US20220122348A1 - Adversarial Optimization Method for Training Process of Generative Adversarial Network - Google Patents
Adversarial Optimization Method for Training Process of Generative Adversarial Network Download PDFInfo
- Publication number
- US20220122348A1 US20220122348A1 US17/288,566 US202017288566A US2022122348A1 US 20220122348 A1 US20220122348 A1 US 20220122348A1 US 202017288566 A US202017288566 A US 202017288566A US 2022122348 A1 US2022122348 A1 US 2022122348A1
- Authority
- US
- United States
- Prior art keywords
- network
- generator
- data
- training
- defence
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 49
- 238000012549 training Methods 0.000 title claims abstract description 45
- 238000005457 optimization Methods 0.000 title claims abstract description 11
- 238000013507 mapping Methods 0.000 claims abstract description 26
- 230000005540 biological transmission Effects 0.000 claims abstract description 17
- 238000005259 measurement Methods 0.000 claims abstract description 17
- 238000007476 Maximum Likelihood Methods 0.000 claims description 4
- MQHWFIOJQSCFNM-UHFFFAOYSA-L Magnesium salicylate Chemical compound [Mg+2].OC1=CC=CC=C1C([O-])=O.OC1=CC=CC=C1C([O-])=O MQHWFIOJQSCFNM-UHFFFAOYSA-L 0.000 abstract description 8
- 208000009119 Giant Axonal Neuropathy Diseases 0.000 abstract 1
- 201000003382 giant axonal neuropathy 1 Diseases 0.000 abstract 1
- 230000006870 function Effects 0.000 description 26
- 238000013528 artificial neural network Methods 0.000 description 8
- 238000010586 diagram Methods 0.000 description 6
- 230000011218 segmentation Effects 0.000 description 5
- 238000004422 calculation algorithm Methods 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 4
- 235000000332 black box Nutrition 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 238000013145 classification model Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 230000009977 dual effect Effects 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 239000002131 composite material Substances 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 238000005315 distribution function Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000002955 isolation Methods 0.000 description 1
- 238000007477 logistic regression Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/11—Complex mathematical operations for solving equations, e.g. nonlinear equations, general mathematical optimization problems
- G06F17/13—Differential equations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G06N3/0454—
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Definitions
- the invention relates to an image processing technology, in particular to an adversarial optimization method for the training process of generative adversarial network.
- a variance is calculated for a given sample in the adversarial samples.
- a deep learning model learns a segmentation plane in a high-dimensional space by training on samples, and different measurements on the segmentation plane serve as different classification discrimination results, as shown in FIG. 1 .
- One or more minor shifts cannot be identified by the human perception, but the samples can cross the segmentation plane in the deep learning result space, causing a change to a decision result of the machine learning model, as shown in FIG. 2 .
- the adversarial samples mainly launch a gradient attack and an encoder attack.
- the attack samples generated by the encoder of neural network are better than those generated by the gradient attack.
- GAN generative adversarial network
- GAN-based attack sample models have emerged from 2018 to 2019.
- their model robustness is too poor because of the instability of GAN. Therefore, based on the optimal transmission theory, this invention addresses the convergence problem of GAN by optimal mapping in two measurements.
- GAN is a generation model that contains two networks (a generator network and a discriminator network). At a given noise source, the generator network generates composite data, while the discriminator network distinguishes the generated data from the real data.
- GAN is affected by the instability of training Most of the latest work on GAN training is to find a stable training method. Currently, most of the commonly used methods rely on heuristics which are very sensitive to modification, and rarely explain this new unstable behavior for the generative adversarial network training from the internal source of the network. This greatly limits the applicability of GAN in image processing applications.
- the invention provides an adversarial optimization method for the training process of generative adversarial network.
- the optimal transmission problem is transformed into solving the elliptic Monge-Ampere partial differential equation (MAPDE) in the generator G.
- MAPDE elliptic Monge-Ampere partial differential equation
- the Neumann boundary conditions are improved and the discretization of MAPDE is extended to obtain the optimal mapping between a generator and a discriminator, which constitutes the adversarial network MAGAN.
- the defence network can obtain a maximum distance between the two measurements and obtain filtered security samples.
- the solution of MAPDE can constitute a new discriminant distribution function instead of the Wasserstein distance of WGAN.
- the technical scheme of the invention is an adversarial optimization method for the training process of generative adversarial network, specifically comprising the steps of:
- the generator has a loss function of
- E x ⁇ Px is an expectation of the probability distribution of real data
- E y ⁇ Py is an expectation of probability distribution of attack sample data
- E x is an expectation for Licpschiz continuous data
- D w is a discriminator network with weights
- D is a discriminator network
- G is a generator network
- ⁇ is a penalty coefficient, a hyper-parameter set by the training network
- E is an expectation
- the defence network has a loss function of
- m is the number of discrete points in each dimension of the network.
- the invention has the beneficial effects that the invention generates an adversarial optimization method for the training process of generative adversarial network, an effective attacking method of GANs is successfully established, and several calculation operators are provided to prove that the precision is improved by 5.3%.
- MAGAN can be stably trained without adjusting hyper-parameters, and the convergence speed of the MAGAN is 317.2% faster than that of the WGAN-GP in the LSUN Bedrooms database.
- the IS value of MAGAN on a CIFAR-10 data set also reaches 8.7, which improves the accuracy of target classification and recognition system for unmanned vehicle.
- FIG. 1 is a schematic diagram of a neural network classification segmentation plane
- FIG. 2 is a schematic diagram of an attack sample cross-domain segmentation plane
- FIG. 3 is a distribution diagram of a numerical initial solution of MAPDE
- FIG. 4 is a schematic diagram of a unit inward normal vector of MAPDE under Neumann boundary conditions
- FIG. 5 is a frame structure diagram of an improved generative adversarial network according to the present invention.
- FIG. 6 is a structure diagram for generative adversarial samples and defence networks according to the present invention.
- FIG. 7 is a flow chart for attacking and defending a neural network of a target classification and a recognition system for unmanned vehicle according to the present invention.
- the neural network implements an algorithm recognition part of an unmanned vehicle target classification and a recognition system for image recognition and classification.
- Convolution neural network (CNN) is mainly used in image processing and target recognition, and under-fitting exists in the space of real data and generated data. Therefore, the proposed attack and defence algorithm has been studied to enhance the robustness of neural networks, but the black-box attack and defence effect is not good.
- ⁇ is a finite open set.
- the transmission cost of optimal transmission mapping T: ⁇ is defined as:
- I is a mapping cost function
- c (x, y) is a distance for mapping x to y
- x, y belong to points within the set.
- Brenier proves that there is a convex function u: ⁇ R (R is a set of all real numbers), whose gradient mapping T:x ⁇ u(x) is the only optimal transmission mapping.
- This convex function is called Brenier potential.
- the Jacobian equation gives that the Brenier potential satisfies the Monge-Ampere PDE (2), and the Jacobian matrix of the gradient mapping is the Hessian matrix of the Brenier potential.
- D is a partial derivative
- det is a determinant
- ⁇ is a measurement
- X and Y are complete sets of x and y, respectively.
- WGAN transforms the original function into the dual problem.
- the constraints of the dual problem are difficult to implement, so sometimes WGAN still produces poor samples or even fail to converge. Therefore, the method of Kantorovich is not a good choice.
- Brenier's method is chosen to obtain the best mapping between the actual data distribution and the generation distribution.
- the transmission cost is the difference square
- the problem can be transformed into solving the Monge-Ampere PDE.
- the Monge-Ampere PDE can avoid Lipschitz constraints completely.
- the scheme of the Monge-Ampere PDE is more accurate than the method of Kantorovich (WGAN-GP).
- the transmission condition (BC) is also named the second boundary value problem or the second kind of boundary condition.
- the boundary conditions can be mapped into the set X by boundary nodes, and Y by the boundary nodes. In view of the gradient appearing in the transmission boundary condition, it is expected to find the Neumann boundary conditions:
- a boundary normal vector n is defined, wherein normal vector components corresponding to n dimensions in the boundary normal vector n are n 1 , n 2 , . . . , n n , where the vector n is perpendicular to the point x (x belongs to a point in the set X because the set X has a boundary, and an unspecified point x on the boundary is also in the set X Therefore, the equation 5 is the boundary conditions meeting the condition, the domain of definition is also X, and x is just the point in the met condition.), ⁇ X is the boundary of the set X, and ⁇ x solves a partial derivative with respect to x.
- ⁇ is a normal vector calculation function, calculated by the equation (7).
- the normal vector component n corresponding to each dimension in the equation (7) will be multiplied by the partial derivative of u in the corresponding dimension, i.e., a difference value between one unit step by which the function u is shifted forward in each dimension and a center value.
- the subscripts of the front and back points are i 1 , (i+1) 1 , i.e. x i 1 , . . . ,i n and x (1+1) 1 , . . . ,i n .
- u k+1 is a solution of the k+1 iteration. Because it is monotonic, the scheme relies only on values within a square. When the dimension of MAPDE is greater than 2, there are many nodes around the boundary that directly affect the value of the solution, as shown in FIG. 3 . This will take more computational time, so more attention should be paid to the boundary, and upper limits of these monotonic methods should be used for all acceptable ranges to ensure that a highly accurate numerical solution is obtained. The monotonicity of the method is preserved. First, a boundary is set on the other side of the square area.
- the present invention defines several finite difference operators that are used to approximate the first and second order partial derivatives using the central difference value. Standard discretization is performed to the equation by the center difference:
- MA is a discrete Monge-Ampere PDE
- D is a partial derivative
- f, g are measurements.
- U is a discrete numerical solution to the Monge-Ampere PDE.
- the partial derivatives of all nodes need to be calculated on the basis of Newton iteration. However, when the dimension is greater than 3, it is difficult to obtain all the partial derivatives in the high dimension.
- the solution of MAPDE can be discretized by setting a variable, the high dimension also makes it difficult for nodes in the grid to define the sequence of the nodes. It can be found that most relevant nodes in each dimension are its forward and backward nodes. The nodes with the central difference in different dimensions are the same node. Therefore, it is suggested to use the surrounding gradient instead of the overall gradients in order to accelerate the high-dimensional convergence.
- Equation 14 is maximum likelihood estimation for solving the real data and the generated data, and the maximum value of Equation 15 is solved by MLE to determine the distribution of the real data and the generated data.
- the best mapping u(x) is obtained by solving the Monge-Ampere PDE, as shown by a dashed rectangle in FIG. 5 .
- An image data training set and random noise are sent into a generator, and the generator outputs data as attack samples and real image data, correspondingly obtaining a brand-new discriminator D formed by entering two data sets X and Y into a dashed rectangle.
- the brand-new discriminator D distinguishes the measurement of the real data from the measurement of the generated data, and the generator G continuously confronts the discriminator D in the training process, thereby performing effective attacks.
- the best mapping u(x) is obtained by solving the Monge-Ampere PDE, and the generation of attack adversarial samples is achieved by the trained generator G in FIG. 5 . That is, the interior of the generator constitutes the attack network.
- the upper bound of the loss function of the generator of the equation 16 will be solved; on the contrary, in the process of training the generator G shown in FIG. 5 , the lower bound of the loss function of the generator of the equation 16 will be solved, finally achieving a good adversarial effect.
- the loss function, as a new divergence, of the Monge-Kantorovich transmission problem replaces the Wasserstein distance of WGAN.
- Implementation steps include:
- Step 1 calculating the probability density ⁇ X of X and the probability density ⁇ Y of Y
- Step 2 solving maximum likelihood estimation maximum values of the real data and the generated data
- Step 3 calculating the measurements of the real data and the generated data
- Step 4 calculating the first type of boundary of the MAPDE
- Step 5 calculating the optimal u(x) between the real data distribution P r and the generation distribution P g ;
- Step 6 iterating gradient values
- Step 7 calculating the loss function
- This network is applied in an adversarial sample generation network by MAGAN.
- a black-box attack and a white-box attack are realized more efficiently by good robustness of the MAGAN, forming an attack network.
- the defence network is trained by the attack samples generated by the generator G, and the robustness is enhanced by the solution of the Monge-Ampere PDE, as shown in FIG. 6 .
- the discriminator D in FIG. 6 corresponds to the structure in a dashed box in FIG. 5 , which contains a numerical solution of the partial differential equation.
- the discriminator is trained and applied to a lower discriminator network of FIG. 6 , the filtered samples are used for training the defence network.
- the samples of the generator G are taken as the input data of the defence network, and since the input training data has strong robustness of black-box and white-box attacks, the cross entropy is calculated by logistic regression on the input data by means of the loss function of the defence network, the real data and the artificial data are distinguished, achieving an isolation effect and a good defence effect.
- the loss function for the generator is
- E x ⁇ Px is an expectation of the probability distribution of real data
- E y ⁇ Py is an expectation of probability distribution of attack sample data
- E x is an expectation for Licpschiz continuous data
- D w is a discriminator network with weights
- D is a discriminator network
- G is a generator network
- ⁇ is a penalty coefficient, a hyper-parameter set by the training network
- E is an expectation
- the defence network has a loss function of
- the defence network in FIG. 6 is implemented by adding an attacked target network, as shown in FIG. 7 .
- the generator G can learn the measurements of the attack samples to calculate the optimal mapping U. Then the generator can obtain attack samples of the target network according to the input data and the random noise so as to carry out effective attacks.
- the target network is a network required to be attacked, i.e. a known network trained by others.
- the left box of FIG. 7 the left box of FIG.
- the output data of the generator G is taken as the input data of the defence network, and the defence network is trained according to the loss function (equation 17) obtained from the solution of the Monge-Ampere PDE and the optimal transmission theory in the right box.
- the equation 16 is a generator network.
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computational Mathematics (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Databases & Information Systems (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Operations Research (AREA)
- Algebra (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
- Complex Calculations (AREA)
Abstract
Description
- The invention relates to an image processing technology, in particular to an adversarial optimization method for the training process of generative adversarial network.
- In recent years, as the core technology of artificial intelligence, deep learning has made a lot of key breakthroughs in the fields of image, speech, natural language processing and so on. Many different methods of generating adversarial examples have been proposed to attack deep neural networks. These methods involve directly calculating pixels of the gradient image, or directly solving for the optimization of the image pixels.
- With the development of deep learning, it is used in more and more fields instead of traditional intelligent algorithms. However, high precision and low risk are required in some fields such as finance, unmanned vehicle and the like, which must have higher security, especially in automatic driving. Therefore, the impact of adversarial samples on the network model cannot be ignored. In deep learning networks, the adversarial samples affect a final learning model through subtle noise. In addition, the adversarial samples adopted by attackers cannot be judged by human senses, so that judgment and defence can be carried out only by the neural network itself. A typical scene contains adversarial samples of an image classification model. By superimposing carefully constructed variations on the image, a misjudgment is produced by the classification model as it is difficult for the naked eye to perceive.
- In principle, a variance is calculated for a given sample in the adversarial samples. A deep learning model learns a segmentation plane in a high-dimensional space by training on samples, and different measurements on the segmentation plane serve as different classification discrimination results, as shown in
FIG. 1 . - One or more minor shifts cannot be identified by the human perception, but the samples can cross the segmentation plane in the deep learning result space, causing a change to a decision result of the machine learning model, as shown in
FIG. 2 . - So far, these optimization problems have been solved in three broad ways:
- (1) by directly using optimizers such as L-BFGS or Adam (Kingma & Ba, 2015), such as Szegedy (2013) and Carlini & Wagner (2016), and these optimizer-based methods tend to be slower and more powerful than other methods.
- (2) by technique approximation based on a single step gradient, such as a fast gradient sign (Goodfellow et al., 2014b) or a most unlikely class (Kurakin et al., 2016a). These methods are fast, requiring only a single forward and backward pass through target classifiers to compute the perturbation.
- (3) by approximation of iterative variants based on gradient techniques (Kurakin et al., 2016a; MoosaviDezfooli et al., 2016a; b). According to these methods, multiple forward and backward passes through the target network are used to move the input more carefully toward the adversarial classification.
- At present, the adversarial samples mainly launch a gradient attack and an encoder attack. Among them, the attack samples generated by the encoder of neural network are better than those generated by the gradient attack. Especially after 2017, with the development of generative adversarial network (GAN), GAN has gradually become a main network generation tool. Therefore, GAN-based attack sample models have emerged from 2018 to 2019. However, their model robustness is too poor because of the instability of GAN. Therefore, based on the optimal transmission theory, this invention addresses the convergence problem of GAN by optimal mapping in two measurements.
- GAN is a generation model that contains two networks (a generator network and a discriminator network). At a given noise source, the generator network generates composite data, while the discriminator network distinguishes the generated data from the real data. However, GAN is affected by the instability of training Most of the latest work on GAN training is to find a stable training method. Currently, most of the commonly used methods rely on heuristics which are very sensitive to modification, and rarely explain this new unstable behavior for the generative adversarial network training from the internal source of the network. This greatly limits the applicability of GAN in image processing applications.
- Aiming at the problem that the convergence of the training of the Generative Adversarial Networks (GANs) used for attack and defence is unstable, the invention provides an adversarial optimization method for the training process of generative adversarial network. The optimal transmission problem is transformed into solving the elliptic Monge-Ampere partial differential equation (MAPDE) in the generator G. To solve MAPDE of n (n>3) dimensions, the Neumann boundary conditions are improved and the discretization of MAPDE is extended to obtain the optimal mapping between a generator and a discriminator, which constitutes the adversarial network MAGAN. In the process of training the defence network, by overcoming the loss function of the optimal mapping, the defence network can obtain a maximum distance between the two measurements and obtain filtered security samples. The solution of MAPDE can constitute a new discriminant distribution function instead of the Wasserstein distance of WGAN.
- The technical scheme of the invention is an adversarial optimization method for the training process of generative adversarial network, specifically comprising the steps of:
- 1) sending an image data training set and random noise into a generator in the generative adversarial network, taking generated data output by the generator as attack samples, putting the generated data and real data of the image data into two data sets X and Y, inputting the two data sets into a discriminator D in the generator, calculating probability densities ρX of X and ρY of Y, and solving maximum likelihood estimation maximum values of the probability densities of the real data and the generated data; calculating the measurements of the real data and the generated data so as to solve a numerical solution of the elliptic Monge-Ampere partial differential equation, obtaining a best mapping between the real data distribution and the generated data distribution, training the generator by calculating a loss function of the generator, forming an attack network in the generator, and finally obtaining an optimal mapping U of the attack sample and the real data to complete the training of the attack network;
- 2) adding the discriminator D trained in the step 1) into a defence network in the generative adversarial network, sending the image data training set and the random noise into the generator in the generative adversarial network, taking the output data of the generator as input data of the defence network, training the defence network by a defence network loss function obtained by a solution of the Monge-Ampere PDE and an optimal transmission theory, and obtaining by the defence network a maximum distance between the two measurements by overcoming a loss function of the optimal mapping in the process of training the defence network; and finally obtaining an output value of the defence network by iterative training, namely filtered safety samples.
- The generator has a loss function of
-
L D =E x˜Px[D w(G(x))]−E y˜Py[D w(G(y))]+λEx (|∇xD(x)|2−1) - wherein x and y correspond to points within the sets X and Y; Ex˜Px is an expectation of the probability distribution of real data; Ey˜Py is an expectation of probability distribution of attack sample data; E
x is an expectation for Licpschiz continuous data; Dw is a discriminator network with weights; D is a discriminator network; G is a generator network; λ is a penalty coefficient, a hyper-parameter set by the training network, and E is an expectation; - the defence network has a loss function of
-
- wherein m is the number of discrete points in each dimension of the network.
- The invention has the beneficial effects that the invention generates an adversarial optimization method for the training process of generative adversarial network, an effective attacking method of GANs is successfully established, and several calculation operators are provided to prove that the precision is improved by 5.3%. In addition, MAGAN can be stably trained without adjusting hyper-parameters, and the convergence speed of the MAGAN is 317.2% faster than that of the WGAN-GP in the LSUN Bedrooms database. The IS value of MAGAN on a CIFAR-10 data set also reaches 8.7, which improves the accuracy of target classification and recognition system for unmanned vehicle.
-
FIG. 1 is a schematic diagram of a neural network classification segmentation plane; -
FIG. 2 is a schematic diagram of an attack sample cross-domain segmentation plane; -
FIG. 3 is a distribution diagram of a numerical initial solution of MAPDE; -
FIG. 4 is a schematic diagram of a unit inward normal vector of MAPDE under Neumann boundary conditions; -
FIG. 5 is a frame structure diagram of an improved generative adversarial network according to the present invention; -
FIG. 6 is a structure diagram for generative adversarial samples and defence networks according to the present invention; -
FIG. 7 is a flow chart for attacking and defending a neural network of a target classification and a recognition system for unmanned vehicle according to the present invention. - In order to assist and realize unmanned vehicle, the neural network implements an algorithm recognition part of an unmanned vehicle target classification and a recognition system for image recognition and classification. Convolution neural network (CNN) is mainly used in image processing and target recognition, and under-fitting exists in the space of real data and generated data. Therefore, the proposed attack and defence algorithm has been studied to enhance the robustness of neural networks, but the black-box attack and defence effect is not good.
- In order to solve the problem, a homeomorphism mapping from a region to itself is searched, meeting two conditions of preserving measurements and minimizing transmission cost. According to the preserving measurements, E c Q for all Boyle sets (Ω is a finite open set). The mapping T maps a probability distribution μ to a probability distribution v, denoted as T*μ=v. The transmission cost of optimal transmission mapping T: Ω→Ω is defined as:
-
I[s]=∫X c(x,s(x))dx (1), -
c(x,y)=½|x−y| 2 (2), - wherein I is a mapping cost function; c (x, y) is a distance for mapping x to y; and x, y belong to points within the set.
- In this case, Brenier proves that there is a convex function u: Ω→R (R is a set of all real numbers), whose gradient mapping T:x→∇u(x) is the only optimal transmission mapping. This convex function is called Brenier potential. The Jacobian equation gives that the Brenier potential satisfies the Monge-Ampere PDE (2), and the Jacobian matrix of the gradient mapping is the Hessian matrix of the Brenier potential.
-
- wherein D is a partial derivative; det is a determinant; ρ is a measurement; X and Y are complete sets of x and y, respectively.
- According to the Kantorovich theorem, WGAN transforms the original function into the dual problem. However, the constraints of the dual problem are difficult to implement, so sometimes WGAN still produces poor samples or even fail to converge. Therefore, the method of Kantorovich is not a good choice. This has prompted the need for more complex methods to efficiently compute the optimal mapping. Brenier's method is chosen to obtain the best mapping between the actual data distribution and the generation distribution. In the case where the transmission cost is the difference square, the problem can be transformed into solving the Monge-Ampere PDE. The Monge-Ampere PDE can avoid Lipschitz constraints completely. For the optimal transmission problem, the scheme of the Monge-Ampere PDE is more accurate than the method of Kantorovich (WGAN-GP).
- In the optimal transmission problem between convex sets, the transmission condition (BC) is also named the second boundary value problem or the second kind of boundary condition. The boundary conditions can be mapped into the set X by boundary nodes, and Y by the boundary nodes. In view of the gradient appearing in the transmission boundary condition, it is expected to find the Neumann boundary conditions:
-
- According to the equation (5), a boundary normal vector n is defined, wherein normal vector components corresponding to n dimensions in the boundary normal vector n are n1, n2, . . . , nn, where the vector n is perpendicular to the point x (x belongs to a point in the set X because the set X has a boundary, and an unspecified point x on the boundary is also in the set X Therefore, the equation 5 is the boundary conditions meeting the condition, the domain of definition is also X, and x is just the point in the met condition.), ∂X is the boundary of the set X, and δx solves a partial derivative with respect to x. ϕ is a normal vector calculation function, calculated by the equation (7). The normal vector component n corresponding to each dimension in the equation (7) will be multiplied by the partial derivative of u in the corresponding dimension, i.e., a difference value between one unit step by which the function u is shifted forward in each dimension and a center value.
- X is n-dimensional, so there are n subscripts for x, i.e., xi
1 , . . . ,in , i.e., i1, i2, i3, . . . in, n in total, representing different dimensions. Because the numerical solution needs to be discretized, there are m discrete points in each dimension, i.e., i={1, 2, 3, . . . , m}. The front and back points, i.e., i+1 and i, are needed to calculate the difference by the five-step method. Assuming that in dimension 1, the subscripts of the front and back points are i1, (i+1)1, i.e. xi1 , . . . ,in and x(1+1)1 , . . . ,in . Rule: the subscript x in the first term u takes backwards, in turn, a next unit step i+1 in dimension. - This updated boundary conditions are used to solve the Monge-Ampere PDE again to obtain a new numerical solution. uk+1 is a solution of the k+1 iteration. Because it is monotonic, the scheme relies only on values within a square. When the dimension of MAPDE is greater than 2, there are many nodes around the boundary that directly affect the value of the solution, as shown in
FIG. 3 . This will take more computational time, so more attention should be paid to the boundary, and upper limits of these monotonic methods should be used for all acceptable ranges to ensure that a highly accurate numerical solution is obtained. The monotonicity of the method is preserved. First, a boundary is set on the other side of the square area. Then, at the corners, the derivation directions in the other dimensions are limited in the form of tilt constraints. Also, limiting the allowed direction to a single quadrant ensures that the required information will continue to remain within the square area. Next, a new approximation is obtained in an inward direction. Finally, as shown inFIG. 4 , the above steps are repeated until a suitable boundary is obtained, which corresponds to considering all supporting hyperplanes at these points. - The present invention defines several finite difference operators that are used to approximate the first and second order partial derivatives using the central difference value. Standard discretization is performed to the equation by the center difference:
-
MA s h[u]=(D x1 x1 u)(D x2 x2 u)−(D x1 x2 u)2−f(x)/g(D x1 u,D x2 u) (8). - MA is a discrete Monge-Ampere PDE; D is a partial derivative; and f, g are measurements. U is a discrete numerical solution to the Monge-Ampere PDE.
- The finite difference operator is
-
- Along these directions v=(1/√{square root over (2)},1/√{square root over (2)}) and v⊥=(1/√{square root over (2)},−1/√{square root over (2)}), the finite difference operator is further computed:
-
- The result of discretization is the variational form of MA operator, and an additional term is added to the expression to further punish non-convexity:
-
- The partial derivatives of all nodes need to be calculated on the basis of Newton iteration. However, when the dimension is greater than 3, it is difficult to obtain all the partial derivatives in the high dimension. Although the solution of MAPDE can be discretized by setting a variable, the high dimension also makes it difficult for nodes in the grid to define the sequence of the nodes. It can be found that most relevant nodes in each dimension are its forward and backward nodes. The nodes with the central difference in different dimensions are the same node. Therefore, it is suggested to use the surrounding gradient instead of the overall gradients in order to accelerate the high-dimensional convergence.
- An iterative expression of MAPDE under Neumann boundary conditions and an initialization expression at the beginning of solving the equation can be obtained:
-
- MAPDE can be solved to obtain the best mapping u(x) between the actual data distribution and the generation distribution. Then, according to the following equation, a cost function for the OT problem can be obtained, which can train the generator how to efficiently generate real data. Equation 14 is maximum likelihood estimation for solving the real data and the generated data, and the maximum value of Equation 15 is solved by MLE to determine the distribution of the real data and the generated data.
-
- The best mapping u(x) is obtained by solving the Monge-Ampere PDE, as shown by a dashed rectangle in
FIG. 5 . - An image data training set and random noise are sent into a generator, and the generator outputs data as attack samples and real image data, correspondingly obtaining a brand-new discriminator D formed by entering two data sets X and Y into a dashed rectangle. The brand-new discriminator D distinguishes the measurement of the real data from the measurement of the generated data, and the generator G continuously confronts the discriminator D in the training process, thereby performing effective attacks. The best mapping u(x) is obtained by solving the Monge-Ampere PDE, and the generation of attack adversarial samples is achieved by the trained generator G in
FIG. 5 . That is, the interior of the generator constitutes the attack network. In the process of training the discriminator D shown inFIG. 5 , the upper bound of the loss function of the generator of the equation 16 will be solved; on the contrary, in the process of training the generator G shown inFIG. 5 , the lower bound of the loss function of the generator of the equation 16 will be solved, finally achieving a good adversarial effect. The loss function, as a new divergence, of the Monge-Kantorovich transmission problem replaces the Wasserstein distance of WGAN. - At the start of the training, only the real data of the image is used to obtain the probability density PX of X The probability density ρY of the attack sample Y causes the generation distribution Pg=the real data distribution Pr generated data, since the solution of MAPDE causes Pg to tend towards Pr. The appropriate boundary can then be calculated by the equation (13) using Pr and Pg according to the Neumann boundary conditions. Next, the equation set F[u]=0 is obtained by using the finite difference method (FDM) and solved by the Newton iteration method.
-
Algorithm 1 Monge-Ampere Adversarial Network Training Process Requirements: Adam optimizer hyper-parameters α, β1 , and β2; evaluation iteration number N of each generator iteration; a finite step size difference method h with a batch M; Output: a trained generator parameter θ; 1: while θ does not converge, do 2: for i = 1; i < N; i + + do 3: calculate the probability density ρX of X and the probability density ρY of Y 4: L({circumflex over (θ)}; y) = supθ∈Θ L(θ; y) 5: 6: 7: calculate the optimal u(x) between the real data distribution Pr and the generation distribution Pg 8: un + 1 = un − ∇F (un) − 1 F (un) 9: 10: use the cost function of the Monge-Kantorovich transmission problem to train the generator 11: I [s] = ∫X c (x, s (x))dx 12: end for 13: end while - Implementation steps include:
- Step 1, calculating the probability density ρX of X and the probability density ρY of Y
- Step 2, solving maximum likelihood estimation maximum values of the real data and the generated data;
- Step 3, calculating the measurements of the real data and the generated data;
- Step 4, calculating the first type of boundary of the MAPDE;
- Step 5, calculating the optimal u(x) between the real data distribution Pr and the generation distribution Pg;
- Step 6, iterating gradient values;
- Step 7, calculating the loss function;
- Repeat the steps until the cost function converges.
- This network is applied in an adversarial sample generation network by MAGAN. As shown in
FIG. 5 , a black-box attack and a white-box attack are realized more efficiently by good robustness of the MAGAN, forming an attack network. In order to better defend against attacks from the adversarial samples, the defence network is trained by the attack samples generated by the generator G, and the robustness is enhanced by the solution of the Monge-Ampere PDE, as shown inFIG. 6 . The discriminator D inFIG. 6 corresponds to the structure in a dashed box inFIG. 5 , which contains a numerical solution of the partial differential equation. When the discriminator is trained and applied to a lower discriminator network ofFIG. 6 , the filtered samples are used for training the defence network. The samples of the generator G are taken as the input data of the defence network, and since the input training data has strong robustness of black-box and white-box attacks, the cross entropy is calculated by logistic regression on the input data by means of the loss function of the defence network, the real data and the artificial data are distinguished, achieving an isolation effect and a good defence effect. - The loss function for the generator is
-
L D =E x˜Px[D w(G(x))]−E y˜Py[D w(G(y))]+λEx (|∇xD(x)|2−1)2 (16), - wherein x and y correspond to points within the sets X and Y; Ex˜Px is an expectation of the probability distribution of real data; Ey˜Py is an expectation of probability distribution of attack sample data; E
x is an expectation for Licpschiz continuous data; Dw is a discriminator network with weights; D is a discriminator network; G is a generator network; λ is a penalty coefficient, a hyper-parameter set by the training network, and E is an expectation; - the defence network has a loss function of
-
- When the attack network in
FIG. 5 is trained, the defence network inFIG. 6 is implemented by adding an attacked target network, as shown inFIG. 7 . In the right box ofFIG. 7 , by setting the loss function of the generator, the generator G can learn the measurements of the attack samples to calculate the optimal mapping U. Then the generator can obtain attack samples of the target network according to the input data and the random noise so as to carry out effective attacks. The target network is a network required to be attacked, i.e. a known network trained by others. In the left box ofFIG. 7 , the output data of the generator G is taken as the input data of the defence network, and the defence network is trained according to the loss function (equation 17) obtained from the solution of the Monge-Ampere PDE and the optimal transmission theory in the right box. The equation 16 is a generator network. When the defence network is trained, the loss function of the generator does not participate in the training. In the process of training the defence network, by overcoming the loss function of the optimal mapping, the defence network can obtain a maximum distance between the two measurements. Finally, the output values of the defence network can be obtained by iterative training, and the target classification and recognition system for unmanned vehicle can be well helped to judge the attack samples.
Claims (2)
L D =E x˜Px[D w(G(x))]−E y˜Py[D w(G(y))]+λE
Applications Claiming Priority (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010113638.1 | 2020-02-04 | ||
CN202010113638.1A CN111401138B (en) | 2020-02-24 | 2020-02-24 | Countermeasure optimization method for generating countermeasure neural network training process |
PCT/CN2020/118698 WO2021169292A1 (en) | 2020-02-24 | 2020-09-29 | Adversarial optimization method for training process of generative adversarial neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
US20220122348A1 true US20220122348A1 (en) | 2022-04-21 |
US11315343B1 US11315343B1 (en) | 2022-04-26 |
Family
ID=71430417
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US17/288,566 Active US11315343B1 (en) | 2020-02-24 | 2020-09-29 | Adversarial optimization method for training process of generative adversarial network |
Country Status (3)
Country | Link |
---|---|
US (1) | US11315343B1 (en) |
CN (1) | CN111401138B (en) |
WO (1) | WO2021169292A1 (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20220277187A1 (en) * | 2021-03-01 | 2022-09-01 | Robert Bosch Gmbh | Concept-based adversarial generation method with steerable and diverse semantics |
CN115047721A (en) * | 2022-05-31 | 2022-09-13 | 广东工业大学 | Method for rapidly calculating mask near field by using cyclic consistency countermeasure network |
CN115276766A (en) * | 2022-07-19 | 2022-11-01 | 西安电子科技大学 | Cooperative Unmanned Aerial Vehicle (UAV) assisted interference power and trajectory joint optimization method |
CN115984792A (en) * | 2022-09-30 | 2023-04-18 | 北京瑞莱智慧科技有限公司 | Countermeasure test method, system and storage medium |
CN116071797A (en) * | 2022-12-29 | 2023-05-05 | 北华航天工业学院 | Sparse face comparison countermeasure sample generation method based on self-encoder |
CN116545767A (en) * | 2023-06-27 | 2023-08-04 | 北京天云海数技术有限公司 | Automatic XSS attack load generation method and system based on generation countermeasure network |
Families Citing this family (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111401138B (en) * | 2020-02-24 | 2023-11-07 | 上海理工大学 | Countermeasure optimization method for generating countermeasure neural network training process |
CN112801297B (en) * | 2021-01-20 | 2021-11-16 | 哈尔滨工业大学 | Machine learning model adversity sample generation method based on conditional variation self-encoder |
CN112989359B (en) * | 2021-03-17 | 2022-06-14 | 华南理工大学 | Backdoor attack method for pedestrian re-identification model based on triple loss |
CN113395653A (en) * | 2021-06-08 | 2021-09-14 | 南京工业大学 | Fingerprint positioning Radio Map expansion method based on DC-CGAN |
CN113947704A (en) * | 2021-10-09 | 2022-01-18 | 北京建筑大学 | Confrontation sample defense system and method based on attention ranking |
CN114219778B (en) * | 2021-12-07 | 2024-04-02 | 北京工业大学 | Data depth enhancement method based on WGAN-GP data generation and poisson fusion |
CN114154250B (en) * | 2021-12-20 | 2024-06-14 | 北京航空航天大学 | DCCGAN-based hypersonic aircraft flow thermosetting coupling physical field solving method |
CN114117333B (en) * | 2022-01-20 | 2022-05-17 | 南湖实验室 | Countermeasure reconstruction network design, training method and detection method for anomaly detection |
CN114969785B (en) * | 2022-05-27 | 2024-06-18 | 哈尔滨工业大学(深圳) | Carrier-free image steganography method based on reversible neural network |
CN115064250A (en) * | 2022-06-06 | 2022-09-16 | 大连理工大学 | Method for adjusting distribution of stay in hospital and related product |
CN114708974A (en) * | 2022-06-06 | 2022-07-05 | 首都医科大学附属北京友谊医院 | Method for predicting hospitalization duration of new coronary pneumonia patient and related product |
US11895344B1 (en) | 2022-12-09 | 2024-02-06 | International Business Machines Corporation | Distribution of media content enhancement with generative adversarial network migration |
CN117012204B (en) * | 2023-07-25 | 2024-04-09 | 贵州师范大学 | Defensive method for countermeasure sample of speaker recognition system |
CN117669651B (en) * | 2024-01-31 | 2024-05-14 | 山东省计算中心(国家超级计算济南中心) | ARMA model-based method and ARMA model-based system for defending against sample black box attack |
Family Cites Families (41)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110059793B (en) | 2017-10-26 | 2024-01-26 | 辉达公司 | Gradual modification of a generative antagonistic neural network |
US20190147343A1 (en) * | 2017-11-15 | 2019-05-16 | International Business Machines Corporation | Unsupervised anomaly detection using generative adversarial networks |
US11137761B2 (en) * | 2017-11-20 | 2021-10-05 | At&T Intellectual Property I, L.P. | Object modeling with adversarial learning |
US10592779B2 (en) * | 2017-12-21 | 2020-03-17 | International Business Machines Corporation | Generative adversarial network medical image generation for training of a classifier |
KR102176335B1 (en) * | 2018-02-07 | 2020-11-10 | 어플라이드 머티리얼즈 이스라엘 리미티드 | Method and system for generating a training set usable for testing semiconductor specimens |
US10970765B2 (en) * | 2018-02-15 | 2021-04-06 | Adobe Inc. | Generating user-customized items using a visually-aware image generation network |
US11069030B2 (en) * | 2018-03-22 | 2021-07-20 | Adobe, Inc. | Aesthetics-guided image enhancement |
US10810754B2 (en) * | 2018-04-24 | 2020-10-20 | Ford Global Technologies, Llc | Simultaneous localization and mapping constraints in generative adversarial networks for monocular depth estimation |
CN108711138B (en) * | 2018-06-06 | 2022-02-11 | 北京印刷学院 | Gray level picture colorizing method based on generation countermeasure network |
GB201809604D0 (en) * | 2018-06-12 | 2018-07-25 | Tom Tom Global Content B V | Generative adversarial networks for image segmentation |
CN109389166A (en) * | 2018-09-29 | 2019-02-26 | 聚时科技(上海)有限公司 | The depth migration insertion cluster machine learning method saved based on partial structurtes |
CN109584178A (en) * | 2018-11-29 | 2019-04-05 | 腾讯科技(深圳)有限公司 | Image repair method, device and storage medium |
US11087170B2 (en) * | 2018-12-03 | 2021-08-10 | Advanced Micro Devices, Inc. | Deliberate conditional poison training for generative models |
KR20200075344A (en) * | 2018-12-18 | 2020-06-26 | 삼성전자주식회사 | Detector, method of object detection, learning apparatus, and learning method for domain transformation |
KR20200093910A (en) * | 2019-01-29 | 2020-08-06 | 삼성전자주식회사 | Method for providing data assocatied with original data, electronic device and storage medium therefor |
JP7268367B2 (en) * | 2019-01-30 | 2023-05-08 | 富士通株式会社 | LEARNING DEVICE, LEARNING METHOD AND LEARNING PROGRAM |
US11024013B2 (en) * | 2019-03-08 | 2021-06-01 | International Business Machines Corporation | Neural network based enhancement of intensity images |
CN109919251B (en) * | 2019-03-21 | 2024-08-09 | 腾讯科技(深圳)有限公司 | Image-based target detection method, model training method and device |
US11120526B1 (en) * | 2019-04-05 | 2021-09-14 | Snap Inc. | Deep feature generative adversarial neural networks |
KR20200132665A (en) * | 2019-05-17 | 2020-11-25 | 삼성전자주식회사 | Attention layer included generator based prediction image generating apparatus and controlling method thereof |
CN110222628A (en) * | 2019-06-03 | 2019-09-10 | 电子科技大学 | A kind of face restorative procedure based on production confrontation network |
JP7016835B2 (en) * | 2019-06-06 | 2022-02-07 | キヤノン株式会社 | Image processing method, image processing device, image processing system, learned weight manufacturing method, and program |
KR20200142374A (en) * | 2019-06-12 | 2020-12-22 | 삼성전자주식회사 | Method for selecting artificial intelligience model based on input data and disaply apparatus for performing the same method thereof |
US11068753B2 (en) * | 2019-06-13 | 2021-07-20 | Visa International Service Association | Method, system, and computer program product for generating new items compatible with given items |
US11373093B2 (en) * | 2019-06-26 | 2022-06-28 | International Business Machines Corporation | Detecting and purifying adversarial inputs in deep learning computing systems |
US10496809B1 (en) * | 2019-07-09 | 2019-12-03 | Capital One Services, Llc | Generating a challenge-response for authentication using relations among objects |
WO2021025217A1 (en) * | 2019-08-08 | 2021-02-11 | 엘지전자 주식회사 | Artificial intelligence server |
US20210049452A1 (en) * | 2019-08-15 | 2021-02-18 | Intuit Inc. | Convolutional recurrent generative adversarial network for anomaly detection |
EP3798917A1 (en) * | 2019-09-24 | 2021-03-31 | Naver Corporation | Generative adversarial network (gan) for generating images |
KR20190119548A (en) * | 2019-10-02 | 2019-10-22 | 엘지전자 주식회사 | Method and apparatus for processing image noise |
US11232328B2 (en) * | 2020-01-31 | 2022-01-25 | Element Ai Inc. | Method of and system for joint data augmentation and classification learning |
CN111401138B (en) | 2020-02-24 | 2023-11-07 | 上海理工大学 | Countermeasure optimization method for generating countermeasure neural network training process |
US11961219B2 (en) * | 2020-02-27 | 2024-04-16 | KLA Corp. | Generative adversarial networks (GANs) for simulating specimen images |
KR20210136706A (en) * | 2020-05-08 | 2021-11-17 | 삼성전자주식회사 | Electronic apparatus and method for controlling thereof |
US11651292B2 (en) * | 2020-06-03 | 2023-05-16 | Huawei Technologies Co., Ltd. | Methods and apparatuses for defense against adversarial attacks on federated learning systems |
US20210383241A1 (en) * | 2020-06-05 | 2021-12-09 | Nvidia Corporation | Training neural networks with limited data using invertible augmentation operators |
US20210397198A1 (en) * | 2020-06-18 | 2021-12-23 | Ford Global Technologies, Llc | Enhanced vehicle operation |
JP7419178B2 (en) * | 2020-07-01 | 2024-01-22 | 株式会社東芝 | Learning devices, methods and programs |
US20220027490A1 (en) * | 2020-07-24 | 2022-01-27 | Siemens Aktiengesellschaft | Gan-based data obfuscation decider |
US12056880B2 (en) * | 2020-08-03 | 2024-08-06 | Korea Advanced Institute Of Science And Technology | Method of classifying lesion of chest x-ray radiograph based on data normalization and local patch and apparatus thereof |
US11328410B2 (en) * | 2020-08-03 | 2022-05-10 | KLA Corp. | Deep generative models for optical or other mode selection |
-
2020
- 2020-02-24 CN CN202010113638.1A patent/CN111401138B/en active Active
- 2020-09-29 WO PCT/CN2020/118698 patent/WO2021169292A1/en active Application Filing
- 2020-09-29 US US17/288,566 patent/US11315343B1/en active Active
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20220277187A1 (en) * | 2021-03-01 | 2022-09-01 | Robert Bosch Gmbh | Concept-based adversarial generation method with steerable and diverse semantics |
US11763135B2 (en) * | 2021-03-01 | 2023-09-19 | Robert Bosch Gmbh | Concept-based adversarial generation method with steerable and diverse semantics |
CN115047721A (en) * | 2022-05-31 | 2022-09-13 | 广东工业大学 | Method for rapidly calculating mask near field by using cyclic consistency countermeasure network |
CN115276766A (en) * | 2022-07-19 | 2022-11-01 | 西安电子科技大学 | Cooperative Unmanned Aerial Vehicle (UAV) assisted interference power and trajectory joint optimization method |
CN115984792A (en) * | 2022-09-30 | 2023-04-18 | 北京瑞莱智慧科技有限公司 | Countermeasure test method, system and storage medium |
CN116071797A (en) * | 2022-12-29 | 2023-05-05 | 北华航天工业学院 | Sparse face comparison countermeasure sample generation method based on self-encoder |
CN116545767A (en) * | 2023-06-27 | 2023-08-04 | 北京天云海数技术有限公司 | Automatic XSS attack load generation method and system based on generation countermeasure network |
Also Published As
Publication number | Publication date |
---|---|
CN111401138A (en) | 2020-07-10 |
WO2021169292A1 (en) | 2021-09-02 |
US11315343B1 (en) | 2022-04-26 |
CN111401138B (en) | 2023-11-07 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US11315343B1 (en) | Adversarial optimization method for training process of generative adversarial network | |
CN112364885B (en) | Confrontation sample defense method based on interpretability of deep neural network model | |
US11783198B2 (en) | Estimating the implicit likelihoods of generative adversarial networks | |
Balduzzi et al. | Neural taylor approximations: Convergence and exploration in rectifier networks | |
CN109165735A (en) | Based on the method for generating confrontation network and adaptive ratio generation new samples | |
CN114239685B (en) | Method and device for evaluating robustness of neural network image classification model | |
CN112580728B (en) | Dynamic link prediction model robustness enhancement method based on reinforcement learning | |
CN111291810B (en) | Information processing model generation method based on target attribute decoupling and related equipment | |
CN113935396A (en) | Manifold theory-based method and related device for resisting sample attack | |
WO2019234156A1 (en) | Training spectral inference neural networks using bilevel optimization | |
CN112686249B (en) | Grad-CAM attack method based on anti-patch | |
Kang et al. | Interpretability for reliable, efficient, and self-cognitive DNNs: From theories to applications | |
CN111950635B (en) | Robust feature learning method based on layered feature alignment | |
CN113935496A (en) | Robustness improvement defense method for integrated model | |
Lin et al. | Fluid flow mass transport for generative networks | |
Moreels et al. | Common-frame model for object recognition | |
Bénézet et al. | Learning conditional distributions on continuous spaces | |
CN112800871B (en) | Automatic driving image recognition method based on attention mechanism and relation network | |
CN115510986A (en) | Countermeasure sample generation method based on AdvGAN | |
Millan et al. | Explaining regression based neural network model | |
Maeda et al. | Neural network maximizing ordinally supervised multi-view canonical correlation for deterioration level estimation | |
WO2022236647A1 (en) | Methods, devices, and computer readable media for training a keypoint estimation network using cgan-based data augmentation | |
WO2022077345A1 (en) | Method and apparatus for neural network based on energy-based latent variable models | |
WO2022167077A1 (en) | A normalization scheme for self-attention neural networks | |
Sornam et al. | An efficient morlet function based adaptive method for faster backpropagation for handwritten character recognition |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AS | Assignment |
Owner name: YUNWU NETLINK (SUZHOU) INTELLIGENT TECHNOLOGY CO., LTD., CHINA Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:PEI, SONGWEN;SHEN, TIANMA;REEL/FRAME:056030/0876 Effective date: 20210325 Owner name: UNIVERSITY OF SHANGHAI FOR SCIENCE AND TECHNOLOGY, CHINA Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:PEI, SONGWEN;SHEN, TIANMA;REEL/FRAME:056030/0876 Effective date: 20210325 |
|
FEPP | Fee payment procedure |
Free format text: ENTITY STATUS SET TO UNDISCOUNTED (ORIGINAL EVENT CODE: BIG.); ENTITY STATUS OF PATENT OWNER: SMALL ENTITY |
|
FEPP | Fee payment procedure |
Free format text: ENTITY STATUS SET TO SMALL (ORIGINAL EVENT CODE: SMAL); ENTITY STATUS OF PATENT OWNER: SMALL ENTITY |
|
STCF | Information on status: patent grant |
Free format text: PATENTED CASE |