Network congestion link diagnosis method based on adversarial auto-encoding

Abstract

The application relates to a network congestion link diagnosis method based on an adversarial automatic encoder and belongs to the technical field of network communication, which comprises the following steps: generating a data set containing link states and path states according to the congestion prior probability of a link; feeding the data set into an automatic encoder, superimposing a generative adversarial network on the encoder or the decoder of the automatic encoder, inputting the path states or the link states into the encoder, and reconstructing the link states or the path states; training the generative adversarial network, enabling the generative adversarial network to learn the congestion prior probability information of the link, and outputting a link state solution with the maximum congestion posterior probability; and reversely propagating the reconstruction loss and the generative adversarial loss, and iteratively updating network parameters. The application applies the adversarial automatic encoder, learns the unique mapping relationship from the link to the path through the link states and the path states, does not need to measure the network topology, and meanwhile, the adversarial automatic encoder is data-driven, has strong adaptability and has certain attack resistance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of network communication, in particular to a network congestion link diagnosis method based on adversarial auto-encoding. BACKGROUND

[0002] Network congestion diagnosis methods are divided into two types, direct measurement and indirect measurement. Direct measurement, commonly known as using Simple Network Management Protocol (SNMP) to monitor network performance data, the polling mechanism of SNMP takes minutes to collect data when facing large and complex networks, which in turn contributes to network communication congestion. Directly measuring all nodes in the network not only takes time and causes network conditions, but also faces the problem of not willing to cooperate and share sensitive information such as link-level performance parameters between nodes due to the development of the Internet towards distribution, heterogeneity and privacy.

[0003] Network tomography in indirect measurement breaks these limitations, does not require all nodes to follow a special network protocol, does not require all nodes to install measurement devices, and does not require nodes to cooperate with each other. Network tomography regards all nodes inside the target network as black boxes, only needs to obtain end-to-end measurement results, and inversely infers the characteristics inside the network, such as link-level parameters: link delay, link packet loss rate and link utilization, to locate the congestion link in the network.

[0004] A linear relationship can be established between the link packet loss rate and the path packet loss rate, and the link packet loss rate can be quantitatively estimated to identify the congestion link, but the measurement difficulty of quantitative estimation is large, the error tolerance is small, and the link state and path state also have a linear relationship in Boolean algebra, so only the qualitative estimation of the link state as congestion or good is needed, therefore qualitative estimation is commonly used at home and abroad.

[0005] There are multiple solutions for congestion link diagnosis in network tomography. Given the path state and topology structure, there is only one link state solution when all columns in the routing matrix are linearly independent, i.e. a link forms a path, which obviously does not conform to the actual network. The actual situation is that the number of paths is less than the number of links. Therefore, the estimation problem has multiple link state solutions.

[0006] Most of the current network tomography-based congestion link diagnosis methods, such as Smallest Common Failure Set (SCFS), Congested Link Identification (CLINK) and Maximum A Posteriori Estimate (MAP) algorithm, need to measure the network topology in advance, and are empirical models with high abstraction and neglect many details such as nonlinear factors, and have general adaptability. In addition, the premise of these algorithms is a safe network environment, but the Internet has security vulnerabilities and network attacks emerge in an endless stream, which requires the algorithm to have certain anti-attack ability. However, the performance of these algorithms is severely degraded when facing network attacks or data noise. SUMMARY

[0007] In order to solve the problems of the traditional congestion link diagnosis method that needs to measure the network topology in advance and the performance is severely degraded when facing network attacks, the present application provides a network congestion link diagnosis method based on adversarial auto-encoder. The method applies an adversarial auto-encoder to learn the unique mapping relationship from link state to path state without measuring the network topology, and the adversarial auto-encoder is data-driven, has strong adaptability and certain anti-attack ability.

[0008] To solve the above problems, the present application adopts the following technical scheme:

[0009] A network congestion link diagnosis method based on adversarial auto-encoder, the method comprising the following steps:

[0010] Step S1: generating a data set containing link state and path state according to the congestion prior probability of the link;

[0011] Step S2: inputting the data set into an auto-encoder, in a supervised scheme, superimposing a generative adversarial network on the decoder of the auto-encoder, taking the decoder as the generator of the generative adversarial network, inputting the link state into the encoder, and the encoder outputs the predicted path state, the predicted path state is input into the decoder, and the decoder outputs the reconstructed link state; in an unsupervised scheme, superimposing a generative adversarial network on the encoder of the auto-encoder, taking the encoder as the generator of the generative adversarial network, inputting the path state into the encoder, and the encoder outputs the predicted link state, and the path state is directly reconstructed according to the predicted link state;

[0012] Step S3: training the generative adversarial network under the supervised scheme and the unsupervised scheme respectively, so that the generative adversarial network learns the congestion prior probability information of the link, and outputs the link state solution with the maximum congestion posterior probability;

[0013] Step S4: iteratively updating network parameters by backpropagation reconstruction loss and generative adversarial loss.

[0014] Compared with the prior art, the present application has the following beneficial effects:

[0015] The present application provides a network congestion link diagnosis method based on adversarial auto-encoding, which can be applied in the scenario of non-cooperative internal nodes of the network, including supervised and unsupervised schemes, both of which infer the location of the congested link through the congestion prior probability of the link. The auto-encoder architecture can learn the unique mapping relationship from the link state to the path state, avoiding the measurement of the network topology, saving diagnosis time, and at the same time, the adversarial auto-encoder relies on data-driven, has strong adaptability, and has stronger attack resistance than traditional algorithms. The generative adversarial network is responsible for learning the congestion prior probability of the link, so that the link state with the maximum output congestion posterior probability is obtained. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The flowchart of the network congestion link diagnosis method based on adversarial auto-encoding described in the embodiments of the present application;

[0017] Figure 2 The adversarial auto-encoding network structure diagram under the unsupervised scheme;

[0018] Figure 3 The adversarial auto-encoding network structure diagram under the supervised scheme;

[0019] Figure 4 The network topology diagram for experiment;

[0020] Figure 5 The performance evaluation index curve result diagram of various algorithms under different attack frequencies. DETAILED DESCRIPTION

[0021] Adversarial auto-encoder (AAE) is one of the unsupervised model methods of deep learning, which can learn the data features with representation constraints using unlabeled data, has good adaptive ability, and has been verified and applied in semi-supervised classification, unsupervised clustering, image and style disentanglement. Adversarial auto-encoder AAE can be regarded as a structure combining auto-encoder (AE) and generative adversarial network (GAN).

[0022] The autoencoder mainly consists of an encoder and a decoder. The encoder learns the implicit features of the input data, and the decoder reconstructs the original input data based on the implicit features. The function of the autoencoder is to extract features and reduce reconstruction error. The generative adversarial network has at least two modules: a generator G (Generative Model) and a discriminator D (Discriminative Model). The generator and the discriminator learn from each other to produce good outputs. The generator tries to generate real samples to deceive the discriminator, and the discriminator tries to distinguish the generated samples from the real samples. The generator and the discriminator are trained alternately to obtain optimization.

[0023] The application applies the adversarial autoencoder to the congestion link diagnosis, wherein the autoencoder is responsible for establishing the mapping relationship between the link and the path, and can learn the unique mapping relationship from the link to the path through the link state and the path state without the network topology. The generative adversarial network is responsible for learning the congestion prior probability of the link, so as to make the link state with the maximum output congestion posterior probability. In order to make the purpose, technical scheme and advantages of the application more clear and explicit, the embodiments of the application will be further described below with reference to the drawings.

[0024] Figure 1 The flowchart of the network congestion link diagnosis method based on the adversarial autoencoder proposed in the embodiment is shown. The network congestion link diagnosis method of the embodiment includes the following steps:

[0025] Step S1: generating a data set containing link states and path states according to the congestion prior probability of the link.

[0026] Specifically, the link state generation process is as follows:

[0027] Step S11: generating random numbers consistent with the number of links and randomly distributed in the [0, 1] interval. If the random number is less than the prior probability of the link, it is judged as good and its value is 0. Otherwise, if the random number is greater than the prior probability of the link, it is judged as congested and its value is 1. Thus, the link state X is obtained.

[0028] The path state generation process is as follows:

[0029] Step S12: combining the link state X and the topology structure A, and obtaining the path state Y according to the following formula: The calculation formula is as follows:

[0030]

[0031] Wherein, n p is the number of paths, n c is the number of links, A ik is the network routing matrix, and the rows in the matrix correspond to the paths and the columns correspond to the links.

[0032] Link state and Path state is a variable composed of multiple binary tags 0 and 1, 0 represents a good state, and 1 represents a congestion state. The link state and the path state are one-to-one correspondence encapsulated into a dictionary and stored in a csv table to form a data set, which is convenient for subsequent calling.

[0033] Step S2: The data set is sent into the autoencoder, and the mapping relationship between the path state and the link state is constructed by the autoencoder.

[0034] In the unsupervised scheme, as shown in Figure 2 , a generative adversarial network is superimposed on the encoder of the autoencoder, and the encoder is used as the generator of the generative adversarial network. The path state is input into the encoder of the autoencoder, and the encoder maps the path state Y to a hidden layer, which refers to the predicted link state X'. The predicted link state X' is not passed through the decoder, but is directly reconstructed into the path state Y' according to the operation formula .

[0035] In the supervised scheme, as shown in Figure 3 , a generative adversarial network is superimposed on the decoder of the autoencoder, and the decoder is used as the generator of the generative adversarial network. The link state is input into the encoder of the autoencoder, and the encoder maps the link state X to a hidden layer, which refers to the predicted path state Y'. The known link state and path state train the encoder so that the link state X can be completely correctly mapped to the corresponding path state Y'. The encoder outputs the predicted path state Y', and the predicted path state Y' is input into the decoder as input, and the link state X' is obtained by reconstruction through the decoder.

[0036] Step S3: The generative adversarial network under the supervised scheme and the unsupervised scheme is trained respectively, so that the generative adversarial network learns the congestion prior probability information of the link, and outputs the link state solution with the maximum congestion posterior probability.

[0037] Still referring to Figure 2 , the generative adversarial network under the unsupervised scheme is trained, including the following steps:

[0038] (1) According to the path state Y input into the encoder, a plurality of link state solutions X1…X n are generated, and the congestion prior probability p k of the link is combined, so that each link state solution is calculated by The congestion posterior probability of the link

[0039] (2) For the predicted link state X' output by the generator, the predicted congestion posterior probability P X' ;

[0040] (3) In the congestion posterior probability P of the link, P X is sampled as a positive sample into the discriminator, and the predicted congestion posterior probability P X' is input into the discriminator as a negative sample;

[0041] (4) The generator outputs negative samples as close to the positive samples as possible, the discriminator discriminates the positive samples as true and the negative samples as false as possible, the generator and the discriminator game with each other to reach Nash equilibrium, thereby inducing the generative adversarial network to learn the congestion prior probability information of the link and output the link state solution with the maximum congestion posterior probability.

[0042] Still referring to Figure 3 , when training the generative adversarial network under the supervised scheme, the following steps are included:

[0043] (1) A plurality of link state solutions X1…X n are generated according to the predicted path state Y' output by the encoder, and the congestion prior probability p k of the link is combined, and the congestion posterior probability P of each link state solution is calculated by the formula

[0044] (2) For the predicted link state X' output by the generator, the predicted congestion posterior probability P X' is calculated by the formula in step (1);

[0045] (3) In the congestion posterior probability P of the link, P X is sampled as a positive sample into the discriminator, and the predicted congestion posterior probability P X' is input into the discriminator as a negative sample;

[0046] (4) The generator outputs negative samples as close to the positive samples as possible, the discriminator discriminates the positive samples as true and the negative samples as false as possible, the generator and the discriminator game with each other to reach Nash equilibrium, thereby inducing the generative adversarial network to learn the congestion prior probability information of the link and output the link state solution with the maximum congestion posterior probability.

[0047] Step S4: Backpropagation of reconstruction loss and generative adversarial loss, and iterative update of network parameters.

[0048] The reconstruction loss and the generative adversarial loss in this step are both represented by a binary cross entropy loss (BCELoss), the reconstruction loss ensures that the predicted link state is the correct solution of the input path state, and the generative adversarial loss ensures that the predicted link state is the solution with the maximum congestion posterior probability of the input path state. The reconstruction loss and the generative adversarial loss are back propagated, and the network parameters are optimized.

[0049] Further, the network congestion link diagnosis method based on the adversarial automatic coding of the embodiment further includes step S5: selecting a performance evaluation index and an experimental network topology, and verifying the performance of the adversarial automatic coding network after updating the network parameters on the test set.

[0050] In step S5, the detection rate (DR), the false positive rate (FPR), and the balanced average F1 (Balanced Score) are taken as evaluation indexes. The formula calculation of the three indexes is as follows: Wherein, TP refers to an actual congestion link predicted as congestion, FP refers to an actual good link predicted as congestion, FN refers to an actual congestion link predicted as good, and TN refers to an actual good link predicted as good. The DR in the application refers to the proportion of the actual congestion links judged as congestion (i.e. correct judgment), the FPR refers to the proportion of the actual good links judged as congestion (i.e. incorrect judgment), and the F1 refers to the balanced index, which is the error weight of balancing the missed detection and the false detection.

[0051] The experimental network topology adopted is two kinds of topological designs, as shown in Figure 4 , wherein Figure 4 (a) is a 2x3 network topology, Figure 4 (b) is a 3x5 network topology. When evaluating the performance of the AAE algorithm, the large topology, i.e. the 3x5 network topology, is adopted. When evaluating the attack resistance of the AAE model, the small topology, i.e. the 2x3 network topology, is adopted. The 2x3 network topology contains two paths y1, y2 and three links x1, x2, x3, and there are four path state combinations and eight link state combinations. The congestion prior probability of the link x1 is 0.12, the congestion prior probability of the link x2 is 0.22, and the congestion prior probability of the link x3 is 0.32. The 3x5 network topology contains three paths y1, y2, y3 and five links x1, x2, x3, x4, x5, and there are eight path state combinations and thirty-two link state combinations. The congestion prior probability of the link x1 is 0.10, the congestion prior probability of the link x2 is 0.20, the congestion prior probability of the link x3 is 0.30, the congestion prior probability of the link x4 is 0.40, and the congestion prior probability of the link x5 is 0.50.

[0052] The data from the test set is fed into both unsupervised and supervised schemes in the adversarial autoencoder to complete autoencoder training and generative adversarial training. The unsupervised scheme is as follows: Figure 2 As shown, the path state Y is input to the encoder, and the output is the predicted link state X'. A decoder is not needed because there is a unique mapping between the link state and the path state; the path state Y' can be directly reconstructed using the formula Y' = A·X'. If matrix multiplication is used, it is important to ensure that any element in Y' greater than 1 is set to 1, or to perform a logical OR operation. At this point, a reconstruction loss exists between Y and Y'. The binary cross-entropy loss function BCELoss is chosen to describe this loss. BCELoss is a loss function for binary classification problems; each path can be viewed as a binary classification problem of congestion or good performance, and multiple paths are the superposition of multiple binary classification problems. Here, the goal of BCELoss is to reduce the reconstruction error, making the reconstructed path state Y' approximate the input path state Y, indirectly ensuring that the predicted link state X' is definitely a solution to the input path state Y. The prior probability p of link congestion is known. k And the predicted link state X', according to the formula Obtain the posterior probability P of congestion for the predicted link. X' There are many link X solutions that can cause path state Y, therefore there are multiple posterior congestion probabilities. Can a true P be obtained by sampling from itself as the probability? X Sampling P X and predicted P X' As positive and negative samples are input into the discriminator, a single output value is obtained. The discriminator's loss consists of two parts, half of which is the positive sample P. X The BCELoss loss between the output value of the input discriminator and real(1) is the other half of the negative sample P. X' The BCELoss loss between the input discriminator's output value and fake(0) means that the discriminator hopes to distinguish positive samples P. X The result is true (1), and the negative sample P is true. X' The result is judged as false (0), while the generator of a Generative Adversarial Network (GAN) aims to generate a sample P that is close to a positive sample. X negative sample P X' The generator and discriminator work against each other, leading to optimization. Here, the generator in a GAN is essentially the encoder in an AAE (Advanced Image Processing Unit), so the encoder part should include not only the reconstruction loss but also the generator loss, i.e., P. X'The loss of the value input into the discriminator and real(1) and the loss of the generator ensure that the predicted link state is the maximum congestion posterior probability solution of the input path state. Because the reconstruction loss ensures that the predicted link state is the correct solution, which is the premise and basis, the loss of the generator ensures that the predicted link state is the solution with the maximum congestion posterior probability, so the reconstruction loss accounts for a large proportion, and the loss of the generator accounts for a small proportion.

[0053] The encoder and the discriminator can be encapsulated by the Adam (Adaptive Moment Estimation) optimizer as G and D, and the loss function of G and D is described as:

[0054] G loss = 0.1·BCELoss(D(P X' ), real) + 0.9·BCELoss(Y', Y)

[0055] D loss = 0.5·BCELoss(D(P X ), real) + 0.5·BCELoss(D(P X' ), fake)

[0056] Wherein, D(P X' ) represents the output value of P X' input into the discriminator, and D(P X ) represents the output value of P X input into the discriminator.

[0057] The loss is back-propagated, and the gradient is updated for each small batch of data. The gradient is optimized once in each round. When the loss hovers around a certain value, the AAE training is completed.

[0058] The supervised diagnosis scheme based on the adversarial autoencoder AAE is as follows: Figure 3 Compared with the unsupervised scheme, the network topology is not used, and the topology measurement time is saved. The self-encoder builds the mapping relationship of the link and the path: the measured link state X is input into the encoder, and the predicted path state Y' is output. The predicted path state Y' is input into the decoder to output the reconstructed link state X'. Under the supervision, the known link state X and path state Y can train the encoder with unique X→Y mapping. At this time, the decoder board is stacked with the above similar generative adversarial network, so that the predicted path state Y' can output the link state X' with the maximum congestion posterior probability.

[0059] After the training, the performance evaluation experiment of AAE algorithm is carried out to evaluate the performance of AAE method. The AAE unsupervised scheme, AAE supervised scheme, maximum a posteriori estimation (MAP) algorithm, congestion link positioning algorithm (CLINK) and smallest consistent failure set (SCFS) algorithm are applied to the test set to obtain their DR, FPR and F1 values respectively. As shown in Table 1, the unsupervised scheme of AAE has the same performance indicators as MAP and CLINK, and according to the comprehensive indicator F1, the 0.846 of the unsupervised scheme of AAE is obviously better than the 0.777 of SCFS algorithm, which realizes the best diagnostic performance in unsupervised mode. In addition, by checking the link state prediction of AAE, it is found that AAE and MAP both identify the link state with the maximum congestion posterior probability, which proves that the generative adversarial network has learned the congestion prior probability information of the link. The performance indicators of the supervised scheme of AAE are the largest DR, the smallest FPR and the largest F1 among the other four methods, which indicates that the performance is the best. In addition, by checking the path state predicted by the encoder, it is proved that the link is correctly mapped to the corresponding path through the encoder, which effectively avoids the measurement of network topology.

[0060] Table 1 Performance evaluation index results of different algorithms

[0061] DR FPR F1 AAE unsupervised 0.849 0.141 0.846 AAE supervised 0.870 0.025 0.864 MAP 0.849 0.141 0.846 CLINK 0.849 0.141 0.846 SCFS 0.754 0.074 0.777

[0062] Further, after the model is trained, the anti-attack ability experiment of AAE model is carried out to evaluate the anti-attack performance of AAE method. This experiment is carried out in a 2x3 network topology, and the attack of flipping the label value in the Byzantine attack is selected, that is, the state value changes from 1 to 0 (congestion to good) or from 0 to 1 (good to congestion). In the unsupervised scenario, one path y1 in the attack topology is attacked, and the attack frequency is set to 10%, 12.5%,..., 40% in turn, with an attack frequency set every 2.5%, a total of 13 attack frequencies. Observe how the performance of AAE unsupervised, MAP and CLINK algorithms changes under different attack frequencies, which is mainly judged by the performance evaluation index curve, including the detection rate (DR) curve, the false positive rate (FPR) curve and the balanced average number (F1) curve. The results are shown in Figure 5 As shown in Table 1, the DR curves and F1 curves of AAE, MAP and CLINK all gradually decrease with the increase of attack frequency, and the FPR curve gradually rises with the increase of attack frequency, indicating that the attack causes the diagnostic performance to decrease, and the more frequent the attack, the more serious the performance degradation. It is difficult to say that AAE performance is better than others combined with DR and FPR curves, and F1 as a comprehensive indicator can help to judge that the F1 curve of AAE is almost higher than the F1 curves of MAP and CLINK. Therefore, it can be concluded that the anti-attack ability of AAE is stronger than that of MAP and CLINK.

[0063] The application provides a network congestion link diagnosis method based on an adversarial automatic coding, which can be applied to a scenario where internal nodes of a network do not cooperate, and includes two schemes of supervised and unsupervised, and the two schemes both infer the position of a congestion link through a congestion prior probability of a link.

[0064] The technical features of the above-described embodiments can be combined in any manner, and to make the description concise, all possible combinations of the technical features in the above-described embodiments are not described, but as long as the combinations of the technical features do not exist, they should be considered as the scope of the description.

[0065] The above-described embodiments only express several implementation manners of the application, the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the application, a number of modifications and improvements can be made, which are all within the protection scope of the application. Therefore, the protection scope of the patent of the application should be subject to the appended claims.

Claims

1. A network congestion link diagnosis method based on adversarial auto-encoding, characterized in that, The method comprises the following steps: Step S1: generating a data set containing link states and path states according to the prior probability of link congestion; Step S2: inputting the data set into an autoencoder, in a supervised scheme, superimposing a generative adversarial network on the decoder of the autoencoder, taking the decoder as the generator of the generative adversarial network, inputting the link state into the encoder, and outputting the predicted path state from the encoder, inputting the predicted path state into the decoder, and outputting the reconstructed link state from the decoder; in an unsupervised scheme, superimposing a generative adversarial network on the encoder of the autoencoder, taking the encoder as the generator of the generative adversarial network, inputting the path state into the encoder, and outputting the predicted link state from the encoder, and directly reconstructing the path state according to the predicted link state; Step S3: training the generative adversarial network in the supervised scheme and the unsupervised scheme respectively, so that the generative adversarial network learns the prior probability information of the link congestion, and outputs the link state solution with the maximum posterior probability of congestion; Step S4: iteratively updating the network parameters by backpropagating the reconstruction loss and the generative adversarial loss. 2.The network congestion link diagnosis method based on adversarial auto-encoding according to claim 1, wherein, Step S1 comprises: Step S11: generate random numbers randomly distributed in the interval [0, 1] and consistent in number with the number of links, and judge that the link is good if the random number is less than the prior probability of congestion of the link, the value of the link state is 0, and the link is congested if the random number is greater than the prior probability of congestion of the link, the value of the link state is 1, and the link state is obtained Step S12: According to the link state X and the network routing matrix A ik The path state is calculated The calculation formula is as follows: where n p is the number of paths, n c is the number of links, and A ik is the network routing matrix. 3.The network congestion link diagnosis method based on adversarial auto-encoding according to claim 2, characterized in that, When training the generative adversarial network in the supervised scheme, the following steps are included: (1) generating a plurality of link state solutions X1...Xn from the predicted path state Y' output by the encoder n , in combination with the prior probability p of congestion of the link k , each link state solution is calculated via the formula to obtain the posterior probability of congestion of the link (2) For the predicted link state X' output by the generator, the predicted congestion posterior probability P can also be calculated according to the formula in step (1) X′ ; (3) the congestion posterior probability of the link P is obtained by itself as a probability sampling X The predicted congestion posterior probability P is input into the discriminator as a positive sample X′ The predicted congestion posterior probability P is input into the discriminator as a negative sample (4) outputting the link state solution with the maximum posterior probability of congestion. 4.The network congestion link diagnosis method based on adversarial auto-encoding according to claim 2, wherein, When training the generative adversarial network in the unsupervised scheme, the following steps are included: (1) generating a plurality of link state solutions X1...Xn from path state Y inputted into the encoder n , in combination with the prior probability p of congestion of the link k , each link state solution is calculated via the formula to obtain the posterior probability of congestion of the link (2) For the predicted link state X' output by the generator, the predicted congestion posterior probability P can also be calculated according to the formula in step (1) X′ ; (3) the congestion posterior probability of the link P is obtained by itself as a probability sampling X The predicted congestion posterior probability P is input into the discriminator as a positive sample X′ The discriminator is input into the discriminator as a negative sample; (4) outputting the link state solution with the maximum posterior probability of congestion. 5.The network congestion link diagnosis method based on adversarial auto-encoding according to claim 1, wherein, Both the reconstruction loss and the generative adversarial loss use a binary cross-entropy loss function. 6.The network congestion link diagnosis method based on adversarial auto-encoding according to claim 1, wherein, Further comprising the following steps: Step S5: selecting a performance evaluation index and an experimental network topology, and verifying the performance of the adversarial autoencoder after updating the network parameters on the test set.

7. The method of claim 6, wherein the method further comprises: The selected performance evaluation index includes detection rate, false positive rate and balanced average number, and the experimental network topology uses a 3x5 network topology. 8.The network congestion link diagnosis method based on adversarial auto-encoding according to claim 1, wherein, Further comprising the following steps: Step S6: in a 2x3 network topology, using the flip label value in the Byzantine attack to attack a path in the network topology at different frequencies, and recording the performance evaluation index curve of the adversarial autoencoder under different attack frequencies.

9. The method of claim 8, wherein the method further comprises: The starting value of the attack frequency is 10%, and the terminal value is 40%, and an attack frequency is set every 2.5%. 10.The network congestion link diagnosis method based on adversarial auto-encoding according to claim 8, characterized in that, The performance evaluation index curve includes a detection rate curve, a false positive rate curve and a balanced average number curve.

Images