Method and terminal for active suppression of denial of service attack in high dynamic crowd intelligence network

Abstract

The application discloses a high-dynamic crowd wisdom network active inhibition denial of service attack method and a terminal, relates to the technical field of information security, and comprises a cloud server and a plurality of multi-system high-dynamic crowd wisdom network clients. The method calculates the relative weight of the client by using a truth discovery algorithm on the parameters uploaded by the client, and iteratively divides the client by using the relative weight, so that the purpose of detecting malicious clients and actively inhibiting denial of service attacks is achieved. The method solves the problems that the existing defense methods cannot defend against some attacks, the numerical statistical defense method has poor effect, and malicious clients cannot be detected and distinguished.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information security, and particularly relates to a high-dynamic crowd intelligence network active inhibition denial of service attack method and terminal. BACKGROUND

[0002] With the popularity of mobile Internet and smart devices, more and more data is generated and scattered in various devices and terminals, which makes the traditional centralized machine learning method face the problems of data centralization and privacy leakage. Therefore, distributed machine learning technology gradually replaces centralized machine learning, among which, federated learning is a common distributed machine learning technology, which provides distributed model training between multiple data sources with local data. Without exchanging local individual or sample data, only by exchanging model parameters or intermediate results, a global model based on virtual fusion data is constructed, so as to realize the balance between data privacy protection and data sharing calculation. With the rapid development of big data and AI, the demand for more efficient and faster machine learning methods is increasing, and the federated learning method of multi-institutional high-dynamic crowd intelligence network has the advantages of distribution, parallelism, real-time, etc., which can well meet these demands. As a multi-institutional high-dynamic crowd intelligence network machine learning method, federated learning has become one of the current research hotspots and has a wide application prospect.

[0003] Federated learning is a multi-institutional high-dynamic crowd intelligence network machine learning method. Multi-institutional means that federated learning involves multiple entities (such as devices, nodes, users, etc.), each entity has its own local data and computing power. These entities work together in the federated learning process to jointly train and improve the model. Different entities may have different operating systems, hardware devices, network connections, and other different technical architectures, which makes federated learning a multi-institutional learning method. High-dynamic means that federated learning is a highly dynamic process, entities can join or exit the federated learning network at any time, and local data and computing power may also change at any time. Crowd intelligence network means that federated learning is a collective wisdom process, through the collaborative work of multiple entities to achieve better learning effect. These entities can be individuals or organizations distributed in different geographical locations, different industries, and different fields. Federated learning takes advantage of this crowd intelligence network to collect the local data and knowledge of different entities to achieve better model training and improvement.

[0004] In a traditional denial of service attack, the attacker usually uses a large number of requests to occupy the resources of the server, causing the server to be unable to normally provide services. Similarly, in federated learning, an attacker can poison the entire model by manipulating the data on some client devices and uploading false data to the federated learning system, so that the federated learning system cannot normally provide services, thereby achieving the effect of denial of service attack. The denial of service attack of the multi-institutional high-dynamic crowd wisdom network refers to that the malicious client submits malicious parameters for poisoning attack, so that the precision of the global model of the federated learning decreases or the model is unavailable, so as to achieve the purpose of denial of service attack. The existing denial of service attack defense methods of the multi-institutional high-dynamic crowd wisdom network have the following problems: (1) some defense methods can only defend against specific attacks (2) some defense methods have poor defense effect based on numerical statistics methods (such as: median, average) (3) some defense methods cannot detect malicious clients or need additional information to judge malicious clients.

[0005] Therefore, it is urgent to propose a high-dynamic crowd wisdom network active inhibition denial of service attack method and terminal to solve the above problems. SUMMARY

[0006] In view of the problems existing in the prior art, the present application provides a high-dynamic crowd wisdom network active inhibition denial of service attack method and terminal, which uses a truth discovery algorithm to calculate the parameters uploaded by the client, obtains the relative weight, and uses the obtained relative weight to iteratively divide each client, so as to detect the malicious client and actively inhibit the denial of service attack.

[0007] The technical scheme of the present application is as follows:

[0008] The high-dynamic crowd wisdom network active inhibition denial of service attack method comprises a cloud server and a plurality of clients; the method steps are as follows:

[0009] Step 1, cloud server initialization, model parameter distribution: the cloud server initializes the global model ω0 according to the task, and sends ω0 to all clients;

[0010] Step 2, client local training, model parameter uploading: the client receives ω0 and trains using the local data set;

[0011] Step 3, cloud server detects malicious clients and benign clients: the cloud server uses a truth discovery algorithm to calculate the relative weight of the client, calculates a threshold value and groups the clients in combination with the relative weight;

[0012] Step 4, the cloud server aggregates the benign model and distributes it to each benign client;

[0013] Step 5, repeat step 2 to step 4 until the global model converges.

[0014] As a further optimization of the above scheme, in step 2, the process of training is represented as:

[0015]

[0016] where F(ω) represents the objective function, represents the gradient of the t-th round of training model, η represents the learning rate, and t represents the number of local training.

[0017] As a further optimization of the above scheme, in step 2, the objective function F(ω) is determined by the task of the cloud server and the selected model; for example, training a neural network model for image classification, a stochastic gradient descent (SGD) algorithm can be run on the client to iteratively minimize the classification error of all training images available to the client.

[0018] As a further optimization of the above scheme, the detailed steps of step 3 are as follows:

[0019] Step 3-1, the cloud server randomly selects 3 clients;

[0020] Step 3-2, the cloud server uses the truth discovery algorithm to calculate the relative weight of the client, calculates the threshold and groups the clients according to the relative weight; the grouping basis is:

[0021]

[0022] where u∈N, N is the number of clients, w u is the relative weight of the u-th client, G1, G2 are client groups, T is the threshold and T = mean(w u )-α, α is a control parameter, |w| indicates the number of clients selected;

[0023] Step 3-3, select the group with the most clients as G1, and the other group as G2; G1 randomly selects m clients that have not been detected to form a client group G, that is, G = {G1∪m}; the relative weight of the clients in G is calculated using the truth discovery algorithm, and then divided by the threshold; where m = min{|G1|-1, U}, U is the number of clients that have not been detected, and the division method is G 2' is merged with G2 to form a new G2.

[0024] Step 3-4, repeat step 3-3 until all clients are grouped.

[0025] As a further optimization of the above scheme, in step 4, the convergence method of the benign model is represented as Where Agg is an aggregation function, is a new model after aggregation.

[0026] As a further optimization of the above scheme, in step 4, the aggregation function Agg is one of any federated aggregation functions; for example, in the Fedavg algorithm, Where, D u is the number of data sets of the u-th local client.

[0027] As a further optimization of the above scheme, the detailed steps of the truth discovery algorithm are as follows:

[0028] Step S1, calculate the estimated aggregation value:

[0029]

[0030] Where N is the number of clients, w u is the relative weight of the u-th client, i represents the i-th dimension, represents the i-th dimension value of the u-th client.

[0031] Step S2, update the relative weight of the client through the estimated aggregation value:

[0032]

[0033] Where p represents the dimension, d() is the distance function, i.e. d(x,y) = (x-y) 2 ;

[0034] Step S3, repeat step S1 and step S2 until the estimated aggregation value converges.

[0035] As a further optimization of the above scheme, in step 3, the benign client is a group of clients whose number is greater than 50% after grouping, and the remaining clients are all malicious clients; in the attack and defense scene of the multi-system high dynamic crowd wisdom network, the malicious clients are generally set to be a minority, i.e. the proportion of the number of clients is < 50%.

[0036] As a further optimization of the above scheme, a plurality of said clients form a multi-system, high-dynamic crowd wisdom network.

[0037] The application also discloses a terminal comprising a storage device for storing a plurality of instructions and a processor for executing each instruction in the storage device; the instructions are suitable for being loaded and executed by the processor to implement the active suppression denial of service attack method for the high-dynamic crowd wisdom network.

[0038] Compared with the prior art, the application has the following beneficial effects:

[0039] The application provides a high-dynamic crowd intelligence network-oriented active denial-of-service attack suppression method and terminal, which has many advantages.

[0040] (1) The existing denial-of-service attack method can be prevented only by the parameter information uploaded by the client;

[0041] (2) The malicious client and the benign client can be distinguished without additional information;

[0042] (3) Strong compatibility, which can be compatible with the existing multi-system high-dynamic crowd intelligence network without changing the system framework of the multi-system high-dynamic crowd intelligence network. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The application provides a high-dynamic crowd intelligence network-oriented active denial-of-service attack suppression method and terminal, which has many advantages.

[0044] Figure 2 The application provides a high-dynamic crowd intelligence network-oriented active denial-of-service attack suppression method and terminal, which has many advantages. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical scheme and advantages of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only a part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0046] As shown in Figure 1 , Figure 2 The application discloses a high-dynamic crowd intelligence network-oriented active denial-of-service attack suppression method, which comprises a cloud server and a plurality of clients, and the plurality of clients form a multi-system high-dynamic crowd intelligence network. The method steps are as follows:

[0047] Step 1, cloud server initialization, model parameter issuing: the cloud server initializes a global model ω0 according to a task, and sends ω0 to all clients;

[0048] Step 2, client local training, model parameter uploading: the client receives ω0, and trains by using a local data set; wherein, the training process is represented as:

[0049]

[0050] Wherein, F(ω) represents a target function, denotes the gradient of the t-th round of training model, η denotes the learning rate, t denotes the number of local training; the objective function F(ω) is determined by the task of the cloud server and the selected model; for example, training a neural network model for image classification, a stochastic gradient descent (SGD) algorithm can be run on the client to iteratively minimize the classification error of all training images available to the client;

[0051] Step 3, the cloud server detects malicious clients and benign clients: the cloud server uses the truth discovery algorithm to calculate the relative weights of the clients, calculates the threshold value, and groups the clients in combination with the relative weights;

[0052] In this embodiment, the detailed steps of step 3 are as follows:

[0053] Step 3-1, the cloud server randomly selects 3 clients;

[0054] Step 3-2, the cloud server uses the truth discovery algorithm to calculate the relative weights of the clients, calculates the threshold value, and groups the clients in combination with the relative weights; the grouping basis is:

[0055]

[0056] wherein u is a number and u∈N, u∈N, N is the number of clients, w u denotes the relative weight of the u-th client, G1, G2 are client groups, which store client numbers, T is a threshold value and T=mean(w)-α, α is a control parameter, w denotes the number of clients to be screened; the threshold value T is the mean value of the client weights minus the control parameter α, when the client weights are similar, the clients are all divided into a group by setting α;

[0057] Step 3-3, select the group with the largest number of clients as G1, and the other group as G2; G1 randomly selects m undetected clients to form a client group G, that is, G={G1∪m}; the relative weights of the clients in G are calculated by using the truth discovery algorithm, and then divided by the threshold value; wherein m=min{|G1|-1,U}, U is the number of undetected clients, and the division method is G obtained by dividing 2' and G2 are merged to form a new G2;

[0058] Step 3-4, repeat step 3-3 until all clients are grouped; wherein the benign clients are a group with more than 50% of the number of clients after grouping, and the remaining clients are malicious clients; in the attack and defense scene of the multi-system high dynamic crowd wisdom network, the malicious clients are generally set to be a minority, that is, the proportion of the number of clients is <50%;

[0059] The detailed steps of the truth discovery algorithm are as follows:

[0060] Step S1, calculating an estimated aggregation value:

[0061]

[0062] Wherein, N is the number of clients, w u is the relative weight of the u-th client, i represents the i-th dimension, represents the i-th dimension value of the u-th client;

[0063] Step S2, updating the relative weight of the client through the estimated aggregation value:

[0064]

[0065] Wherein, p represents the dimension, d() is a distance function, i.e. d(x,y) = (x-y) 2 ;

[0066] Step S3, repeating step S1 and step S2 until the estimated aggregation value converges;

[0067] Step 4, the cloud server aggregates the benign model and delivers it to each benign client; in this embodiment, the convergence method of the benign model is represented as Wherein, Agg is an aggregation function, Agg is one of any federated aggregation functions, is the new model after aggregation; for example, in the Fedavg algorithm, Wherein, D u is the number of data sets of the u-th local client;

[0068] Step 5, repeating steps 2 to 4 until the global model converges.

[0069] The embodiment also discloses a terminal comprising a storage device storing a plurality of instructions and a processor for executing each instruction in the storage device; the instructions are suitable for being loaded and executed by the processor to implement the active suppression denial of service attack method for high dynamic crowd wisdom network.

[0070] In order to facilitate understanding of the implementation of the technical solution, in this embodiment, specific data is used for demonstration:

[0071] Step 1, cloud server initialization, model parameter distribution: the cloud server initializes the global model ω0 according to the task, and sends ω0 to all clients; In order to facilitate the description, the model is instantiated as ω0=[8.7036e-02,3.2684e-03,1.5362e-02,-1.0842e-01,-7.1150e-02], There are 10 clients, of which 6 are benign clients and 4 are malicious clients. The clients are numbered, 0-3 are malicious clients, and 4-9 are benign clients;

[0072] Step 2, local training of client, upload model parameters: the client receives ω0, and trains using the local data set; In order to facilitate the description, the parameters uploaded by the 6 benign clients are [[-8.5535e-02, 2.5329e-01, 5.5006e-02, 4.4094e-03, 7.1967e-02], [-7.7181e-02, 2.6037e-01, 5.3830e-02, -7.9036e-04, 7.1707e-02], [-8.6689e-02, 2.5315e-01, 5.0962e-02, -6.6478e-04, 7.3116e-02], [-8.6570e-02, 2.5349e-01, 5.4062e-02, 7.1972e-04, 7.0153e-02], [-8.6373e-02, 2.5493e-01, 5.3346e-02, -1.4878e-04, 7.2126e-02], [-8.3871e-02, 2.5935e-01, 5.6294e-02, 8.0938e-04, 7.0597e-02]], The malicious parameters uploaded by the 4 malicious clients are [[8.5535e-02, -2.5329e-01, -5.5006e-02, -4.4094e-03, -7.1967e-02], [7.7181e-02, -2.6037e-01, -5.3830e-02, 7.9036e-04, -7.1707e-02], [8.6689e-02, -2.5315e-01, -5.0962e-02, 6.6478e-04, -7.3116e-02], [8.6570e-02, -2.5349e-01, -5.4062e-02, -7.1972e-04, -7.0153e-02]];

[0073] Step 3, cloud server detects malicious clients and benign clients: the cloud server uses the truth discovery algorithm to calculate the relative weight of the client, calculates the threshold and combines the relative weight to group the clients; Including the following steps:

[0074] Step 3.1, randomly select 3 clients [5 0 7];

[0075] Step 3.2, calculate the relative weight of the clients using the truth discovery algorithm, and group by the relative weight: divide the clients by weight by setting a threshold, which can be formally expressed as:

[0076]

[0077] Where T = mean(w) - a, a = 0.5;

[0078] Step 3-3, select the group with the most clients as G1, and the other group as G2; G1 randomly selects m undetected clients to form a client group G, that is, G = {G1∪m}; calculate the relative weight of the clients in G using the truth discovery algorithm, and then divide by threshold; where m = min{|G1|-1, U}, U is the number of undetected clients, and the division method is The G obtained by dividing 2' Merge with G2 to form a new G2;

[0079] Step 3-4, repeat step 3-3 until all clients are grouped;

[0080] In step 3, the relative weight of the clients is calculated by the truth discovery algorithm, [9.1179e+00,2.1873e-04,9.1239e+00], the division result G1 = [5 7], G2 = [0];

[0081] G1 selects client [3], G = [5 7 3] calculates the relative weight, [0.9.1173e+00,0.9.1233e+00,0.0002.1884e-04], the division result G1 = [5 7], G2 = [0 3];

[0082] G1 selects client [8], G = [5 7 8] calculates the relative weight, [0.0290,4.1294,4.3799], after division and adjustment, the result is G1 = [0 3 5], G2 = [7 8];

[0083] G1 selects clients [4 6], G = [0 3 5 4 6] calculates the relative weight, [0.7033,0.7039,5.5481,5.7232,5.7363], after division and adjustment, the result is G1 = [7 8 0 3], G2 = [5 4 6];

[0084] G1 selects the client

[192] , G = [780 3192] calculates the relative weight, [1.1578, 1.1510, 4.4288, 4.4332, 4.3471, 1.1366, 4.4346], after division and adjustment, the result is G1 = [546 789], G2 = [03 12];

[0085] Since the number of clients of G1 is greater than 50%, G1 is a benign client group, and G2 is a malicious client group;

[0086] Step 4, the cloud server aggregates the benign model and distributes it to each benign client;

[0087] Step 5, repeat steps 2 to 4 until the global model converges, use the aggregation function to converge the model parameters, get the final model

[0088]

[0089] According to the disclosure and teaching of the above description, those skilled in the art of the present application can also make changes and modifications to the above embodiments. Therefore, the present application is not limited to the specific embodiments disclosed and described above, and some modifications and changes of the present application should fall within the protection scope of the claims of the present application. In addition, although some specific terms are used in the specification, these terms are only for convenience of explanation and do not constitute any limitation on the present application.

Claims

1. A method for active suppression of denial of service attacks in high dynamic crowd intelligence networks, comprising a cloud server and a plurality of clients; characterized in that, The method steps are as follows: Step 1, cloud server initialization, model parameter issuing: the cloud server initializes a global model ω0 according to a task, and sends ω0 to all clients; Step 2, client local training, model parameter uploading: the client receives ω0 and trains using a local data set; Step 3, cloud server detection of malicious clients and benign clients: the cloud server uses a truth discovery algorithm to calculate the relative weight of the client, calculates a threshold and groups the clients in combination with the relative weight; the detailed steps are as follows: Step 3-1, the cloud server randomly selects 3 clients; Step 3-2, the cloud server uses a truth discovery algorithm to calculate the relative weight of the client, calculates a threshold and groups the clients in combination with the relative weight; the grouping basis is: wherein u∈N, N is the number of clients, w u G1, G2 are client groups, T is a threshold and T = mean(w) - a, a is a control parameter, |w| indicates the number of clients being filtered; Step 3-3, select the group with the largest number of clients as G1, and the other group as G2; G1 randomly selects m undetected clients to form a client group G, i.e. G = {G1∪m}; the relative weights of the clients in G are calculated by the truth discovery algorithm, and then divided by a threshold; wherein m = min{|G1|-1, U}, U is the number of undetected clients, and the division method is G 2' Merging with G2 to form a new G2; Step 3-4, repeat step 3-3 until all clients are grouped; wherein the benign client is a group with more than 50% of the number of clients after grouping, and the remaining clients are malicious clients; Step 4, the cloud server aggregates the benign model and issues it to each benign client; Step 5, repeat steps 2 to 4 until the global model converges.

2. The method of claim 1, wherein the method is characterized by, In step 2, the training process is represented as: where F(ω) represents the objective function, denotes the gradient of the t-th round of training model, η represents the learning rate, and t represents the number of local training.

3. The method of claim 2, wherein the method is characterized by, In step 2, the objective function F(ω) is determined by the task of the cloud server and the selected model.

4. The high-mobility crowd network-oriented active suppression denial-of-service attack method of claim 1, wherein, In Step 4, the convergence method of the benign model is represented as where Agg is an aggregation function, is the new model after aggregation.

5. The high-mobility crowd network active suppression denial-of-service attack method of claim 4, wherein, In step 4, the aggregation function Agg is one of any federated aggregation functions.

6. The high-mobility crowd network-oriented active suppression denial-of-service attack method according to claim 1, characterized in that, The detailed steps of the truth discovery algorithm are as follows: Step S1, calculate the estimated aggregation value: where N is the number of clients, w u is the relative weight of the u-th client, i denotes the i-th dimension, denotes the i-th dimension value of the u-th client; Step S2, update the relative weight of the client through the estimated aggregation value: where p represents the dimension, d() is a distance function, i.e. d(x,y) = (x-y) 2 ; Step S3, repeat step S1 and step S2 until the estimated aggregation value converges.

7. The high-mobility crowd network-oriented active suppression denial-of-service attack method according to claim 1, characterized in that, A plurality of the clients form a multi-system, high-dynamic crowd intelligence network.

8. A terminal comprising a storage device storing a plurality of instructions and a processor for executing the instructions in the storage device; characterized in that, The instructions are suitable for the processor to load and execute the active suppression denial of service attack method for a high-dynamic crowd intelligence network as claimed in any one of claims 1 to 7.

Images