Robustness Authentication Method for Intelligent Traffic Analysis Against Multimodal Disturbances

Abstract

This application relates to the field of network traffic analysis technology, and in particular to a robust authentication method for intelligent traffic analysis oriented towards multimodal adversarial attacks. The method includes: randomly selecting data packet sequences multiple times from the original traffic data; generating smoothed samples based on the selected data packet sequences after each random selection; constructing a smoothed sample set based on the smoothed samples; processing the smoothed samples in the smoothed sample set into the data format required by the base classifier; inputting the processed smoothed sample set into the smoothed classifier; and outputting the classification results of the smoothed sample set. The corresponding robust radius is then derived based on the classification results. This solves the problems of related technologies, which, when dealing with adversarial attacks in traffic analysis, only consider a single type of disturbance, cannot effectively handle common combined disturbance scenarios in reality, and lack model compatibility, making it difficult to meet the security requirements of real-world network environments.

Description

Technical Field

[0001] This application relates to the field of network traffic analysis technology, and in particular to a robust authentication method for intelligent traffic analysis oriented towards multimodal adversarial disturbances. Background Technology

[0002] With the widespread adoption of internet applications, the scale and complexity of network traffic are constantly increasing, making traffic analysis systems, as a core means of ensuring network security, increasingly important. Current mainstream traffic analysis systems mostly rely on deep learning models, which can effectively identify potential threats in encrypted traffic; however, these models generally suffer from insufficient robustness.

[0003] Adversarial attacks can significantly affect model output by applying small perturbations to input traffic (such as packet padding, delays, or packet insertion / deletion), inducing misjudgments and thus posing security risks. Existing research has attempted to improve model robustness in two directions: one is empirical methods, such as data augmentation, feature engineering, and model structure optimization, but these methods cannot provide theoretical robustness guarantees; the other is verifiable robustness methods, especially stochastic smoothing techniques, which can provide mathematical guarantees for the perturbation robustness radius of deep model outputs. However, most existing stochastic smoothing methods are designed for continuous spatial tasks such as images and cannot effectively cope with adversarial attacks in the traffic domain that involve both additive and discrete perturbations.

[0004] Although existing work has introduced certifiable robustness into traffic analysis, it only considers a single type of disturbance, fails to cover complex attack scenarios in reality, has a significantly reduced robustness when facing combined disturbances, and does not have good model compatibility, making it difficult to meet the security requirements of real network environments. Summary of the Invention

[0005] This application provides a robust authentication method for intelligent traffic analysis oriented towards multimodal adversarial disturbances, in order to solve the problems that related technologies only consider a single type of disturbance when dealing with adversarial attacks in traffic analysis, which cannot effectively handle common joint disturbance scenarios in reality, and lack model compatibility, making it difficult to meet the security requirements in real network environments.

[0006] The first aspect of this application provides a robustness authentication method for intelligent traffic analysis against multimodal disturbances, comprising the following steps: randomly selecting data packet sequences multiple times from the original traffic data; generating smoothed samples based on the selected data packet sequences after each random selection; constructing a smoothed sample set based on the smoothed samples; processing the smoothed samples in the smoothed sample set into the data format required by the base classifier; inputting the processed smoothed sample set into the smoothed classifier; outputting the classification result of the smoothed sample set from the smoothed classifier; and deriving the corresponding robust radius based on the classification result.

[0007] Optionally, generating a smoothed sample based on the selected data packet sequence includes: obtaining the packet length sequence and time interval sequence of the data packet sequence; determining random noise conforming to an exponential distribution based on the packet length sequence and time interval sequence; and adding the random noise to the data packet sequence to obtain a smoothed sample.

[0008] Optionally, a smooth classifier is constructed from the base classifier, wherein the classification process of the smooth classifier includes: using the base classifier to identify the category of each smooth sample; generating a classification result set based on the category of each smooth sample; and taking the category that appears most frequently in the classification result set as the classification output result.

[0009] Optionally, before using the most frequent category in the classification result set as the classification output, the method includes: calculating the frequency of occurrence of each category in the classification result set; and determining the most frequent category based on the frequency of occurrence of each category.

[0010] Optionally, the corresponding robust radius is derived based on the classification output, including: using discrete smoothing mechanism and additive smoothing mechanism to derive the robust region of the smoothing classifier in at least one of discrete perturbation and additive perturbation, and determining the robust radius within the robust region.

[0011] Optionally, discrete perturbations include perturbations applied to the flow at the packet level, while additive perturbations include perturbations applied to the metadata of the packet at the packet level.

[0012] A second aspect of this application provides an intelligent traffic analysis robustness authentication device for multimodal adversarial disturbances, comprising: a selection module, which randomly selects data packet sequences multiple times from the original traffic data; after each random selection, it generates smoothed samples based on the selected data packet sequences and constructs a smoothed sample set based on the smoothed samples; a processing module, which processes the smoothed samples in the smoothed sample set into the data format required by the base classifier, inputs the processed smoothed sample set into the smoothed classifier, and the smoothed classifier outputs the classification result of the smoothed sample set; and a derivation module, which derives the corresponding robust radius based on the classification output result.

[0013] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the program to implement the robust authentication method for intelligent traffic analysis oriented against multimodal disturbances as described in the first aspect.

[0014] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement a robust authentication method for intelligent traffic analysis oriented against multimodal disturbances, as described in the first aspect.

[0015] The fifth aspect of this application provides a computer program product, which, when executed, is used to implement the intelligent traffic analysis robustness authentication method for multimodal adversarial disturbances as described in the first aspect.

[0016] Therefore, this application has the following beneficial effects:

[0017] This application's embodiments involve repeatedly randomly selecting data packet sequences from the original traffic data. After each random selection, smoothed samples and a smoothed sample set are generated based on the selected sequence. The smoothed samples in the smoothed sample set are processed, and the processed smoothed sample set is input into a smoothing classifier. The smoothing classifier outputs the classification result of the smoothed sample set, and finally, the corresponding robust radius is derived based on the classification result. This approach has low computational cost, requires no understanding or modification of the model's internal structure, and can adapt to multimodal real-world attacks, providing better robustness assurance for traffic analysis systems. It also exhibits higher model versatility, widely applicable to models with various architectures and input formats. Therefore, it solves the problems of related technologies that, when dealing with adversarial attacks in traffic analysis, only consider a single type of disturbance, cannot effectively handle common combined disturbance scenarios in reality, and lack model compatibility, making it difficult to meet the security requirements of real network environments.

[0018] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0019] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0020] Figure 1 This is a flowchart illustrating a robustness authentication method for intelligent traffic analysis against multimodal disturbances provided in an embodiment of this application.

[0021] Figure 2 This is a logical schematic diagram of a traffic analysis robustness authentication method for multimodal disturbance resistance provided according to an embodiment of this application;

[0022] Figure 3 This is a schematic diagram illustrating the robustness derivation of the discrete smoothing mechanism in a traffic analysis robustness authentication method for multimodal disturbance resistance provided according to an embodiment of this application;

[0023] Figure 4 This is a schematic diagram illustrating the robustness derivation of the additive stochastic smoothing mechanism in a traffic analysis robustness authentication method for multimodal disturbance resistance provided according to an embodiment of this application;

[0024] Figure 5This is a schematic diagram illustrating the derivation of the joint robust region in a traffic analysis robustness authentication method for multimodal disturbance resistance provided according to an embodiment of this application;

[0025] Figure 6 This is a block diagram of an intelligent traffic analysis robustness authentication device for multimodal disturbance resistance provided according to an embodiment of this application;

[0026] Figure 7 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0027] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0028] The following describes an embodiment of the robustness authentication method for intelligent traffic analysis oriented towards multimodal adversarial disturbances, with reference to the accompanying drawings. Addressing the issues mentioned in the background art, which consider only a single disturbance type when dealing with adversarial attacks in traffic analysis, failing to effectively handle common combined disturbance scenarios in practice, and lacking model compatibility, making it difficult to meet security requirements in real network environments, this application provides a robustness authentication method for intelligent traffic analysis oriented towards multimodal adversarial disturbances. In this method, multiple random selections of data packet sequences are performed on the original traffic data. After each random selection, smoothed samples and a smoothed sample set are generated based on the selected data packet sequence. The smoothed samples in the smoothed sample set are processed, and the processed smoothed sample set is input into a smoothing classifier. The smoothing classifier outputs the classification result of the smoothed sample set. Finally, the corresponding robust radius is derived based on the classification output result. This method has low computational complexity, requires no understanding or modification of the model's internal structure, and can adapt to real multimodal attacks, providing better robustness assurance for traffic analysis systems. It also has higher model versatility and is widely applicable to models with various architectures and input formats. This solves the problems that related technologies, when dealing with adversarial attacks in traffic analysis, only consider a single type of disturbance, cannot effectively handle common combined disturbance scenarios in reality, and lack model compatibility, making it difficult to meet the security needs of real network environments.

[0029] Specifically, Figure 1 This is a flowchart illustrating a robustness authentication method for intelligent traffic analysis against multimodal disturbances, provided as an embodiment of this application.

[0030] like Figure 1As shown, the robustness certification method for intelligent traffic analysis against multimodal disturbances includes the following steps:

[0031] In step S101, the original traffic data is randomly selected multiple times for data packet sequences. After each random selection of data packet sequences, a smoothed sample is generated based on the selected data packet sequence, and a smoothed sample set is constructed based on the smoothed sample.

[0032] The raw traffic data refers to the actual traffic data transmitted during network communication, which usually flows in the network in the form of data packets. It generally includes information such as the timestamp of the data packet, source IP address, destination IP address, port number, protocol type, and payload size. The process of randomly selecting a data packet sequence each time is as follows: based on a flow x containing n data packets, d data packets are randomly selected from the n data packets as the data packet sequence, and the other data packets are deleted. The method for generating smooth samples will be described in detail below and will not be repeated here.

[0033] It is understood that, in this embodiment of the application, based on a stream x containing n data packets, d data packets are randomly selected from the n data packets as a data packet sequence, and other data packets are deleted. Then, a smoothed sample is generated based on the obtained data packet sequence, and a smoothed sample set is constructed based on the smoothed sample. Specifically:

[0034] In this embodiment of the application, generating a smoothed sample based on the selected data packet sequence includes: obtaining the packet length sequence and time interval sequence of the data packet sequence; determining random noise conforming to an exponential distribution based on the packet length sequence and time interval sequence; and adding the random noise to the data packet sequence to obtain a smoothed sample.

[0035] Among them, packet length sequence refers to the effective payload size of each data packet in the data packet sequence, which is a numerical sequence formed by arranging them in chronological order in bytes; time interval sequence refers to the time difference between adjacent data packets, which is usually a numerical sequence formed by arranging them in chronological order in milliseconds or microseconds; exponential distribution is a continuous probability distribution, characterized by "memorylessness", that is, future behavior is unrelated to the past; random noise refers to small-amplitude perturbation signals added artificially, used to simulate additive perturbations commonly used in adversarial attacks.

[0036] It is understood that, in the embodiments of this application, for a selected data packet sequence, the packet length and time interval information are first extracted to form two numerical sequences; then, based on the statistical characteristics of these two sequences, random noise following an exponential distribution is generated; finally, these noises are added to the corresponding packet length and time interval respectively to construct new samples with perturbation characteristics, i.e., smoothed samples. The above process of randomly selecting and retaining a portion of data packets and introducing random noise following an exponential distribution into the packet length sequence and time interval sequence of the selected data packets is repeated several times to obtain a smoothed sample set of all smoothed samples.

[0037] In step S102, the smoothed samples in the smoothed sample set are processed into the data format required by the base classifier, and the processed smoothed sample set is input into the smoothed classifier. The smoothed classifier outputs the classification result of the smoothed sample set.

[0038] Among them, the base classifier f is the traffic classification model in the original traffic analysis system that does not employ a verifiable robustness method. The smooth classifier will be described in detail below and will not be repeated here.

[0039] It is understood that in this embodiment, the smoothed samples in the smoothed sample set are preprocessed to conform to the data format required by the base classifier, and the processed smoothed samples are sent into the smoothed classifier. The smoothed classifier will perform a classification operation on each sample and output the corresponding classification output result. That is, in this embodiment, the smoothed samples are first processed to adapt to the requirements of the base classifier which does not have additional robustness guarantee, and finally the classification results are summarized for subsequent robustness analysis.

[0040] In this embodiment of the application, a smooth classifier is constructed from the base classifier. The classification process of the smooth classifier includes: using the base classifier to identify the category of each smooth sample; generating a classification result set based on the category of each smooth sample; and taking the category that appears most frequently in the classification result set as the classification output result.

[0041] The smooth classifier is obtained based on the base classifier and the smoothing function. Specifically, the standard definition of the smooth classifier in this application embodiment is: given the base classifier f, let... for Given a set of random variables, and a smoothing function f will be a collection of streams Mapping to a collection of categories Above, using the base classifier f and the smoothing function ψ, we can obtain the smoothing classifier g. g always returns the most likely classification result of the base classifier f for the smoothed sample ψ(x), that is... Where, p y (x) is The abbreviation for represents the probability that the output of the base classifier f of the random variable z is class y, given that the random variable z follows a distribution defined by a smooth function ψ(x) with x as input. For input x, the classification result g(x) given by the smooth classifier is denoted as y. A The corresponding probability Noted as p A Since there is actually no way to know the true p A Therefore, the Monte Carlo method is used to estimate p with confidence level α. A lower bound p A As an alternative, getp A Then, the robust region of the smooth classifier g with input x can be determined. Estimation is performed to determine the smooth classifier g for any input within the robust region R(x). The classification results are all the same as g(x), that is:

[0042] It is understood that the embodiments of this application construct a smooth classifier from the base classifier. The smooth classifier first identifies the category of each smooth sample through the base classifier, performs statistics on the smooth samples, generates a classification result set, and counts the frequency of each category in the classification result set. The dominant category with the highest frequency is taken as the classification output result of the smooth classifier.

[0043] In this embodiment of the application, before using the most frequent category in the classification result set as the classification output result, the method includes: calculating the occurrence frequency of each category in the classification result set; and determining the most frequent category based on the occurrence frequency of each category.

[0044] It is understood that, before using the category that appears most frequently in the classification result set as the final classification output, the embodiments of this application first need to calculate the frequency of each category in the classification result set, that is, count the number of times each category appears in the classification result set, and determine which category appears most frequently based on these frequencies. Specifically, it is to traverse the classification result set, record the number of times each category appears, and find the category with the highest frequency, and use this category with the highest frequency as the final classification output result of the smooth classifier.

[0045] In step S103, the corresponding robust radius is derived based on the classification output results.

[0046] The robust radius refers to the maximum amount of disturbance that the input data can withstand without changing the model's prediction results at a given confidence level. In the embodiments of this application, it represents the maximum perturbation boundary that the classifier can maintain stable classification decisions when facing adversarial perturbations.

[0047] It is understood that, after determining the classification output result in the previous step, the embodiments of this application derive the corresponding robust radius based on the result, ensuring that the model can provide reliable classification results even when network traffic is subjected to a certain degree of adversarial disturbance.

[0048] In this embodiment of the application, the corresponding robust radius is derived based on the classification output result, including: using discrete smoothing mechanism and additive smoothing mechanism to derive the robust region of the smoothing classifier in at least one of discrete perturbation and additive perturbation, and determining the robust radius within the robust region.

[0049] Discrete smoothing mechanism refers to introducing perturbations by randomly selecting and retaining some data packets in the original traffic, simulating the effect of discrete perturbations (such as data packet insertion, deletion, or replacement); additive smoothing mechanism refers to introducing perturbations by adding random noise that conforms to an exponential distribution to the metadata (such as packet length, arrival time interval, etc.) of selected data packets, simulating the effect of additive perturbations (such as packet length padding or arrival time delay); robust region refers to the maximum range of perturbations within which the model can maintain its classification decision unchanged under a specific type of perturbation.

[0050] It is understood that embodiments of this application may use discrete smoothing mechanisms, additive smoothing mechanisms, or a combination of discrete and additive smoothing mechanisms to derive the robust region of the classifier, specifically:

[0051] stream collection Mapping to a collection of categories The robust region of the discrete smoothing mechanism can be derived as follows: there exists a pair of flows in By inserting n into the original stream x ins Packet, replace n sub Each package and delete n del The package is obtained; a smoothing function exists. It will randomly select d packets (d≤n) from stream x, while preserving the original order of these d packets. Let be a set of positive integers, representing the range of values ​​for the number d of data packets that can be selected. for A set of random variables on; smooth classifier g del According to the standard definition, assuming a smooth classifier g del Classification results for x and smooth classifier g del Classify x as y A probability p A ∈[1 / 2,1] satisfies g del (x)=y A And p A ≥ p a ≥1 / 2, then when When the conditions are met, then in It is a simplified form of the combinatorial formula n! / (r!(nr)!).

[0052] Similarly, the robust region of the additive smoothing mechanism can be derived as follows: there exists a pair of flows Where x contains n packets, and the packet length sequence is l = (l1, l2, ..., ln). nThe time interval sequence is t = (t1, t2, ..., t...). n ), This can be achieved by adding non-negative noise vectors to l and t. and Thus, it is obtained that a smooth function exists. in, Let be the set of positive real numbers, and let β be the value of β. l The value of β must be positive. for A set of random variables on; smoothing function ψ add (x,β l ,β t )to l i and t i Add random variables respectively (1≤i≤n) and Smooth classifier g add According to the standard definition, assuming and p A ∈[1 / 2,1] satisfies g add (x)=y A And p A ≥ p A ≥1 / 2, then when the condition is... When established, Where the robust radius r add =-(β) l +β t log 2(1- p A ).

[0053] It should be noted that, based on the two mechanisms mentioned above, the embodiments of this application can further derive the joint robust region after combining the discrete smoothing mechanism and the additive smoothing mechanism, and finally obtain the robust radius of the classifier. The joint robust radius can be derived using the following method:

[0054] There is a pair of streams Where x contains n packets, and the packet length sequence is l = (l1, l2, ..., ln). n The time interval sequence is t = (t1, t2, ..., t...). n ), This can be obtained in two steps: First, add non-negative perturbation vectors to l and t respectively. and There exists a smoothing function The function first randomly selects d packets from x while preserving their order, and then processes each selected packet x... i ,to l i and t iAdd random variables respectively (1≤i≤n) and Smooth classifier g jnt Based on the above standard definition, assuming and p A ∈[1 / 2,1] satisfies g jnt (x)=y A And p A ≥ p A ≥1 / 2, then the following condition holds true:

[0055]

[0056] but in and respectively by δ l and δ t according to The result is obtained by sorting in descending order.

[0057] Specifically, the packet sequence in stream x is first defined as X = {x1, x2, ..., x}. n},but The packets in stream x can be divided into the following three categories: the original packets in stream x It was derived from the replacement of some packages in XV. The package obtained through insertion The smoothing function ψ jnt (x,β l ,β t d) The set of all possible smoothed samples generated is defined as S. For each element s∈S in S, s can be obtained from x in the following two steps: randomly select d packets from x. At the same time, their order is preserved, as s′, for each selected package Towards and Add respectively by and Generated random noise; if Then we have: set Defined as S sel S can be sel Divided into two subsets By ψ jnt (x,β l ,β t ,d) and From the distribution, we can see that for all elements All of them are: And for all elements All of them are: Where i * yes In (x1, x2, ..., x n The serial number in ) is defined Similar to the derivation in additive smoothing, we can obtain The lower bound:

[0058]

[0059] definition assumed and They are respectively composed of δ l and δ t according to Arranged in descending order. Let K be the denoted K. Similar to the derivation in the discrete smoothing mechanism, we can obtain... The lower bound: Will With the above The lower bound, By linking the lower bounds together, we can obtain... lower bound Finally, by Substitute into the inequality This allows us to obtain the robust radius under the desired joint perturbation.

[0060] In the embodiments of this application, discrete perturbation includes perturbation applied to the flow at the packet level, and additive perturbation includes perturbation applied to the metadata of the packet at the packet level.

[0061] Discrete disturbances refer to disturbances caused by operations on data packets in network traffic. These disturbances typically involve operations such as inserting, deleting, or replacing data packets, for example, inserting extra data packets into a flow, deleting some data packets, or replacing some data packets with others; metadata refers to packet length, arrival time interval, etc. Additive disturbances refer to disturbances caused by modifications to the metadata of data packets at the data packet level, typically manifested as padding the length of data packets (adding extra bytes) or delaying the arrival time of data packets (adjusting timestamps).

[0062] The robustness authentication method for intelligent traffic analysis oriented towards multimodal adversarial disturbances proposed in this application involves randomly selecting data packet sequences multiple times from the original traffic data. After each random selection, smoothed samples and a set of smoothed samples are generated based on the selected data packet sequences. The smoothed samples in the set of smoothed samples are processed, and the processed set of smoothed samples is input into a smoothing classifier. The smoothing classifier outputs the classification results of the set of smoothed samples. Finally, the corresponding robust radius is derived based on the classification results. This method has low computational cost, requires no understanding or modification of the internal structure of the model, and can adapt to real multimodal attacks, providing better robustness assurance for traffic analysis systems. It also has higher model versatility and is widely applicable to models with different architectures and input formats.

[0063] The robustness certification method for intelligent traffic analysis against multimodal disturbances in this application will be further described below through a specific embodiment, such as... Figure 2 As shown, it includes:

[0064] Step 1: Randomly select packets from the raw traffic data. Select a portion of data packets from the raw flow and retain them. Introduce random noise that conforms to an exponential distribution into the packet length sequence and time interval sequence of the selected data packets to generate smooth samples.

[0065] Specifically, based on a stream x containing n data packets, d data packets are randomly selected from the n data packets, the other data packets are deleted, and exponentially distributed random noise is added to the packet length and time interval sequence δ,t of the remaining d data packets to obtain a smoothed sample s. The above operation is repeated to obtain a smoothed sample set S.

[0066] Step 2: Process the smoothed samples into the data format required by the base classifier, and then use the base classifier to classify the smoothed sample set.

[0067] For each smooth sample s in set S, it is preprocessed into the data form required by the base classifier f and the classification result f(s) is obtained. The classification results are statistically analyzed to obtain the set of classification results f(S) of the base classifier f on set S.

[0068] Step 3: Calculate the frequency of each category in the classification results, and take the dominant category with the highest frequency as the classification output of the smooth classifier. Combine the frequency of the dominant category to further derive the corresponding robust radius.

[0069] Specifically, the class that appears most frequently in the classification result f(S) of set S is taken as the classification output of the smooth classifier g, that is, g(x) = y. A ;

[0070] By calculating y AThe robust radius is derived from the proportion in f(S). The mathematical proof of robustness certification and the derivation of the robust radius are described later. The robust radius derived by this method is effective for discrete perturbations, additive perturbations and joint perturbations of the two.

[0071] To derive the robustness radius of the smooth classifier obtained through the above process under a specific input, it is first necessary to provide a formal mathematical description of this implementation method. Consider the stream set... The classification problem on the above, given a base classifier f, f will stream the set Mapping to a collection of categories Above. Let's assume... for Given a set of random variables, and a smoothing function This function is... The stream x generates a smoothed sample z. Based on the base classifier f and the smoothing function ψ, a smoothing classifier g can be obtained. g always returns the most likely classification result of the base classifier f for the smoothed sample ψ(x), i.e. Where, p y (x) is The abbreviation for represents the probability that the output of the base classifier f of the random variable z is class y, given that the random variable z follows a distribution defined by a smooth function ψ(x) with x as input. For input x, the classification result g(x) given by the smooth classifier is denoted as y. A The corresponding probability Noted as p A Since there is actually no way to know the true p A Therefore, the Monte Carlo method is used to estimate p with confidence level α. A lower bound p A As an alternative. p A Then, the robust region of the smooth classifier g with input x can be determined. Estimation is performed to determine the smooth classifier g for any input within the robust region R(x). The classification results are all the same as g(x), that is: This is the standard definition of the smooth classifier in this embodiment, and will not be repeated below. Record If Except for y A Treating all other classes as a single major category simplifies the multi-class classification problem into a binary classification problem. This yields the key conditions for proving robustness and deriving robust regions in the stochastic smoothing method.

[0072] Since this embodiment is aligned with the perturbation form of real adversarial attacks, it adopts a joint perturbation form of discrete perturbation and additive perturbation. Therefore, the robustness of discrete perturbation and additive perturbation will be derived separately in the following text, and the robust radius under joint perturbation will be obtained in the end.

[0073] Discrete smoothing mechanisms are implemented based on random packet selection. Given an input stream x with n packets, random packet selection randomly chooses d packets (d≤n) from these n packets, while preserving the original order of these d packets. Through discrete smoothing mechanisms, the robust region of the smoothing classifier against discrete perturbations (packet insertion, replacement, and deletion) can be obtained. The robust region of the discrete smoothing mechanism can be derived using the following lemma:

[0074] There exists a pair of streams in By inserting n into the original stream x ins Packet, replace n sub Each package and delete n del Each packet is obtained. A smoothing function exists. It will randomly select d packets (d≤n) from stream x, while preserving the original order of these d packets. Let be a set of positive integers, representing the range of values ​​for the number d of data packets that can be selected. for A set of random variables. A smooth classifier g. del Obtained according to the standard definition. Assume... and p A ∈[1 / 2,1] satisfies g del (x)=y A And p A ≥ p A ≥1 / 2, then when When the conditions are met, then in It is a simplified form of the combinatorial formula n! / (r!(nr)!).

[0075] Specifically, the packet sequence in stream x is first defined as X = {x1, x2, ..., x}. n},but The packets in stream x can be divided into the following three categories: the original packets in stream x It was derived from the replacement of some packages in XV. And the package obtained through insertion Now select nn from the original stream x sub -n del individual packages At the same time, their original order is preserved, denoted as v. It can be divided into two parts, S1 and S2, as follows: like Figure 3 As shown. (By ψ) sel (x,d) and The probability distribution can then be obtained: Will Let it be K. The lower bound can be obtained from the following inequality:

[0076]

[0077] Will Substitute into the inequality Then you can get This means that when When the conditions are met, then

[0078] Additive smoothing is achieved by adding exponentially distributed random noise to the metadata (such as packet length, time interval, etc.) of all data packets in the input x. Through additive smoothing, a robust region for the smoothing classifier against additive perturbations (packet padding, delay) can be obtained. The robust region of the additive smoothing mechanism can be derived using the following lemma:

[0079] There exists a pair of streams Where x contains n packets, and the packet length sequence is l = (l1, l2, ..., ln). n ), time interval sequence

[0080] Let t = (t1, t2, ..., t n ). This can be achieved by adding non-negative noise vectors to l and t. and Thus, a smooth function exists. in, Let be the set of positive real numbers, and let β be the value of β. l The value of β must be positive. for A set of random variables on the smoothing function ψ add (x,β l ,β t )to l i and t i Add random variables respectively (1≤i≤n) and Smooth classifier g add Obtained according to the standard definition. Assume... and p A ∈[1 / 2,1] satisfies g add (x)=y A And p A ≥ p A ≥1 / 2, then when the condition is... When established, Where the robust radius r add =-(β) l +β t log 2(1- p A ).

[0081] Specifically, firstly, the smoothing function ψ add (x,β l ,β t The set of all possible smooth samples is defined as S, where the sequence of packet lengths of streams s ∈ S is given by... The time interval sequence is The smoothed sample set S is now divided into two parts as follows: like Figure 4 As shown. From ψ add (x,β l ,β t )and The probability distribution can be obtained that for all elements s∈S1 in S1, we have: For all elements s∈S2 in S2, we have: Will Let it be K. The lower bound can be obtained from the following inequality:

[0082]

[0083] By Substitute into the inequality Then the solution can be obtained. That is when When the conditions are met, Where r add =-(β) l +β t log 2(1- p A ).

[0084] This embodiment uses both discrete smoothing and additive smoothing mechanisms. Therefore, based on the above derivation results, the joint robust region after combining the discrete smoothing and additive smoothing mechanisms can be further derived, ultimately obtaining the robust radius of the smoothing classifier given in this embodiment. The joint robust radius can be derived using the following theorem:

[0085] There is a pair of streams Where x contains n packets, and the packet length sequence is l = (l1, l2, ..., ln). n The time interval sequence is t = (t1, t2, ..., t...).n ). This can be obtained in two steps: First, add non-negative perturbation vectors to l and t respectively. and There exists a smoothing function The function first randomly selects d packets from x while preserving their order, and then processes each selected packet x... i ,to l i and t i Add random variables respectively (1≤i≤n) and Smooth classifier g jnt Obtained according to the standard definition. Assume... and p A ∈[1 / 2,1] satisfies g jnt (x)=y A And p A ≥ p a ≥1 / 2, then the following condition holds true:

[0086]

[0087] but in and respectively by δ l and δ t according to The result is obtained by sorting in descending order.

[0088] Specifically, the packet sequence in stream x is first defined as X = {x1, x2, ..., x}. n},but The packets in stream x can be divided into the following three categories: the original packets in stream x It was derived from the replacement of some packages in XV. The package obtained through insertion The smoothing function ψ jnt (x,β l ,β t d) The set of all possible smoothed samples generated is defined as S. For each element s∈S in S, s can be obtained from x in the following two steps: randomly select d packets from x. At the same time, their order is preserved, as s′, for each selected package Towards and Add respectively by and Generated random noise. If Then we have: set Defined as S selS can be sel Divided into two subsets like Figure 5 As shown. From ψ jnt (x,β l ,β t ,d) and From the distribution, we can see that for all elements All of them are: And for all elements All of them are: Where i * yes In (x1, x2, ..., x n The sequence number in ) is defined. Similar to the derivation in additive smoothing, we can obtain The lower bound:

[0089]

[0090] definition assumed and They are respectively composed of δ l and δ t according to Arranged in descending order. Let K be the denoted K. Similar to the derivation in the discrete smoothing mechanism, we can obtain... The lower bound: Will With the above The lower bound, By linking the lower bounds together, we can obtain... lower bound By Substitute into the inequality This allows us to obtain the robust radius under the desired joint perturbation.

[0091] Next, referring to the accompanying drawings, a robust authentication device for intelligent traffic analysis oriented against multimodal disturbances is described according to an embodiment of this application.

[0092] Figure 6 This is a block diagram of an intelligent traffic analysis robustness authentication device for multimodal disturbance resistance according to an embodiment of this application.

[0093] like Figure 6 As shown, the intelligent traffic analysis robustness certification device 10 for multimodal disturbance resistance includes: a selection module 201, a processing module 202, and a derivation module 203.

[0094] The selection module 201 randomly selects data packet sequences from the original traffic data multiple times. After each random selection, a smoothed sample is generated based on the selected data packet sequence, and a smoothed sample set is constructed based on the smoothed sample. The processing module 202 processes the smoothed samples in the smoothed sample set into the data format required by the base classifier, inputs the processed smoothed sample set into the smoothed classifier, and the smoothed classifier outputs the classification result of the smoothed sample set. The derivation module 203 derives the corresponding robust radius based on the classification output result.

[0095] In this embodiment, the selection module 201 is further configured to: obtain the packet length sequence and time interval sequence of the data packet sequence; determine random noise conforming to an exponential distribution based on the packet length sequence and time interval sequence; and add the random noise to the data packet sequence to obtain a smoothed sample.

[0096] In this embodiment of the application, the processing module 202 is further configured to: identify the category of each smoothed sample using a base classifier; generate a classification result set based on the category of each smoothed sample; and take the category that appears most frequently in the classification result set as the classification output result.

[0097] In this embodiment of the application, a calculation module is also included, wherein the calculation module is further configured to: calculate the frequency of occurrence of each category in the classification result set before using the category that appears most frequently in the classification result set as the classification output result; and determine the category that appears most frequently based on the frequency of occurrence of each category.

[0098] In this embodiment of the application, the derivation module 203 is further configured to: derive the robust region of the smoothing classifier in at least one of the discrete perturbation and the additive perturbation using the discrete smoothing mechanism and the additive smoothing mechanism, and determine the robust radius within the robust region.

[0099] In the embodiments of this application, discrete perturbation includes perturbation applied to the flow at the packet level, and additive perturbation includes perturbation applied to the metadata of the packet at the packet level.

[0100] It should be noted that the foregoing explanation of the embodiment of the intelligent traffic analysis robustness authentication method for multimodal adversarial disturbances also applies to the intelligent traffic analysis robustness authentication device for multimodal adversarial disturbances in this embodiment, and will not be repeated here.

[0101] The intelligent traffic analysis robustness authentication device for multimodal adversarial disturbances proposed in this application involves randomly selecting data packet sequences multiple times from the original traffic data. After each random selection, smoothed samples and a smoothed sample set are generated based on the selected data packet sequence. The smoothed samples in the smoothed sample set are processed, and the processed smoothed sample set is input into a smoothing classifier. The smoothing classifier outputs the classification result of the smoothed sample set, and finally, the corresponding robust radius is derived based on the classification output result. This device has low computational cost, requires no understanding or modification of the internal structure of the model, and can adapt to real multimodal attacks, providing better robustness assurance for the traffic analysis system. It also has higher model versatility and is widely applicable to models with different architectures and input formats.

[0102] Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include:

[0103] The memory 301, the processor 302, and the computer program stored on the memory 301 and capable of running on the processor 302.

[0104] When the processor 302 executes the program, it implements the robustness authentication method for intelligent traffic analysis against multimodal disturbances provided in the above embodiments.

[0105] Furthermore, electronic devices also include:

[0106] Communication interface 303 is used for communication between memory 301 and processor 302.

[0107] The memory 301 is used to store computer programs that can run on the processor 302.

[0108] The memory 301 may include high-speed RAM (Random Access Memory) memory, and may also include non-volatile memory, such as at least one disk storage.

[0109] If the memory 301, processor 302, and communication interface 303 are implemented independently, then the communication interface 303, memory 301, and processor 302 can be interconnected via a bus to complete communication between them. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 7 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0110] Optionally, in a specific implementation, if the memory 301, processor 302, and communication interface 303 are integrated on a single chip, then the memory 301, processor 302, and communication interface 303 can communicate with each other through an internal interface.

[0111] Processor 302 may be a CPU (Central Processing Unit), an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement embodiments of this application.

[0112] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for robust authentication of intelligent traffic analysis against multimodal disturbances.

[0113] This application also provides a computer program product, including a computer program or instructions, which, when executed, implement the above-described robust authentication method for intelligent traffic analysis against multimodal disturbances.

[0114] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0115] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0116] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0117] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any of the following techniques known in the art, or a combination thereof: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0118] Those skilled in the art will understand that all or part of the steps of the methods implementing the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0119] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A robustness verification method for intelligent traffic analysis oriented towards multimodal disturbances, characterized in that, Includes the following steps: The raw traffic data undergoes multiple random selections of data packet sequences. After each random selection, a smoothed sample is generated based on the selected data packet sequence. A smoothed sample set is then constructed based on the smoothed samples. The process of generating smoothed samples based on the selected data packet sequences includes: Obtain the packet length sequence and time interval sequence of the data packet sequence; Random noise conforming to an exponential distribution is determined based on the packet length sequence and the time interval sequence; The random noise is added to the data packet sequence to obtain the smoothed sample; The smoothed samples in the smoothed sample set are processed into the data format required by the base classifier. The processed smoothed sample set is then input into the smoothed classifier, and the smoothed classifier outputs the classification result of the smoothed sample set. The corresponding robust radius is derived based on the classification output results, including: By utilizing discrete smoothing and additive smoothing mechanisms, the robust region of the smoothing classifier in at least one of discrete perturbation and additive perturbation is derived, and the robust radius is determined within the robust region.

2. The robustness certification method for intelligent traffic analysis against multimodal disturbances according to claim 1, characterized in that, A smooth classifier is constructed from the base classifier, wherein the classification process of the smooth classifier includes: The base classifier is used to identify the category of each smoothed sample; A set of classification results is generated based on the category of each smoothed sample; The category that appears most frequently in the classification result set is taken as the classification output result.

3. The robustness certification method for intelligent traffic analysis against multimodal disturbances according to claim 2, characterized in that, Before using the category that appears most frequently in the classification result set as the classification output result, the following steps are included: Calculate the frequency of occurrence of each category in the classification result set; The most frequently occurring category is determined based on the frequency of occurrence of each category.

4. The robustness certification method for intelligent traffic analysis against multimodal disturbances according to claim 1, characterized in that, The discrete perturbation includes perturbations applied to the flow at the packet level, and the additive perturbation includes perturbations applied to the metadata of the packet at the packet level.

5. A robust authentication device for intelligent traffic analysis against multimodal disturbances, characterized in that, include: The selection module randomly selects data packet sequences from the raw traffic data multiple times. After each random selection, a smoothed sample is generated based on the selected data packet sequence. A smoothed sample set is then constructed based on the smoothed samples. The process of generating smoothed samples based on the selected data packet sequences includes: Obtain the packet length sequence and time interval sequence of the data packet sequence; Random noise conforming to an exponential distribution is determined based on the packet length sequence and the time interval sequence; The random noise is added to the data packet sequence to obtain the smoothed sample; The processing module is used to process the smoothed samples in the smoothed sample set into the data format required by the base classifier, input the processed smoothed sample set into the smoothed classifier, and the smoothed classifier outputs the classification result of the smoothed sample set. The derivation module is used to derive the corresponding robust radius based on the classification output result, wherein it includes: By utilizing discrete smoothing and additive smoothing mechanisms, the robust region of the smoothing classifier in at least one of discrete perturbation and additive perturbation is derived, and the robust radius is determined within the robust region.

6. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, the processor executing the program to implement the robust authentication method for intelligent traffic analysis against multimodal disturbances as described in any one of claims 1-4.

7. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed, they implement the robust authentication method for intelligent traffic analysis oriented against multimodal disturbances as described in any one of claims 1-4.

8. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed, they implement the robust authentication method for intelligent traffic analysis oriented against multimodal disturbances as described in any one of claims 1-4.

Images