A method for anti-detection encrypted traffic obfuscation

Abstract

The present application relates to a kind of anti-detection encrypted traffic obfuscation method, belong to anonymous communication technical field in computer network security.This application uses the pairing mode of many-to-many to train, divide multiple websites into a group, so that attacker can identify at most which website group user traffic belongs to, but cannot confirm the specific website category visited by user.The many-to-many pairing mode increases the number of each group of websites, and the corresponding number of website group pairs is reduced, i.e.the number of cycles in single training is reduced, which significantly reduces the training time cost consumption.At the same time, since the traffic cutting stage sends traffic to two unrelated networks, the bandwidth consumption of a single network is reduced.Compared with the prior art, the present method has strong defense effect and can improve the defense effect under the premise of less defense cost.

Description

TECHNICAL FIELD

[0001] The application relates to an anti-detection encrypted traffic obfuscation method based on adversarial instances and multi-host networks, and belongs to the technical field of anonymous communication in computer network security. BACKGROUND

[0002] Nowadays, network communication gradually becomes an important part of people's daily life, and website fingerprinting attacks pose a threat to user privacy security.

[0003] Three important elements are involved in website fingerprinting attacks: users, attackers, and privacy enhancing technologies (PETs). Users use PETs to protect privacy when accessing websites to achieve anonymity in communication. However, since PETs are publicly available, attackers can collect the traffic of each website of PETs to form a traffic set, and train a classifier based on this. The classifier obtained by training can accurately identify the website category that the user is accessing. This process is called website fingerprinting attack.

[0004] Traditional website fingerprinting attack techniques improve recognition accuracy by manually extracting features, which requires researchers to invest a lot of time and effort to ensure classification accuracy. With the development of deep learning technology, researchers use deep neural networks (DNNs) to design website fingerprinting attacks. This technology can automatically extract features from original website traffic, reducing the human burden while improving recognition accuracy and robustness.

[0005] In view of the above attack mode, the existing encrypted traffic obfuscation scheme can be divided into the following two types according to the technical principle:

[0006] ①Traffic deformation: by filling in false data packets, increasing noise or cutting traffic, etc. to change the timing, length, direction and other characteristics of network traffic, so as to achieve the purpose of destroying the attacker's intention.

[0007] ②Adversarial instances: carefully designed inputs with slight perturbations to mislead the attacker's classifier to make wrong judgments.

[0008] However, the above methods have the following defects:

[0009] 1. Poor defense effect: since there is a one-to-one correspondence between the obfuscated traffic and the original traffic, the attacker can re-collect the traffic set to train the classifier, so as to perform high-precision classification again.

[0010] 2. High defense cost: in the process of training the obfuscated traffic, a large amount of time and bandwidth cost will be consumed. SUMMARY

[0011] The purpose of the present application is to solve the technical problems of the prior art, such as the insufficient defense effect of the current encrypted traffic obfuscation technology and the high defense cost, and creatively proposes a new anti-detection encrypted traffic obfuscation method, which can improve the defense effect under the premise of reasonable defense cost.

[0012] The purpose of the present application is to solve the technical problems of the prior art, such as the insufficient defense effect of the current encrypted traffic obfuscation technology and the high defense cost, and creatively proposes a new anti-detection encrypted traffic obfuscation method, which can improve the defense effect under the premise of reasonable defense cost.

[0013] An anti-detection encrypted traffic obfuscation method comprises the following steps:

[0014] Step 1: pre-training phase.

[0015] A classifier is trained based on monitoring all k website traffic in the monitoring set, which aims to assist formal training.

[0016] Step 2: grouping and pairing phase.

[0017] In this phase, all websites in the monitoring set are randomly divided into n groups, n = 2x i, i is a positive integer, and each group has p websites,

[0018] The n groups of websites are randomly paired, and each pair is used as the input of two generators. The websites included in each website group are defined as the grouping strategy (GS), and the pairing between the website groups is defined as the pairing strategy (PS).

[0019] Step 3: training phase.

[0020] In this phase, the network structure is a generative adversarial network with two generators, one discriminator and one classifier. Each pair of website groups is input into two generators, and is trained with the assistance of a discriminator and a simple classifier. The output of this phase is a traffic set with added adversarial perturbations.

[0021] Step 4: cutting phase.

[0022] Using a multi-host network, the traffic with added adversarial perturbations is randomly sent to two completely unrelated networks, and a number of virtual data packets are randomly added, thereby improving the interference to the attacker.

[0023] Advantages

[0024] Compared with the prior art, the method of the present application has the following advantages:

[0025] 1. Strong defense effect: the present application first uses a many-to-many pairing method for training, divides multiple websites into a group, and makes the attacker identify the user traffic belonging to which website group at most, but cannot confirm the specific website category accessed by the user.

[0026] 2. Low Cost: Traditional one-to-one training requires iterative training on each website pair during a single training session, resulting in excessively high training time costs. Many-to-many pairing increases the number of websites in each group, correspondingly reducing the number of website pairs and thus the number of iterations per training session, significantly reducing training time costs. Furthermore, because traffic is split across two unrelated networks during the traffic splitting phase, bandwidth consumption on a single network is reduced. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the structure of the present invention;

[0028] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0029] The present invention will now be described in further detail with reference to the accompanying drawings.

[0030] A method for obfuscating encrypted traffic to resist detection includes the following steps:

[0031] Step 1: Pre-training phase.

[0032] Based on the traffic data of all k websites in the monitoring set, a classifier SC is trained. The purpose of this classifier is to assist in the formal training, and its output is the website category corresponding to the input traffic.

[0033] Step 2: Grouping and pairing phase.

[0034] First, all websites in the monitoring set are randomly grouped, with each group containing the same number of websites.

[0035] Then, the websites are randomly paired up and used as input for the training phase.

[0036] The monitoring center contains k websites: W1, W2, ..., W k The websites are randomly divided into n groups, each containing p websites. The total number of website groups is even, to prepare for subsequent pairing training. Each website group is denoted as Set1, Set2, ..., Set... n Set n Let n represent the nth website group. During the pairing phase, the above n website groups are randomly paired up and participate in training.

[0037] Step 3: Training phase.

[0038] This stage consists of a Generative Adversarial Network (GAN), containing two generators A and B, one discriminator D, and one classifier SC, as follows:Figure 1 as shown.

[0039] First, the Gaussian noise z is taken as the input of the two generators A and B, and the output of the generator is the adversarial perturbation vector. The adversarial perturbation vector specifies the position and number of added virtual data packets.

[0040] Secondly, the generated adversarial perturbation vector and the original traffic are combined to obtain the adversarial instance. The original traffic refers to the traffic corresponding to the website group obtained after the grouping and pairing stage, which contains the characteristics of multiple website traffic.

[0041] The addition process of the adversarial perturbation vector is as follows:

[0042] The first element of the perturbation vector is added in the first positive direction data packet burst in the tracking. If the starting burst of the tracking is a negative direction data packet, the perturbation vector is right shifted to be added to the positive burst, otherwise the perturbation vector will cut off the negative burst, causing excessive time delay and affecting the function of the traffic itself. The value of the adversarial perturbation is positive, and its direction depends on the original direction of the traffic tracking to which it is added.

[0043] Finally, under the assistance of the classifier SC, the input adversarial instance is judged by the discriminator D to correspond to the website category, and the generator is iteratively trained until the output traffic tracking TA SetA (z) and the traffic tracking TA SetB (z) are very close. The role of the classifier SC is to avoid the adjustment of the generated data of only one generator in the iteration process.

[0044] Step 4: Cutting stage.

[0045] The complete traffic sent by the user to a website is divided into N tracks, each containing M data packets. Each time the traffic is cut, a network is randomly selected, and the number of data packets sent continuously on the current network this time is randomly obtained. Considering that continuous data packet sending is used to avoid excessive time delay, the number of data packets sent each time is a value extracted from a geometric distribution, which is the only memoryless discrete distribution. This will make it difficult for the attacker to predict the sending rule.

[0046] Before sending, randomly decide whether to add a dummy packet containing null value, the dummy packet only contains direction information, the purpose is to destroy the attacker through the intercepted traffic accurately analyze the traffic characteristics. In order to avoid causing excessive cost consumption, the number of additional dummy packets is very small, this number is also from the geometric distribution. In this method, the purpose of the website traffic cutting is only two. Although increasing the number of websites can greatly reduce the identification accuracy of the attacker, but the user's use of the habit of most in two websites at the same time connected, such as mobile devices through the simultaneous connection of WIFI and cellular network to improve network access performance.

[0047] As shown in Figure 2 In order to complete the complete defense settings, this method will run multiple grouping pairing stage, training stage and cutting stage, to generate a variety of different grouping strategy GS and pairing strategy PS, GS and PS one to one.

[0048] Grouping strategy GS defines the number of websites p in each website group and the specific grouping method;

[0049] Pairing strategy PS defines the pairing of each website group.

[0050] When implementing this method, the user is randomly assigned to a pairing strategy, corresponding to the confused traffic. When the attacker collects the traffic set, it is difficult to obtain the same GS and PS as the user, and the trained classifier does not conform to the traffic distribution used by the user, so it cannot correctly classify the website information that the user is accessing.

[0051] When the attacker happens to collect the same GS and PS as the user, the trained classifier can only correctly locate which website group the user is currently accessing, but cannot determine which specific website in the website group the user is currently accessing. At the same time, affected by the multi-host network, the attacker is difficult to collect complete traffic, and the accuracy of the classifier is greatly reduced.

[0052] In summary, the above process, this method realizes the complete defense process.

Claims

1. A method for anti-detection encrypted traffic obfuscation, characterized in that, The method comprises the following steps: Step 1: a pre-training phase, a classifier is trained on the basis of monitoring all k website traffic, assisting the formal training, and the output is the website category corresponding to the input traffic; Step 2: Grouping and pairing phase, all websites in the monitoring set are randomly and evenly divided into n groups, n = 2 x i, i is a positive integer, and each group has p websites, Each website group contains the same number of websites; The above n groups of websites are randomly paired two by two, and are respectively input into two generators; the definition of which websites are included in each website group is defined as a grouping strategy GS, and the pairing between the website groups is defined as a pairing strategy PS; Step 3: training phase, the network structure is a generative adversarial network comprising two generators, a discriminator and a classifier; each pair of website groups is input into the two generators, and is trained with the assistance of the discriminator and the simple classifier; The output of this stage is a traffic set added with an adversarial disturbance; Step 4: cutting phase, using a multi-host network, the traffic added with the adversarial disturbance is randomly sent to two completely unrelated networks, and a plurality of virtual data packets are randomly added to improve the interference on the attacker; The grouping and pairing phase, the training phase and the cutting phase are executed multiple times to generate multiple different grouping strategies GS and pairing strategies PS, and the GS and the PS correspond one by one; In implementation, a user is randomly assigned to a pairing strategy, and corresponding confused traffic; When the attacker collects the traffic set, it is difficult to obtain the same GS and PS as the user, and the trained classifier does not conform to the traffic distribution used by the user; when the attacker happens to collect the same GS and PS as the user, the trained classifier can only correctly locate which website group the website currently accessed by the user belongs to, but cannot determine which specific website in the website group the user is currently accessing; meanwhile, affected by the multi-host network, the attacker is difficult to collect complete traffic.

2. The anti-detection encrypted traffic obfuscation method of claim 1, wherein, In step 2, k websites are monitored collectively: W1, W2, …, Wk k , which are randomly and evenly divided into n groups, each group containing p websites, and the total number of website groups is even; Each website group is denoted as Set1, Set2, …, Set n , Set n n represents the nth website group; In the pairing phase, the above n website groups are randomly paired two by two and participate in training.

3. The anti-detection encrypted traffic obfuscation method of claim 1, wherein, In step 3, the network structure comprises a generative adversarial network, including two generators A and B, one discriminator D and one classifier SC; Firstly, the Gaussian noise z is input into the two generators A and B, and the output of the generator is an adversarial disturbance vector; the adversarial disturbance vector specifies the position and number of added virtual data packets; Secondly, the generated adversarial disturbance vector and the original traffic are combined to obtain an adversarial instance; wherein the original traffic refers to the traffic corresponding to the website group obtained after the grouping and pairing phase, and contains the features of multiple website traffic; The adding process of the adversarial disturbance vector is as follows: The first element of the disturbance vector is added in the first positive direction data packet burst in the tracking, if the starting burst of the tracking is a negative direction data packet, the disturbance vector is right shifted to be added to the positive burst; the value of the adversarial disturbance is positive, and its direction depends on the original direction of the traffic tracking to which it is added; Finally, with the aid of the classifier SC, the input of the discriminator D is judged to correspond to the website category of the generator, and the mutual iteration training is performed until the output of the traffic tracking TA SetA (z) is very close. SetB (z) is very close.

4. The anti-detection encrypted traffic obfuscation method of claim 1, wherein, In step 4, the complete traffic sent by the user to a website is divided into N traces, each trace containing M data packets; each time the traffic is cut, a network is randomly selected, and the number of data packets continuously sent on the current network this time is randomly obtained, and the number of data packets sent each time is a value extracted from a geometric distribution, and the geometric distribution is the only non-memory discrete distribution; Before sending, randomly decide whether to add a dummy packet containing null data, which only contains direction information, aiming to destroy the attacker's accurate analysis of traffic characteristics through intercepted traffic.

Images