturbance Epsilon, the custom parameters η of the crossover operation, the mutation rate σ and the upper limit σ of the mutation interval max , the lower limit σ of the mutation interval min ;

[0014] (3) Initialize N individuals as the initial population Q. The coding form of each individual is x i ={Normal Cell, Reduction Cell, Optim, lr}, where Normal Cell and Reduction Cell represent the architecture coding of the offline training neural network, Optim represents the optimizer type of the offline training of the neural network, and lr represents the learning rate of the offline training optimizer of the neural network.

[0015] Normal Cell is represented as A NormalCell contains l chromosomes, and the i-th array represents the i-th chromosome, where 1≤i≤l, operation represents a type of operation, node represents a node to be operated, represents the 1st node of the i-th chromosome, represents the 2nd node of the i-th chromosome, represents the operation on , represents the operation on .

[0016] Wherein operation is randomly selected from the operation set (labeled as OPERATIONS), the composition of OPERATIONS is: [Maxpooling, Avgpooling, Identity, SepConv_3x3, SepConv_5x5, SepConv_7x7, DilConv_3x3, DilConv_5x5, Conv_1x7_7x1], which is represented by [0, 1, 2, 3, 4, 5, 6, 7, 8] respectively in the code. Among them, Maxpooling represents the maximum pooling operation, Avgpooling represents the average pooling operation, Identity represents no operation, SepConv_3x3, SepConv_5x5, SepConv_7x7 represent depth separable convolution units with convolution kernel size of 3x3, 5x5, 7x7 respectively, DilConv_3x3, DilConv_5x5 represent the dilated convolution unit with convolution kernel size of 3x3, 5x5 respectively, Conv_1x7_7x1 represents a spatial separable convolution unit composed of two convolution layers with convolution kernel size of 1x7 and 7x1 respectively by dividing a convolution kernel layer with convolution kernel size of 7x7.

[0017] The first chromosome of a Normal Cell contains two initial nodes, respectively labeled as and represents the 1st node of the first chromosome in the Normal Cell, represents the 2nd node of the first chromosome in the Normal Cell, which is respectively from the output of the previous two cells. By randomly selecting the existing node, performing operation on it, a new node is generated. A chromosome contains two operations and two nodes. By performing operation operation on the node, two new nodes are generated until the number of nodes in the chromosome is equal to l. The nodes that have not been operated (except the initial nodes) are combined as the output of the entire Normal cell.

[0018] The Reduction Cell is composed of A Reduction Cell contains l chromosomes, the i-th array represents the i-th chromosome, wherein 1≤i≤l, execution represents a type of operation, knot represents a node to be operated, represents the 1st node of the i-th chromosome, represents the 2nd node of the i-th chromosome, execution represents the operation of represents the operation of execution is randomly selected from OPERATIONS, the first chromosome of a Reduction Cell contains two initial nodes, marked as and represents the 1st node of the first chromosome of the Reduction Cell, represents the 2nd node of the first chromosome of the Reduction Cell, which comes from the output of the previous two cells. By randomly selecting an existing knot, performing execution on it, a new knot is generated. A chromosome contains two executions and two knots. By performing execution on the knot, two new knots are generated. Until the number of knots in the chromosome is equal to l. The knots that have not been executed (except for the initial nodes) are combined as the output of the entire Reduction Cell. The difference between the Reduction Cell and the Normal Cell is that the convolution step size Stride of the convolution-related operation in each Reduction cell needs to be changed to 2.

[0019] Optimizer is randomly selected from the set Optimizer = [Adadelta, Adam, SGD, RMSprop, Adagrad, AdamW, Adamax, ASGD, Rprop], and the corresponding code of Optimizer is [0, 1, 2, 3, 4, 5, 6, 7, 8]. Among them, AdaDelta represents an improved adaptive learning rate optimizer, Adam represents an adaptive moment estimation optimizer, SGD represents a stochastic gradient descent optimizer, Adagrad represents an adaptive learning rate optimizer, Rmsprop represents an RMSprop optimizer, AdamW represents an adaptive moment estimation optimizer with weight decay, Adamax represents an adaptive moment estimation optimizer based on the infinity norm, ASGD represents an asynchronous stochastic gradient descent optimizer, and Rprop represents an elastic backpropagation optimizer. The learning rate lr of Optim is randomly selected from the set LearningRate = [0.1, 0.01, 0.001, 0.0001], and the corresponding code of LearningRate is [0, 1, 2, 3].

[0020] An x i ​The instantiation process involves stacking Normal Cells and Reduction Cells in an ordered manner until the number of Cells equals C. A compressed excitation network module (labeled as the SE layer) is added after each cell. After the cells are stacked, a Global pooling layer and a Non-Local layer are added at the end of the network. The Global pooling layer represents the global pooling module, and the Non-Local layer represents the non-local network module.

[0021] (4) Evaluate the performance of the initial population Q. Based on the architectural information represented by each individual in Q, convert it into the corresponding neural network model, and use the offline training dataset X. o-train As the data input for each convolutional neural network model, it undergoes offline training for Epoch rounds, and is used on the validation set X. o-valid Randomly select one of the four adversarial attacks to launch an attack, generating an adversarial verification set X. o-adv-valid Verification tests were conducted, and the results were calculated on the verification set X according to formulas (2) and (3) respectively. o-valid and X o-adv-valid The target values ​​are f1 and f2, and the four adversarial attacks are: Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), Gaussian Noise (GN), and Carlini & Wagner attack (CW). Formulas (2) and (3) are as follows:

[0022]

[0023] Where N err For X o-valid The number of incorrectly predicted samples, N all For X o-valid The total number of samples.

[0024]

[0025] Where N adv-err For X o-adv-valid The number of adversarial examples predicted incorrectly, where Ave and Std represent the number of examples predicted incorrectly in X. o-adv-valid The average and standard deviation of the test accuracy obtained from multiple test models under adversarial attacks.

[0026] (5) Calculate the f1 and f2 values ​​for each individual in population Q. If x i Both f1 and f2 values ​​are better than x. j If the f1 and f2 values ​​are given, then x is called x. i Disposable xj x j The number of dominated individuals n j While adding 1, x j Joined by x i Dominant set of individuals s i In the middle, compare the dominance among individuals to obtain the number of dominated individuals n and the set of dominated individuals s for each individual. Find the individual in the population where n is 0 and store it in set F1; iterate through the set of individuals s dominated by each individual in F1, and perform n operations on each individual g in s. g =n g -1, if there exists n g Individuals with a value of 0 are added to set F2; the above operation is repeated on F2 until the entire population Q is divided into sets F of different levels.

[0027] (6) For a population Q that has already been ranked and classified by dominance, for each set of ranks F k Where 1≤k≤N, sort in ascending order according to f1 and f2 respectively, and then sort F... k Each individual m in the process is subjected to crowding degree calculation according to formula (4).

[0028]

[0029] Where, m d This represents the crowding degree of individual m. f represents the target value of individual m+1. j , f represents the target value of individual m-1. j .

[0030] (7) Perform a traversal operation based on the order of individuals in population Q, selecting two individuals as parent individuals each time, and marking them as follows: and An individual contains two types of cell codes: Normal Cell code and Reduction Cell code. A cell contains l chromosomes. In Normal Cell code, 4 bits are used to represent two nodes and two operations on a chromosome; in Reduction Cell code, 4 bits are used to represent two knots and two executions on a chromosome. n is the number of gene codes in a cell, with a value of 8l. Crossover calculation between two parent individuals is performed according to formula (5), resulting in two offspring individuals, namely... and

[0031]

[0032] where β is dynamically and randomly determined by distribution factor η according to formula (6);

[0033]

[0034] where rand represents a random number generated in the range of 0 to 1, and η is a self-defined parameter, and the greater the value of η, the closer the generated offspring individual to the parent individual. The generated offspring c 1 and c 2 are stored in a new population set P, and the above operation is repeated until the size of set P is equal to the size of population Q.

[0035] (8) Traverse the individuals in the new population P and perform mutation operation on them. For an individual a 0-1 floating point number p is randomly generated, and the relationship between p and mutation rate σ is judged, if p > σ, the individual does not need to be mutated; otherwise, the individual is mutated according to formula (7).

[0036]

[0037] wherein, is the kth encoding of individual x i in population P, is the kth encoding of offspring individual x im obtained according to , l is the number of individual chromosomes, and u k is the kth number in σ max , and l k is the kth number in σ min , wherein δ is calculated according to formula (8).

[0038]

[0039] wherein λ m is a distribution index, rand is a random number between 0 and 1, and δ1 and δ2 are calculated according to formula (9) and formula (10) respectively.

[0040] δ1 = (v k -l k ) / (u k -l k ) (9)

[0041] δ2 = (u k -l k ) / (v k -l k ) (10)

[0042] (9) According to the performance evaluation method described in step (4), the f1 and f2 target values of the population P after crossover and mutation are calculated. The population P is combined with the population Q to obtain the population R, and the population R is non-dominantly sorted according to step (5), and the crowded distance of the population R after non-dominant sorting is calculated according to step (6). In each set F, the first N individuals of the sorted population R are selected according to the size of the crowded degree from small to large, to form a new offspring population Q n ;

[0043] (10) Unconditionally accept Q = Q n ;

[0044] (11) Repeat steps (4) to (9) until the set number of iteration optimization I of architecture optimization is met;

[0045] (12) The individual with the smallest crowded degree in the final F1 set obtained in the offline optimization training module is taken as the optimal individual X best , and the optimal neural network model in the SAR image detection module is constructed according to the optimal architecture information represented by X best , and the parameters of the optimal neural network model are set, including the number of rounds of online detection Epoch2, the learning rate, the type of optimizer and the momentum setting, etc.;

[0046] (13) The data acquisition module of the SAR image detection method obtains SAR image information from SAR real-time monitoring, and after the center of the target to be detected, image slicing and data normalization processing, it is taken as the online detection data set of the SAR image detection module based on the optimal neural network model, marked as X t , FGSM, PGD, CW, GN, and self-supervised perturbation attack (marked as DeepFool) based on orthogonal projection to the decision boundary are performed on X t , respectively, to obtain the corresponding adversarial data set X adv ;

[0047] (14) The optimal neural network model is run by the online detection module to perform online SAR image detection on the online detection data set X t , and the test accuracy of the SAR image is calculated according to formula (11); the online SAR image detection is performed on the online adversarial detection data set X adv , and the adversarial accuracy of the SAR image is calculated according to formula (12).

[0048]

[0049]

[0050] wherein, N true represents the number of correctly predicted normal SAR image samples, N all represents the total number of SAR image samples, N adv-true represents the number of correctly predicted SAR image samples under different adversarial attacks.

[0051] The beneficial effects of the present application are: compared with the prior art, the present application not only realizes the automatic generation and optimized design of the neural network architecture for the SAR image detection system, improves the intelligent design level and design efficiency of the SAR image detection system, but also improves the adversarial robustness and accuracy of the complex SAR image detection system, reduces the complexity of the SAR recognition model, and greatly reduces the required resources when the SAR recognition model is deployed. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 is a schematic diagram of the principle of the SAR image detection system and method based on multi-objective optimization of neural network architecture optimization;

[0053] Figure 2 (a) in is a structural schematic diagram of the max-pooling layer, Figure 2 (b) in is a structural schematic diagram of the average-pooling layer;

[0054] Figure 3 is a structural schematic diagram of depthwise separable convolution, wherein n is the channel number of the input data set;

[0055] Figure 4 (a) in is a structural schematic diagram of dilated convolution with a convolution kernel size of 3x3, Figure 4 (b) in is a structural schematic diagram of dilated convolution with a convolution kernel size of 5x5;

[0056] Figure 5 is a structural schematic diagram of spatial separable convolutions with a convolution kernel size of 1x7 and 7x1 respectively;

[0057] Figure 6 is a structural schematic diagram of the squeeze-and-excitation (SE) layer;

[0058] Figure 7 is a structural schematic diagram of the non-local network layer;

[0059] Figure 8 is an example diagram of the encoding of a Normal Cell and a Reduction Cell in an individual generated randomly, wherein (a) represents the digital encoding of the Normal Cell and the Reduction Cell, (b) represents the neural network unit architecture corresponding to the digital encoding of the Normal Cell, (c) represents the neural network unit architecture corresponding to the digital encoding of the Reduction Cell, add represents the result of the addition of two nodes, concat represents the result of the addition of multiple nodes, inp1 and inp2 represent two different inputs of the cell respectively, and out represents the output of the cell;

[0060] Figure 9 is an example diagram of the conversion of the encoding information of an individual into a corresponding neural network model;

[0061] Figure 10 is an example diagram of a cell crossover operation in the multi-objective neural network architecture optimization process;

[0062] Figure 11 is an example diagram of a cell mutation operation in the multi-objective neural network architecture optimization process;

[0063] Figure 12 is a comparison diagram of the test accuracy of the present technology and the prior art AlexNet (a deep convolutional neural network named after the model designer Alex Krizhevsky), VGG16 (Visual Geometry Group Network 16), DenseNet201 (Dense Convolutional Network 201), ResNet18 (Deep Residual Network 18), ResNet50 (Deep Residual Network 50) and ResNet101 (Deep Residual Network 101) on normal SAR images;

[0064] Figure 13 is a comparison diagram of the model parameter size of the present technology and the prior art AlexNet, VGG16, DenseNet201, ResNet18, ResNet50 and ResNet101;

[0065] Figure 14 is a comparison diagram of the adversarial accuracy of the present technology and the prior art AlexNet, VGG16, DenseNet201, ResNet18, ResNet50 and ResNet101 on SAR images under the adversarial attack FGSM;

[0066] Figure 15 is a comparison chart of the adversarial accuracy of the present technology and the prior art AlexNet, VGG16, DenseNet201, ResNet18, ResNet50 and ResNet101 on SAR images under adversarial attack PGD;

[0067] Figure 16 is a comparison chart of the adversarial accuracy of the present technology and the prior art AlexNet, VGG16, DenseNet201, ResNet18, ResNet50 and ResNet101 on SAR images under adversarial attack CW;

[0068] Figure 17 is a comparison chart of the adversarial accuracy of the present technology and the prior art AlexNet, VGG16, DenseNet201, ResNet18, ResNet50 and ResNet101 on SAR images under adversarial attack GN;

[0069] Figure 18 is a comparison chart of the adversarial accuracy of the present technology and the prior art AlexNet, VGG16, DenseNet201, ResNet18, ResNet50 and ResNet101 on SAR images under adversarial attack Deepfool;

[0070] Figure 19 is a comparison chart of the adversarial accuracy of the present technology and the prior art AlexNet, VGG16, DenseNet201, ResNet18, ResNet50 and ResNet101 on SAR images under adversarial attack SSP. DETAILED DESCRIPTION

[0071] The purposes and effects of the present application will be more apparent from the following further description of the present application in conjunction with the accompanying drawings.

[0072] Taking the Gaofen-3 SAR image detection system as an embodiment, Figure 1 A principle diagram of a lightweight high-robustness SAR adversarial image detection system and method based on multi-objective neural network architecture optimization is shown. A lightweight high-robustness SAR adversarial image detection system based on multi-objective neural network architecture optimization includes a SAR data acquisition module, an offline optimization training module and an online identification module.

[0073] The data acquisition module obtains the center of the target to be detected from the high-resolution image of the SAR historical database, takes the center of the minimum circumscribed eccentric circle of the target center as the midpoint, expands outward by 256 pixels, obtains a SAR image slice dataset with a size of 512 pixels x 512 pixels, performs normalization processing on the image slice dataset and classifies it, divides it into a training set and a validation set, obtains an adversarial validation set after performing adversarial attack on the validation set, and takes the training set and the adversarial validation set as the input of the offline optimization training module; through SAR, real-time monitored images are obtained, after the determination of the center of the target to be detected, image slicing and normalization processing, an online detection dataset is obtained, which is taken as the input of the online detection model;

[0074] The offline optimization training module is based on a multi-objective neural network architecture optimization solver, obtains the optimal neural network model architecture information with light weight and high robustness, and transmits it to the online detection module;

[0075] The online recognition module deploys the light weight neural network best model online, performs online detection on the real-time acquired SAR images, and timely outputs the class information of SAR image classification and the performance indicators of evaluating image classification.

[0076] The system based on multi-objective neural network architecture optimization of the light weight and high robustness SAR image detection method, comprising the following steps:

[0077] (1) Data set acquisition and preprocessing. The data acquisition module of the SAR image detection obtains the source data set X through the acquisition of the center of the target to be detected, image slicing and class labeling of the high-resolution image generated by the Gaofen-3 satellite, scales the size of each image to 256 pixels x 256 pixels, and performs data normalization processing according to formula (1) to obtain the offline training data set X o , input the offline optimization training module based on multi-objective neural network architecture optimization, divide X o according to the proportion of 4:1, obtain the training set X o-train and the validation set X o-valid ;

[0078]

[0079] Wherein, X oj represents the j-th dimension feature of the offline training data set X o , X j represents the j-th dimension feature of the source data set X, and d represents the maximum value of the dimension of X o ;

[0080] (2) Set the parameter values of the offline optimization training module based on multi-objective neural network architecture optimization, including population size N = 30, the number of cells C = 3 constituting the neural network architecture, the number of chromosomes l = 7 for each cell, the number of iteration optimization I = 20 for architecture optimization, the number of epochs Epoch = 15 for neural network offline training, the size of the adversarial perturbation Epsilon = 16 / 255, the mutation rate σ = 0.02, the upper limit of the mutation interval σ max = [7, 1, 7, 1, 7, 2, 7, 2, 7, 3, 7, 3, 7, 4, 7, 4, 7, 5, 7, 5, 7, 6, 7, 6, 7, 7, 7, 7], the lower limit of the mutation interval σ min = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] and the custom parameter η = 1 of the crossover operation.

[0081] (3) Initialize N = 30 individuals as the initial population Q. The coding form of each individual is x i = {NormalCell, Reduction Cell, Optim, lr}, where NormalCell and Reduction Cell represent the architecture coding of the offline training neural network, Optim represents the optimizer type of the offline training of the neural network, and lr represents the learning rate of the offline training optimizer of the neural network.

[0082] NormalCell is represented as One NormalCell contains l = 7 chromosomes, and the i-th array represents the i-th chromosome, where 1 ≤ i ≤ 7, operation represents a type of operation, node represents a node to be operated, represents the 1st node of the i-th chromosome, represents the 2nd node of the i-th chromosome, represents the operation on , represents the operation on .

[0083] wherein operation is randomly selected from a set of operations (labeled as OPERATIONS), the composition of OPERATIONS is: [Maxpooling, Avgpooling, Identity, SepConv_3x3, SepConv_5x5, SepConv_7x7, DilConv_3x3, DilConv_5x5, Conv_1x7_7x1], which are represented by [0, 1, 2, 3, 4, 5, 6, 7, 8] respectively in the encoding. Among them, Maxpooling represents the maximum pooling operation, as shown in (a) of Figure 2 Avgpooling represents the average pooling operation, as shown in (b) of Figure 2 Identity represents not performing any operation, SepConv_3x3, SepConv_5x5, SepConv_7x7 represent depth separable convolution units with convolution kernel sizes of 3x3, 5x5, 7x7 respectively, Figure 3 a schematic diagram of a depth separable convolution unit is given, DilConv_3x3, DilConv_5x5 represent dilated convolution units with convolution kernel sizes of 3x3, 5x5 respectively, Figure 4 a schematic diagram of a dilated convolution unit is given, Conv_1x7_7x1 represents a spatial separable convolution unit composed of two convolution layers with convolution kernel sizes of 1x7 and 7x1 respectively by dividing a convolution kernel layer with a convolution kernel size of 7x7, Figure 5 a schematic diagram of a spatial separable convolution unit is given.

[0084] The first chromosome of a Normal Cell contains two initial nodes, labeled as and represents the 1st node of the first chromosome in the Normal Cell, represents the 2nd node of the first chromosome in the Normal Cell, which comes from the output of the previous two cells. By randomly selecting the existing node, performing operation on it, a new node is generated. A chromosome contains two operations and two nodes. By performing operation on the node, two new nodes are generated until the number of nodes in the chromosome is equal to 7. The nodes that have not been operated (except the initial nodes) are combined as the output of the entire Normal cell.

[0085] The composition of the Reduction Cell is One Reduction Cell contains l = 7 chromosomes, the i-th chromosome represents the i-th chromosome, where 1≤i≤7, execution represents an operation type, knot represents a node to be operated, represents the 1st node of the i-th chromosome, represents the 2nd node of the i-th chromosome, represents the operation on represents the operation on represents the operation on where execution is randomly selected from OPERATIONS, the first chromosome of a Reduction Cell contains two initial nodes, marked as and represents the 1st node of the first chromosome in the Reduction Cell, represents the 2nd node of the first chromosome in the Reduction Cell, which comes from the output of the previous two cells. By randomly selecting an existing knot, performing execution on it, a new knot is generated. A chromosome contains two executions and two knots. By performing execution on the knots, two new knots are generated. Until the number of knots in the chromosome equals 7. The knots that have not been executed (except for the initial nodes) are combined as the output of the entire Reduction Cell. The difference between the Reduction Cell and the Normal Cell is that the convolution step length Stride of the parameter convolution operation in each convolution-related operation in the Reduction cell becomes 2.

[0086] Optimizer is randomly selected from the set Optimizer = [Adadelta, Adam, SGD, RMSprop, Adagrad, AdamW, Adamax, ASGD, Rprop], and the corresponding code of Optimizer is [0, 1, 2, 3, 4, 5, 6, 7, 8]. Among them, AdaDelta represents an improved adaptive learning rate optimizer, Adam represents an adaptive moment estimation optimizer, SGD represents a stochastic gradient descent optimizer, Adagrad represents an adaptive learning rate optimizer, Rmsprop represents an RMSprop optimizer, AdamW represents an adaptive moment estimation optimizer with weight decay, Adamax represents an adaptive moment estimation optimizer based on infinite norm, ASGD represents an asynchronous stochastic gradient descent optimizer, and Rprop represents an elastic backpropagation optimizer. The learning rate lr of Optim is randomly selected from the set LearningRate = [0.1, 0.01, 0.001, 0.0001], and the corresponding code of LearningRate is [0, 1, 2, 3].

[0087] One x i The instantiation process is to stack the Normal Cell and the Reduction Cell in order until the number of cells is equal to 3. A compressed excitation network module (labeled as an SE layer) is added after each cell, Figure 6 A schematic diagram of the SE layer is given. After the stacking of the cells is completed, a Globalpooling layer and a Non-Local layer are added at the end of the network. The Globalpooling layer represents a global pooling module, and the Non-Local layer represents a non-local network module, Figure 7 A schematic diagram of the Non-Local layer is given.

[0088] In this embodiment, the encoding information corresponding to a randomly generated individual is {[2, 1, 3, 1, 62, 2, 2, 6, 3, 2, 0, 0, 0, 1, 2, 0, 2, 2, 2, 7, 5, 7, 5, 1, 1, 3, 6], [1, 1, 3, 1, 1, 0, 6, 2, 1, 1, 7, 1, 2, 1, 3, 2, 1, 1, 4, 2, 2, 2, 5, 6, 5, 2, 6, 3], 8, 3}, Figure 8 An example diagram of the encoding of the Normal Cell and the Reduction Cell in the individual is given, Figure 9 A schematic diagram of the conversion of the individual encoding information into the corresponding neural network model is given.

[0089] (4) Evaluate the performance of the initial population Q. Based on the architectural information represented by each individual in Q, convert it into the corresponding neural network model, and use the offline training dataset X. o-train As the data input for each convolutional neural network model, it undergoes offline training for Epoch rounds, and is used on the validation set X. o-valid Randomly select one of the four adversarial attacks to launch an attack, generating an adversarial verification set X. o-adv-valid Verification tests were conducted, and the results were calculated on the verification set X according to formulas (2) and (3) respectively. o-valid and X o-adv-valid The target values ​​are f1 and f2, and the four adversarial attacks are: Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), Gaussian Noise (GN), and Carlini & Wagner attack (CW). Formulas (2) and (3) are as follows:

[0090]

[0091] Where N err For X o-valid The number of incorrectly predicted samples, N all For X o-valid The total number of samples.

[0092]

[0093] Where N adv-err For X o-adv-valid The number of adversarial examples predicted incorrectly, where Ave and Std represent the number of examples predicted incorrectly in X. o-adv-valid The average and standard deviation of the test accuracy obtained from multiple test models under adversarial attacks.

[0094] (5) Calculate the f1 and f2 values ​​for each individual in population Q. If x i Both f1 and f2 values ​​are better than x. j If the f1 and f2 values ​​are given, then x is called x. i Disposable x j x j The number of dominated individuals n j While adding 1, x j Joined by x i Dominant set of individuals s iIn the middle, compare the dominance among individuals to obtain the number of dominated individuals n and the set of dominated individuals s for each individual. Find the individual in the population where n is 0 and store it in set F1; iterate through the set of individuals s dominated by each individual in F1, and perform n operations on each individual g in s. g =n g -1, if there exists n g Individuals with a value of 0 are added to set F2; the above operation is repeated on F2 until the entire population Q is divided into sets F of different levels.

[0095] (6) For a population Q that has already been ranked and classified by dominance, for each set of ranks F k Where 1≤k≤30, sort in ascending order according to f1 and f2 respectively, and then sort F... k Each individual m in the process is subjected to crowding degree calculation according to formula (4).

[0096]

[0097] Where, m d This represents the crowding degree of individual m. f represents the target value of individual m+1. j , f represents the target value of individual m-1. j .

[0098] (7) Perform a traversal operation based on the order of individuals in population Q, selecting two individuals as parent individuals each time, and marking them as follows: and An individual contains two types of cell codes: Normal Cell code and Reduction Cell code. A cell contains 7 chromosomes. In Normal Cell code, 4 bits are used to represent two nodes and two operations on a chromosome; in Reduction Cell code, 4 bits are used to represent two knots and two executions on a chromosome. n is the number of gene codes in a cell, with a value of 56. Crossover calculations between two parent individuals are performed according to formula (5), resulting in two offspring individuals, namely... and

[0099]

[0100] Where β is dynamically and randomly determined by the distribution factor η = 1 according to formula (6).

[0101]

[0102] where η is a self-defined parameter, and η = 1. The greater the value of η, the closer the offspring to the parent. The offspring c 1 and c 2 are stored in a new population set P, and the above operations are repeated until the size of the set P is equal to the size of the population Q. Figure 10 An illustrative diagram of the two parents crossing to produce offspring is shown.

[0103] (8) The individuals in the new population P are traversed and subjected to mutation operation. For an individual x A 0-1 floating point number p is randomly generated, and the relationship between p and the mutation rate σ = 0.02 is determined. If p > 0.02, the individual does not need to be mutated; otherwise, the individual is subjected to mutation calculation according to formula (7). Figure 11 An illustrative diagram of the two parents crossing to produce offspring is shown.

[0104]

[0105] where, is the kth encoding of the individual x i in the population P, is the kth encoding of the offspring individual x im obtained according to , l is the number of individual chromosomes, u k is the kth number in σ max , and l k is the kth number in σ min , where δ is calculated according to formula (8).

[0106]

[0107] where λ m is a distribution index, rand is a random number between 0 and 1, δ1 and δ2 are calculated according to formula (9) and formula (10), respectively.

[0108] δ1 = (v k -l k ) / (u k -l k ) (9)

[0109] δ2 = (u k -l k ) / (v k -l k ) (10)

[0110] (9) According to the performance evaluation method described in step (4), the f1 and f2 target values of the population P after crossover and mutation are calculated. The population P is combined with the population Q to obtain the population R, and the population R is non-dominantly sorted according to step (5), and the population R after non-dominant sorting is calculated according to step (6). In each set F, the first 30 individuals of the sorted population R are selected according to the size of the crowding degree from small to large, to form a new offspring population Q n ;

[0111] (10) Unconditionally accept Q = Q n ;

[0112] (11) Repeat steps (4) to (9) until the set iteration optimization number I = 20 of the architecture optimization is met;

[0113] (12) The individual with the smallest crowding degree in the final F1 set obtained in the offline optimization training module is taken as the optimal individual X best , and the optimal architecture information represented by X best {[4, 1, 1, 1, 3, 1, 1, 1, 5, 3, 4, 2, 2, 1, 2, 1, 3, 2, 4, 1, 2, 1, 1, 1, 4, 2, 4, 3], [1, 0, 5, 0, 4, 1, 3, 0, 7, 0, 2, 1, 7, 4, 8, 2, 2, 1, 7, 1, 6, 2, 3, 2, 4, 1, 1, 3], 1, 1} is used to construct the optimal neural network model in the SAR image detection module, and the parameters of the optimal neural network model are set, including the number of rounds of online detection Epoch2 = 30, the learning rate = 0.01, the model optimizer is Adam, and the batch size is 48;

[0114] (13) The data acquisition module of the SAR image detection method obtains SAR image information from SAR real-time monitoring, and after center acquisition, image slicing and data normalization processing of the target to be detected, it is used as the online detection data set of the SAR image detection module based on the optimal neural network model, and is marked as X t , FGSM, PGD, CW, GN, and self-supervised perturbation attack (marked as DeepFool) based on orthogonal projection to the decision boundary are performed on X t , respectively, to obtain the corresponding adversarial data set X adv ;

[0115] (14) The optimal neural network model is run by using the online detection module, and the online SAR image detection is performed on the online detection data set X t , and the test accuracy of the SAR image is calculated according to formula (11); the online adversarial detection data set Xadv The online SAR image detection is performed, and the adversarial accuracy of the SAR image is calculated according to formula (12).

[0116]

[0117]

[0118] wherein N true represents the number of normal SAR image samples correctly predicted, N all represents the total number of SAR image samples, N adv-true represents the number of SAR image samples correctly predicted under different adversarial attacks. In this embodiment, the test accuracy obtained by the present technology is 70.24%, and the accuracy under FGSM, PGD, CW, GN, Deepfool and SSP adversarial attacks is 55.95%, 53.57%, 67.85%, 67.85%, 22.62% and 69.05%, respectively.

[0119] The experimental results of the present technology and the prior art AlexNet (a deep convolutional neural network named after the model designer Alex Krizhevsky), VGG16 (Visual Geometry Group Network 16), DenseNet201 (Dense Convolutional Network 201), ResNet18 (Deep Residual Network 18), ResNet50 (Deep Residual Network 50) and ResNet101 (Deep Residual Network 101) for the above embodiment are compared and analyzed, and the test accuracy comparison results are shown in Figure 12 , and the neural network model parameter size comparison chart is shown in Figure 13 . The adversarial accuracy comparison results of the present technology and the prior art on SAR images under adversarial attacks FGSM, PGD, CW, GN, Deepfool and SSP are shown in Figure 14 , Figure 15 , Figure 16 , Figure 17 , Figure 18 and Figure 19 . We can find that compared with various traditional convolutional neural networks relying on artificial experience design, such as AlexNet, VGG16, DenseNet201, ResNet18, ResNet50, ResNet101 and other prior arts, the test accuracy and adversarial accuracy performance indicators obtained by the present technology are the best, and the neural network model parameters obtained are the least.

[0120] In addition, compared with the prior art, the present application realizes the automatic generation and optimized design of the neural network architecture for the SAR image detection system, and improves the intelligent design level and design efficiency of the SAR image detection system.

[0121] In summary, the present application not only realizes the automatic generation and optimized design of the neural network architecture for the SAR image detection system, but also improves the intelligent design level and design efficiency of the SAR image detection system, and improves the adversarial robustness and accuracy of the complex SAR image detection system, reduces the complexity of the SAR recognition model, and greatly reduces the required resources when the SAR recognition model is deployed.

Claims

1. A lightweight high-robustness SAR adversarial image detection method based on multi-objective neural network architecture optimization, characterized in that, Comprise the following steps: (1) Data set collection and preprocessing: the data collection module of the SAR image detection acquires the high-resolution images generated by the Gaofen-3 satellite, obtains the center of the target to be detected, slices the images, and labels the categories, obtains the source data set X, scales the size of each image to 256 pixels x 256 pixels, and performs data normalization processing according to formula (1) to obtain the offline training data set X o , input the offline optimization training module based on the multi-target neural network architecture optimization, and divide X o according to the proportion of 4:1 to obtain the training set X o-train and the verification set X o-valid ; where X oj represents an offline training dataset X o represents the jth dimension of the feature of the source dataset X j represents the jth dimension of the feature of the source dataset X o represents the maximum value of the dimension; (2) Set the parameter values of the offline optimization training module based on the multi-objective neural network architecture optimization: the parameter values include the population size N, the number of units C constituting the neural network architecture, the chromosome number l of each unit, the iteration optimization number I of the architecture optimization, the round Epoch of the neural network offline training, the custom parameters of the crossover operation η, the mutation rate σ and the upper limit σ of the mutation interval max , the lower limit σ of the mutation interval min ; (3) initialize N individuals as initial population Q: the coding form of each individual is x i ={Normal Cell, Reduction Cell, Optim, lr}, wherein Normal Cell and Reduction Cell represent the architecture coding constituting the offline training neural network, Optim represents the optimizer type of the offline training of the neural network, and lr represents the learning rate of the offline training optimizer of the neural network; Normal Cell is represented as One Normal Cell contains l chromosomes, the i-th array represents the i-th chromosome, where 1≤i≤l, operation represents an operation type, and node represents a node to be operated, represents the first node of the i-th chromosome, represents the second node of the i-th chromosome, represents the operation on , represents the operation on ; Wherein operation is randomly selected from the operation set, the composition of OPERATIONS is: [Maxpooling, Avgpooling, Identity, SepConv_3x3, SepConv_5x5, SepConv_7x7, DilConv_3x3, DilConv_5x5, Conv_1x7_7x1], respectively represented by [0,1,2,3,4,5,6,7,8] in the code;Wherein, Maxpooling represents the maximum pooling operation, Avgpooling represents the average pooling operation, Identity represents not performing any operation, SepConv_3x3, SepConv_5x5, SepConv_7x7 represent depth separable convolution units with convolution kernel size of 3x3, 5x5, 7x7 respectively, DilConv_3x3, DilConv_5x5 represent dilated convolution units with convolution kernel size of 3x3, 5x5 respectively, Conv_1x7_7x1 represents a spatial separable convolution unit composed of two convolution kernels with size of 1x7 and 7x1 divided from a convolution layer with convolution kernel size of 7x7; The first chromosome of a Normal Cell contains two initial nodes, denoted as and denotes the first node of the first chromosome in the Normal Cell, denotes the second node of the first chromosome in the Normal Cell, which is from the outputs of the previous two cells; a new node is generated by randomly selecting an existing node and performing an operation on it; a chromosome contains two operations and two nodes; two new nodes are generated by performing operation on the nodes until the number of nodes in the chromosome equals to / ; the nodes that have not been operated are combined, except the initial nodes, as the output of the entire Normal Cell; The Reduction Cell consists of A reduction cell contains l chromosomes, and the i-th array Let represent the i-th chromosome, where 1 ≤ i ≤ l, execution represents an operation type, and knot represents a node to be operated on. This represents the first node of the i-th chromosome. This represents the second node of the i-th chromosome. Indicates to The operation, Indicates to The operation; where execution is randomly selected from OPERATIONS, the first chromosome of a Reduction Cell contains two initial nodes, labeled as follows: and This represents the first node of the first chromosome in the Reduction Cell. This represents the second node of the first chromosome in the Reduction Cell, which comes from the outputs of the first two cells. By randomly selecting an existing knot and performing an execution operation on it, new knots are generated. A chromosome contains two execution operations and two knots. By performing execution operations on the knots, two new knots are generated in total, until the number of knots in the chromosome equals 1. Knots that have not undergone execution operations are combined, except for the initial node, and used as the output of the entire Reduction Cell. The difference between the Reduction Cell and the Normal Cell is that the convolution stride parameter for convolution-related operations in each Reduction Cell is changed to 2. Optimizer is randomly selected from the set Optimizer=[Adadelta, Adam, SGD, RMSprop, Adagrad, AdamW, Adamax, ASGD, Rprop], and the code corresponding to Optimizer is [0,1,2,3,4,5,6,7,8];Wherein AdaDelta represents an improved adaptive learning rate optimizer, Adam represents an adaptive moment estimation optimizer, SGD represents a stochastic gradient descent optimizer, Adagrad represents an adaptive learning rate optimizer, Rmsprop represents an RMSprop optimizer, AdamW represents an adaptive moment estimation optimizer with weight decay, Adamax represents an adaptive moment estimation optimizer based on infinite norm, ASGD represents an asynchronous stochastic gradient descent optimizer, and Rprop represents an elastic backpropagation optimizer;The learning rate lr of Optim is randomly selected from the set LearningRate=[0.1,0.01,0.001,0.0001], and the code corresponding to LearningRate is [0,1,2,3]; One x i The instantiated process is to stack the Normal Cell and the Reduction Cell in order until the number of cells is equal to C; a compression excitation network module is added after each cell, and after the stacking of the cells is completed, a Global pooling layer and a Non-Local layer are added at the end of the network; wherein the Global pooling layer represents a global pooling module, and the Non-Local layer represents a non-local network module. (4) Performance evaluation on initial population Q: according to the architecture information represented by each individual in Q, the corresponding neural network model is converted, and the offline training data set X o-train As the data input of each convolutional neural network model, the offline training is performed for Epoch rounds, and the validation set X o-valid One of the four kinds of adversarial attacks is randomly selected for attack, and an adversarial validation set X o-adv-valid Validation test is performed, and the target values f1 and f2 on the validation sets X o-valid and X o-adv-valid are calculated according to formulas (2) and (3) respectively, where the four kinds of adversarial attacks are: Fast Gradient Sign Method (FGSM), Projected Gradient Descent (PGD), Gaussian Noise (GN) and Carlini & Wagner attack (referred to as CW); formulas (2), (3) are as follows: Where N err For X o-valid The number of incorrectly predicted samples, N all For X o-valid The total number of samples; where N adv-err is X o-adv-valid the number of adversarial samples that are predicted incorrectly, Ave, Std represent the average and standard deviation of the test accuracy of multiple test models under adversarial attacks, respectively. o-adv-valid the average and standard deviation of the test accuracy of multiple test models under adversarial attacks, respectively. (5) Statistics of each individual in the population Q f1 value and f2 value, if x i f1 value and f2 value of x j x i x j ; x j n j x j x i x i dominated individual number n g 1, while x g x g F2 in the set; F2 in the above repeated operation, until the entire population Q is divided into different levels of set F; (6) For the population Q that has been dominance ordered and partitioned into ranks: for each rank set F k where 1 < k < N, each individual m in F k is ordered according to f1 and f2 in ascending order, respectively, and the crowding distance of each individual m in F (7) For each individual m in F (8) For each individual m in F (9) For each individual m in F (10) For each individual m in F (11) For each individual m in F (12) For each individual m in F (13) For each individual m in F (14) For each individual m in F (15) where m d denotes the crowding distance of individual m, denotes the target value f j , denotes the target value f j of individual m-1. (7) The traversal operation is performed according to the arrangement order of the individuals in the population Q: two individuals are selected as parent individuals each time, and are marked as and An individual contains two cell encodings, namely Normal Cell encoding and Reduction Cell encoding; a cell contains l chromosomes, in the Normal Cell encoding, four-bit encoding is used to represent two nodes and two operations in a chromosome; in the Reduction Cell encoding, four-bit encoding is used to represent two knots and two executions in a chromosome; n is the number of genetic encodings of a cell, and is 8l; according to formula (5), the crossover calculation between the two parent individuals is performed to obtain two offspring individuals, which are and Wherein β is dynamically and randomly determined by the distribution factor η according to formula (6); where rand represents a random number generated in the range of 0 to 1, η is a self-defined parameter, the greater the value of η, the closer the generated offspring individual to the parent individual; the generated offspring c 1 and c 2 are stored in a new population set P, and the above operation is repeated until the size of the set P is equal to the size of the population Q; (8) Traverse the individuals in the new population P, and perform mutation operation on them: for each individual A 0-1 floating point number p is randomly generated, and the relationship between p and mutation rate σ is judged. If p > σ, the individual does not need to be mutated; otherwise, the individual is mutated according to formula (7). wherein is the kth encoding of individual x in population P i is the kth encoding of offspring individual x calculated from im wherein l is the number of individual chromosomes, u k is the kth number in σ max k is the kth number in σ min wherein δ is calculated according to equation (8);​​​ where λ m is the distribution exponent, rand is a random number between [0, 1], and δ1, δ2 are calculated according to formula (9) and formula (10), respectively. δ1= (v k -l k ) / (u k -l k ) (9) δ2= (u k -l k ) / (v k -l k ) (10) (9) According to the performance evaluation method described in step (4), the f1, f2 target value calculation of the population P after crossover and mutation is performed: the population P is combined with the population Q to obtain the population R, the population R is non-dominantly sorted according to step (5), the population R after non-dominant sorting is calculated according to step (6), in each set F, the first N individuals of the sorted population R are selected according to the size of the crowding degree from small to large, to constitute a new offspring population Q n ; (10) unconditionally accept Q = Q n ; (11) Repeat steps (4) to (9) until the set number of iterations I of architecture optimization is met; (12) The individual with the minimum congestion in the final F1 set obtained in the offline optimization training module is taken as the optimal individual X best , and according to the optimal architecture information characterized by X best , the optimal neural network model in the SAR image detection module is constructed, and the parameters of the optimal neural network model are set, including the number of epochs Epoch2, the learning rate, the type of optimizer and the momentum setting, etc. (13) The data acquisition module of the SAR image detection method obtains SAR image information from SAR real-time monitoring. After obtaining the center of the target to be detected, image slicing, and data normalization processing, it serves as the online detection dataset of the SAR image detection module based on the optimal neural network model, marked as X t , X t is subjected to FGSM, PGD, CW, GN, orthogonal projection-based adversarial attack, and self-supervised perturbation adversarial attack (SSP), respectively, to obtain the corresponding adversarial dataset X adv ; (14) using the online detection module to run the optimal neural network model, performing online detection on the online detection data set X t , and calculating the test accuracy of the SAR image according to formula (11); performing online SAR image detection on the online adversarial detection data set X adv , and calculating the adversarial accuracy of the SAR image according to formula (12); where N true represents the number of correctly predicted normal SAR image samples, N all represents the total number of SAR image samples, N adv-true represents the number of correctly predicted SAR image samples under different adversarial attacks.

2. A lightweight robust synthetic aperture radar (SAR) adversarial image detection system based on multi-objective neural network architecture optimization for implementing the method of claim 1, characterized in that, The system comprises a data acquisition module, an offline optimization training module and an online identification module of SAR. The data acquisition module obtains a to-be-detected target center from a high-resolution image of a SAR historical database, takes the center of a minimum circumscribed eccentric circle of the target center as a midpoint, expands 256 pixel points outward, obtains a SAR image slice dataset with a size of 512 pixels x 512 pixels, performs normalization processing on the image slice dataset, performs class labeling, divides the image slice dataset into a training set and a verification set, obtains an adversarial verification set after performing adversarial attack on the verification set, and takes the training set and the adversarial verification set as inputs of an offline optimization training module; an image monitored in real time is obtained through SAR, a to-be-detected target center is determined, image slicing and normalization processing are performed, and an online detection dataset is obtained, which is taken as an input of an online detection model; The offline optimization training module is based on a multi-target neural network architecture optimization solver, obtains optimal neural network model architecture information with light weight and high robustness, and transmits the information to the online detection module; The online identification module deploys the light-weight neural network optimal model online, performs online detection on a real-time acquired SAR image, and timely outputs class information of SAR image classification and performance indicators for evaluating image classification.

Images