A mimetic flow detection method for poisoning attacks in an internet of things environment

Abstract

The application belongs to the technical field of Internet of Things security, and discloses a mimicry flow detection method for poisoning attack in an Internet of Things environment. Based on the mimicry defense concept, a diversified abnormal flow detection framework is constructed. Through random combination of multiple feature extraction technologies and machine learning models, an abnormal flow detection executor with consistent functions but different structures is generated. A multi-mode decision mechanism and a dynamic scheduling strategy based on credibility evaluation are designed. The running state and the execution result of the executor are evaluated by means of a negative feedback mechanism. Based on the evaluation result, the executor with higher attack resistance and higher output result reliability is scheduled to execute the task, and the abnormal execution result is eliminated to reduce its interference with the decision, thereby effectively improving the attack resistance of the Internet of Things abnormal flow detection method and significantly improving the network security level.

Description

Technical Field

[0001] This invention belongs to the field of Internet of Things (IoT) security technology, specifically relating to a method for detecting mimicry traffic in the IoT environment in response to poisoning attacks. Background Technology

[0002] With the massive growth of IoT traffic, machine learning-based anomaly detection methods have become an important research direction in IoT security because they do not rely on prior knowledge in specialized fields. These methods typically include three steps: network traffic collection, traffic feature extraction, and traffic sample classification. They aim to identify abnormal data in IoT traffic, helping network administrators to promptly grasp network anomalies and take further measures to analyze the causes or attack intentions, thereby ensuring stable network operation. However, machine learning models are highly complex and sensitive to IoT traffic data, making them vulnerable to poisoning attacks. The core idea of ​​a poisoning attack is that attackers inject malicious data into the training set during the model training phase, interfering with or even destroying the model's convergence process, ultimately causing the trained model to fail.

[0003] Several security defenses against poisoning attacks have been proposed in existing technologies, but most of these strategies can only defend against a limited number of known attack types. Once attackers become aware of these defense strategies, they can easily circumvent detection. Furthermore, existing methods largely focus on protecting datasets or mitigating the impact of data poisoning on model training, lacking direct and effective protection for the model itself. At the same time, flaws or vulnerabilities are unavoidable in software and hardware design, allowing attackers to exploit other known or unknown vulnerabilities in the Internet of Things (IoT) to launch attacks, further exacerbating security problems. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a mimicry traffic detection method for poisoning attacks in an IoT environment. It introduces DVR features into traditional IoT abnormal traffic detection systems, and through a credibility assessment mechanism and dynamic scheduling strategy, ensures the dynamism within the IoT system, prevents attackers from gaining attack experience by exploiting software and hardware vulnerabilities, and improves the effectiveness and robustness of the system.

[0005] The mimicry traffic detection method for poisoning attacks in an IoT environment described in this invention includes the following steps: S1. Obtain raw IoT traffic data, extract features using the CICFowMeter tool for preprocessing, and distribute traffic detection requests to each traffic detection execution entity in the execution state. S2. Construct an execution body component pool, which contains various components used to form a flow detection execution body. These components are scheduled and combined by the execution mechanism into non-similar redundant execution bodies with the same function. S3. Construct an execution pool, which contains multiple dissimilar redundant execution bodies composed of different combinations of components selected from the execution body component pool by a negative feedback mechanism, and selects execution bodies from the pool to perform traffic detection tasks according to a scheduling strategy. S4. Multi-mode adjudication based on credibility assessment: The credibility index of the executor is used to assess the credibility of the execution results of each abnormal traffic detection executor. The execution results of executors with credibility below the preset threshold are removed and do not participate in multi-mode adjudication. The error information is fed back to the negative feedback mechanism. Then, multi-mode adjudication is performed on the remaining executors based on the historical confidence index of the executor, and the system adjudication result is output. S5. The negative feedback mechanism updates the scheduling strategy based on the feedback information from the multi-mode adjudication mechanism, and performs cleaning, reorganization and restoration operations on the execution entities with low credibility through the dynamic scheduling strategy. S6. Return the traffic detection results after the ruling to the network traffic device.

[0006] Furthermore, in S1, features extracted using the CICFowMeter tool include: source IP, destination IP, destination port, transmission protocol, single packet length, total number of packets in a single stream, payload size of data packets, frequency of sessions with the same IP, packet transmission rate per unit time, number of retransmissions, and number of timeout connections. The extracted features are then normalized.

[0007] Furthermore, S4 specifically refers to: S41. Calculate the credibility of the executor for pre-screening before multi-mode adjudication. Executors with credibility below a preset threshold are removed and not included in multi-mode adjudication. , in, Indicates the credibility of the execution entity; ACC represents the machine learning-based confusion matrix. Indicates accuracy. Indicates recall rate, , , These represent the weights of ACC, P, and R, respectively. S42. Calculate the historical confidence index of the executor for multi-mode adjudication of the remaining executors: , in, Indicating the first in the execution mechanism Historical confidence index of an executive entity This represents the sum of the enforcement results of this round of rulings by the enforcement body. Indicates the execution body In the first round of rulings Execution result.

[0008] Furthermore, in S4, the multi-mode decision specifically includes: (1) If the number of executors entering the multi-mode adjudication is greater than or equal to three and their execution results are the same, then the result is selected as the system adjudication result; (2) If the number of executors entering the multi-mode adjudication is greater than or equal to three, and their execution results are not exactly the same, then the result with the most common result shall be selected as the system adjudication result; (3) If the number of executors entering the multi-mode adjudication is greater than or equal to three, and their execution results are all different, then the result with the higher historical confidence index shall be selected as the system adjudication result. (4) If the number of executors entering the multi-mode decision is equal to two, then the result with the higher historical confidence index is selected as the system decision result; (5) If the number of executors entering the multi-mode decision is one, then the result is selected as the system decision result.

[0009] Furthermore, in S4, multi-mode adjudication also includes: When all executors are removed due to their credibility being below the threshold, the system is deemed to have been successfully attacked, and no data enters the multi-mode adjudication process; when no executors are removed, the multi-mode adjudication mechanism based on credibility assessment is equivalent to the traditional multi-mode adjudication mechanism.

[0010] Furthermore, S5 specifically refers to: S51. Perform system initialization and determine system redundancy. and the execution time limit for each executor. The execution mechanism randomly selects functional components from the execution body component pool, randomly combines them into an abnormal traffic detection execution body, stores it in the execution pool, and sets it to the ready state. S52, Assumption At any given time, when the system receives an abnormal traffic detection task processing request, the negative feedback mechanism will generate a scheduling strategy based on indicators such as heterogeneity, and select an execution entity from the execution pool. An execution set is formed by several abnormal traffic detection executors and set to execution status; the network traffic device then segments the traffic detection task and distributes it to the selected executors for execution. S53. After each round of execution, the system sends the execution results of each abnormal traffic detection execution entity to the adjudication mechanism. The adjudication mechanism evaluates the credibility of the execution results and makes multi-mode adjudication, and sends the adjudication results to the output module and the negative feedback mechanism. S54. After each round of adjudication, the negative feedback mechanism generates feedback information based on the received adjudication results to guide the next round of task scheduling. The negative feedback mechanism will remove, clean, and reorganize executors with a credibility below the threshold during the adjudication phase, and select the same number of ready executors from the executor pool to execute the next round of traffic detection tasks. If there are no executors with a credibility below the threshold during the adjudication phase, the negative feedback mechanism will calculate the cumulative execution time of each executor. When the execution time of execution body i Reaching the system's set execution launch time When this happens, the negative feedback mechanism will take executor i offline to the executor pool, put it into the ready state, clear the execution information, and select a new executor to replace it in performing the task.

[0011] The beneficial effects of this invention are as follows: (1) This invention breaks through the limitations of traditional traffic detection relying on fixed feature extraction schemes and single machine learning models. Based on the idea of ​​dynamic and diverse redundancy of DVR, it constructs a differentiated traffic detection execution body through random coupling of multiple feature extraction algorithms and heterogeneous machine learning models. Relying on the dynamic heterogeneous characteristics of detection architecture and modeling logic, it destroys the attacker's prior cognitive conditions, improves the randomness and uncertainty of the detection system, and raises the threshold for data poisoning and model poisoning attacks from the mechanism. (2) This invention introduces a quantitative evaluation of the credibility of the execution entity and a pre-screening mechanism for abnormal outputs on the basis of the traditional multi-mode voting framework. Before the multi-mode decision, the erroneous outputs of low-credibility execution entities are eliminated, the voting bias caused by poisoned samples is avoided, the accuracy of the decision results is optimized, and the fault tolerance performance and operational robustness of the detection system are effectively improved. Attached Figure Description

[0012] Figure 1 This is a diagram of the mimicry traffic detection architecture for poisoning attacks in the Internet of Things environment of the present invention; Figure 2 This is a flowchart of the method described in this invention; Figure 3 This is a flowchart of the dynamic scheduling strategy of the present invention; Figure 4 This is a schematic diagram illustrating the system effectiveness of the present invention under tag-flipping attacks; Figure 5 This is a schematic diagram illustrating the system effectiveness of the present invention under gradient optimization attacks; Figure 6 This is a schematic diagram comparing the robustness of the system in a tag-flipping attack scenario according to the present invention; Figure 7 This is a schematic diagram illustrating the robustness comparison of the gradient optimization attack scenario system of the present invention. Detailed Implementation

[0013] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0014] like Figure 1 As shown, this invention designs a mimicry traffic detection architecture for poisoning attacks in an IoT environment, mainly including input / output, execution module, data preprocessing module, execution body component pool, adjudication module, and negative feedback mechanism. The input / output module is connected to network traffic devices. The execution module selects appropriate traffic detection tasks from the execution body pool according to different scheduling strategies. The execution body component pool includes three categories: multi-feature extraction component pool, multi-training model component pool, and basic component pool, covering various components constituting the traffic detection execution body. For example, the multi-feature extraction component pool contains Pearson correlation coefficient, tree-based feature selection, recursive feature elimination, and other feature extraction algorithms. The adjudication module adopts a multi-mode adjudication mechanism based on credibility assessment, evaluating the credibility of the execution result from both the reliability of the execution body and the execution effect. The negative feedback mechanism changes the scheduling strategy based on the feedback information from the multi-mode adjudication mechanism, performing cleaning, reorganization, and restoration operations on execution bodies with low credibility.

[0015] Within the execution entity component pool, the multi-feature extraction component pool utilizes the CICFlowMeter tool to parse raw network traffic, automatically extracting over 80 dimensions of traffic statistical features, including flow duration, packet length statistics, flow byte rate, and TCP flag count. Building upon this, to further enhance feature effectiveness and construct heterogeneous feature subsets, the component pool integrates multiple feature extraction algorithms, including Pearson correlation coefficient, tree-based feature selection, and recursive feature elimination, evaluating feature effectiveness from different dimensions and providing a differentiated feature space for subsequent construction of heterogeneous execution entities.

[0016] like Figure 2 As shown, the mimicry traffic detection method for poisoning attacks in the Internet of Things environment described in this invention includes the following steps: S1. IoT Traffic Input: Raw IoT traffic data from network traffic devices is preprocessed, and the CICFlowMeter tool is used to obtain the characteristics of the raw traffic data, improving data validity. Then, traffic detection requests are distributed to various traffic detection execution units within the execution module. S2. Execution Pool Construction: The execution pool contains various non-similar redundant executions. These executions are composed of different components selected from the execution component pool by a negative feedback mechanism. When a traffic detection task arrives, the execution module selects a suitable execution traffic detection task from the execution pool according to different scheduling strategies. S3. Execution Component Pool Construction: The execution component pool contains various components that constitute the flow detection execution body. These components are scheduled and combined by the execution mechanism into non-dissimilar redundant execution bodies with the same function; the heterogeneity of the non-dissimilar redundant execution bodies largely ensures the randomness and unpredictability of the system, thereby hindering attackers from correctly understanding the system and expanding the system's defense surface; S4. Multi-mode adjudication based on credibility assessment: This step assesses the credibility of the execution results from two aspects: the reliability of the executor and the execution effect. First, the credibility of the execution results of each abnormal traffic detection executor is assessed using the executor credibility index, and an executor credibility threshold is set. The credibility index of the executor is compared with the threshold. When the credibility is less than the threshold, the adjudication mechanism considers that the abnormal traffic detection executor that produced the execution result has encountered an unknown error or suffered a poisoning attack. In this case, the execution result of the executor needs to be removed, not participate in the multi-mode adjudication, and the error information is fed back to the negative feedback mechanism.

[0017] After the adjudication mechanism removes all executors with credibility less than the threshold from the adjudication, the remaining executors are adjudicated in a multi-mode manner based on the historical confidence index of the executor. The specific adjudication is as follows: (1) If the number of executors entering the multi-mode adjudication is greater than or equal to three and their execution results are the same, then the result is selected as the system adjudication result; (2) If the number of executors entering the multi-mode adjudication is greater than or equal to three and their execution results are not exactly the same, then the result with the most similar results is selected as the system adjudication result; (3) If the number of executors entering the multi-mode adjudication is greater than or equal to three and their execution results are all different, then the result with the higher historical confidence index is selected as the system adjudication result; (4) If the number of executors entering the multi-mode adjudication is equal to two, regardless of whether their execution results are the same, then the result with the higher historical confidence index is selected as the system adjudication result; (5) If there is only one executor entering the multi-mode adjudication, then the result is selected as the system adjudication result.

[0018] However, several special cases should be noted: (1) If all executors are eliminated when the credibility assessment of the executors is carried out, then no data will enter the multi-mode adjudication, which means that the system has been successfully attacked; (2) If all executors are eliminated when the credibility assessment of the executors is carried out, then the multi-mode adjudication mechanism based on credibility assessment is equivalent to the traditional multi-mode adjudication.

[0019] Step S4 includes the following sub-steps S41-S42, wherein the historical confidence index calculated in S41 is used for the voting rules in multi-mode adjudication, and the executive credibility calculated in S42 is used for the pre-screening step (i.e., credibility assessment and elimination) before multi-mode adjudication.

[0020] S41. Calculate the historical confidence index of the executor. The historical confidence index of an executor reflects its operational stability and is a crucial basis for the DHR system's strategy scheduling and adjudication. Initially, the historical confidence index of each executor in the execution mechanism is zero. After each round of adjudication, the execution results of each executor are evaluated for anomalies to obtain its historical confidence index. The specific calculation formula is as follows: (1) in, Indicating the first in the execution mechanism Historical confidence index of an executive entity This represents the sum of the enforcement results of this round of rulings by the enforcement body. Indicates the execution body In the first round of rulings The execution result is as follows. The higher the historical confidence index of the executor, the fewer random failures or attacks it has received, and the greater the significance of the result for abnormal traffic detection.

[0021] S42. Calculate the credibility of the execution entity. The credibility of the execution entity is an important standard for determining whether its execution result is recognized by the adjudication mechanism. In the scenario of abnormal traffic classification, the most critical issue is whether machine learning can accurately and efficiently identify abnormal traffic. Various model evaluation metrics have been established in the industry to evaluate trained models. This method combines model evaluation metrics from machine learning to measure the credibility of each abnormal traffic detection execution entity.

[0022] Accuracy, the most common and easily understood evaluation metric in machine learning, represents the ratio of correct model predictions to the total number of correct predictions. The confusion matrix, based on machine learning, is calculated using the following formula: Similarly, precision. Recall is the ratio of the number of samples predicted as positive to the number of positive samples in the actual sample. F1 score represents the ratio of the number of positive examples in the actual sample to the number of samples predicted as positive. F1 score can be considered as accuracy. and recall rate The harmonic mean, .

[0023] This method utilizes the Technique for Order Performance by Similarity to Ideal Solution (TOPSIS) algorithm to calculate the weight allocation of the aforementioned indicators. First, the raw data for these indicators is positively oriented to unify the evaluation direction. Then, the sum of squares normalization method is used to standardize each indicator, ensuring the elimination of the influence of differences in indicator dimensions and orders of magnitude. The relative similarity of the aforementioned indicators is defined as follows: , , The weights of the three indicators are obtained according to this ratio. (2) (3) (4) The formula for calculating the credibility of the final executor is as follows: (5)

[0024] S5. Negative Feedback Mechanism. The negative feedback mechanism is the control center of the IoT abnormal traffic detection architecture. This mechanism changes the scheduling strategy based on feedback information from the multi-mode adjudication mechanism, and performs cleaning, reorganization, and restoration operations on executors with low reliability through dynamic scheduling strategies. Regarding step S5, as... Figure 3 As shown, it includes the following sub-steps S51-S54: S51. Perform system initialization and determine system redundancy. and the execution time limit for each executor. The execution mechanism randomly selects functional components from the execution body component pool, randomly combines them into an abnormal traffic detection execution body, stores it in the execution pool, and sets it to the ready state. S52, Assumption At any given time, when the system receives an abnormal traffic detection task processing request, the negative feedback mechanism will generate a scheduling strategy based on indicators such as heterogeneity, and select an execution entity from the execution pool. The abnormal traffic detection executors form an execution set and are set to execution. The input module then segments the traffic detection task and distributes it to the selected executors for execution. S53. After each round of execution, the system sends the execution results of each abnormal traffic detection execution entity to the adjudication module. The adjudication mechanism evaluates the credibility of the execution results and makes multi-mode adjudication, and sends the adjudication results to the output module and the negative feedback mechanism. S54. After each round of adjudication, the negative feedback mechanism generates feedback information based on the received adjudication results to guide the next round of task scheduling. The negative feedback mechanism will remove, clean, and reorganize executors with a credibility below the threshold during the adjudication phase, and select the same number of ready executors from the executor pool to execute the next round of traffic detection tasks. If there are no executors with a credibility below the threshold during the adjudication phase, the negative feedback mechanism will calculate the cumulative execution time of each executor. When the execution time of execution body i Reaching the system's set execution launch time When this happens, the negative feedback mechanism will take executor i offline to the executor pool, put it into the ready state, clear the execution information, and select a new executor to replace it in performing the task; S6. Output the detection results and return the determined traffic detection results to the network traffic device.

[0025] To verify the effectiveness and security of the method of this invention, the present invention uses real network traffic in the network system as the object of detection to construct a mimicry network traffic detection system. This system includes multiple routers, switches, and network traffic detection devices. In the experiment, these devices will be treated as heterogeneous entities, reflecting the heterogeneity of the mimicry defense architecture.

[0026] The dataset for this invention originates from a portion of the data (a total of 18.36 million raw network traffic data entries) exported from the Colasoft backtracking analysis system of China Southern Power Grid during the period of March 23-25, 2023. This dataset is labeled with normal and abnormal traffic, containing 81 features, including IP address, geographic location, and port number. The network dataset was processed using CICFlowMeter to obtain 60,000 data entries, with a training set to test set ratio of 8:2. The training set contains 48,000 traffic data entries, including 36,000 normal traffic entries and 12,000 abnormal traffic entries; the test set contains 12,000 traffic data entries, including 9,600 normal traffic entries and 2,400 abnormal traffic entries. The experiment of this invention uses an Alibaba Cloud server to simulate a network environment and deploy the traffic detection architecture described in this paper. The platform parameters are ecs.e-c1m4.xlarge (4 vCPUs, 16 GiB, economy type e). The terminal is simulated through a local host with parameters of Intel Core i7 3.5 GHz, 8GB RAM, and 64-bit Ubuntu 20.04 operating system.

[0027] To verify that the solution of this invention can effectively protect against poisoning attacks, this invention simulates two poisoning attack methods: data poisoning and model poisoning, namely label flipping attack and gradient optimization attack.

[0028] Label flipping attack: An attacker changes the labels of certain samples in the training dataset. After compromising a traffic detection execution unit, the attacker changes the labels of some abnormal traffic in the unit's training set to normal, making the model unable to accurately identify abnormal traffic after training.

[0029] Gradient optimization attack: An attacker modifies the gradients during model training, causing the model parameters to update in the wrong direction. For example, after compromising a traffic detection execution unit, the attacker changes the parameters during model training, causing the model to converge in the wrong direction.

[0030] Figure 4 and Figure 5 The system effectiveness of this invention was compared with that of PRAD, CTP, and De-Pois under two different poisoning attack methods. Figure 4 It can be seen that the accuracy of these four methods in detecting abnormal traffic under label-flipping attacks is all above 94.6%, indicating that all four methods can effectively counter label-flipping attacks. Furthermore, the method of this invention is only slightly inferior to PRAD, but the effectiveness indicators of each system are all above 98.6%. However, from... Figure 5 It can be seen that the performance of PRAD, CTP, and De-Pois schemes under gradient optimization attacks is significantly lower than that of label flipping attacks. This indicates that the above schemes only focus on protecting the dataset and reducing the impact of outlier data on model training, but are relatively lacking in the ability to protect model security. The scheme proposed in this paper only shows a slight decrease, which fully demonstrates the effectiveness of the method of this invention in dealing with gradient optimization attacks and effectively reduces the threat posed by attackers directly compromising the training model.

[0031] System robustness measures the degree to which a system successfully resists poisoning attacks; from the attacker's perspective, it's the probability of the system not failing. Therefore, the robustness of this invention is calculated based on the system failure rate. When the detection result of the invented method for traffic does not match the category of the original traffic, it means the system has failed. The ratio of the number of failed traffic detections to the total number of traffic is the system failure rate, i.e., the ratio of false negatives (FN) to all traffic in the experimental dataset. The system failure rate also reflects the probability that the system cannot correctly complete the traffic detection task when attacked or when a known or unknown error occurs. The lower the system failure rate, the higher the system robustness, which also reflects the attack resistance performance and reliability of the system of this invention. Furthermore, in reality, attackers often launch multiple attacks and obtain feedback information to strengthen their next attack. Therefore, this invention sets the robustness index as... They conducted multiple attacks to simulate real-world attack scenarios. Figure 6 and Figure 7 The system robustness of the four schemes described above is demonstrated under different attack scenarios.

[0032] from Figure 6 and Figure 7 It can be seen that, regardless of the attack scenario, the robustness of PRAD, CTP, and De-Pois schemes decreases rapidly with the increase of the cumulative number of attacks. Furthermore, by the fourth attack, the robustness of these three schemes is almost below 0.1, approaching zero. This is because attackers gain feedback based on the training model's results within the first three attacks and accumulate experience that can enhance subsequent attacks, thereby improving attack efficiency and accuracy. In contrast, the robustness of the method in this invention remains stable at around 93% under any attack scenario. This is because the adjudication module and credibility assessment mechanism of this invention promptly detect the attacked execution unit and replace and clean up the infected execution unit through negative feedback and scheduling mechanisms, ensuring that there are always normal execution units running within the system at any given time. This internal dynamic nature of the system keeps the system a black box to attackers, preventing them from gaining experience and significantly enhancing the system's security and reliability.

[0033] The above description is merely a preferred embodiment of the present invention and is not intended to further limit the present invention. All equivalent changes made based on the description and drawings of the present invention are within the protection scope of the present invention.

Claims

1. A method for detecting mimicry traffic in an Internet of Things (IoT) environment in response to poisoning attacks, characterized in that, Includes the following steps: S1. Obtain raw IoT traffic data, extract features using the CICFowMeter tool for preprocessing, and distribute traffic detection requests to each traffic detection execution entity in the execution state. S2. Construct an execution body component pool, which contains various components used to form a flow detection execution body. These components are scheduled and combined by the execution mechanism into non-similar redundant execution bodies with the same function. S3. Construct an execution pool, which contains multiple dissimilar redundant execution bodies composed of different combinations of components selected from the execution body component pool by a negative feedback mechanism, and selects execution bodies from the pool to perform traffic detection tasks according to a scheduling strategy. S4. Multi-mode adjudication based on credibility assessment: The credibility index of the executor is used to assess the credibility of the execution results of each abnormal traffic detection executor. The execution results of executors with credibility below the preset threshold are removed and do not participate in multi-mode adjudication. The error information is fed back to the negative feedback mechanism. Then, multi-mode adjudication is performed on the remaining executors based on the historical confidence index of the executor, and the system adjudication result is output. S5. The negative feedback mechanism updates the scheduling strategy based on the feedback information from the multi-mode adjudication mechanism, and performs cleaning, reorganization and restoration operations on the execution entities with low credibility through the dynamic scheduling strategy. S6. Return the traffic detection results after the ruling to the network traffic device.

2. The method for detecting mimicry traffic in an IoT environment resistant to poisoning attacks according to claim 1, characterized in that, In S1, the features extracted using the CICFowMeter tool include: source IP, destination IP, destination port, transmission protocol, single packet length, total number of packets in a single stream, payload size of data packets, frequency of sessions with the same IP, packet transmission rate per unit time, number of retransmissions, and number of timeout connections; and the extracted features are normalized.

3. The method for detecting mimicry traffic in an IoT environment resistant to poisoning attacks according to claim 1, characterized in that, S4 specifically refers to: S41. Calculate the credibility of the executor for pre-screening before multi-mode adjudication. Executors with credibility below a preset threshold are removed and not included in multi-mode adjudication. ， in, Indicates the credibility of the execution entity; ACC represents the machine learning-based confusion matrix. Indicates accuracy. Indicates recall rate, , , These represent the weights of ACC, P, and R, respectively. S42. Calculate the historical confidence index of the executor for multi-mode adjudication of the remaining executors: ， in, Indicating the first in the execution mechanism Historical confidence index of an executive entity This represents the sum of the enforcement results of this round of rulings by the enforcement body. Indicates the execution body In the first round of rulings Execution result.

4. The mimicry traffic detection method for poisoning attacks in an IoT environment according to claim 1, characterized in that, In S4, the multi-mode decision includes: (1) If the number of executors entering the multi-mode adjudication is greater than or equal to three and their execution results are the same, then the result is selected as the system adjudication result; (2) If the number of executors entering the multi-mode adjudication is greater than or equal to three, and their execution results are not completely the same, then the result with the most identical results shall be selected as the system adjudication result; (3) If the number of executors entering the multi-mode adjudication is greater than or equal to three, and their execution results are all different, then the result with the higher historical confidence index is selected as the system adjudication result; (4) If the number of executors entering the multi-mode decision is equal to two, then the result with the higher historical confidence index is selected as the system decision result; (5) If the number of executors entering the multi-mode decision is one, then the result is selected as the system decision result.

5. The method for detecting mimicry traffic in an IoT environment resistant to poisoning attacks according to claim 1, characterized in that, In S4, multi-mode adjudication also includes: When all executors are removed due to their credibility being below the threshold, the system is deemed to have been successfully attacked, and no data enters the multi-mode adjudication process; when no executors are removed, the multi-mode adjudication mechanism based on credibility assessment is equivalent to the traditional multi-mode adjudication mechanism.

6. The method for detecting mimicry traffic in an IoT environment resistant to poisoning attacks according to claim 1, characterized in that, S5 specifically refers to: S51. Perform system initialization and determine system redundancy. and the execution time limit for each executor. The execution mechanism randomly selects functional components from the execution body component pool, randomly combines them into an abnormal traffic detection execution body, stores it in the execution pool, and sets it to the ready state. S52, Assumption At any given time, when the system receives an abnormal traffic detection task processing request, the negative feedback mechanism will generate a scheduling strategy based on heterogeneity and historical confidence index, and select an execution entity from the execution pool. An execution set is formed by several abnormal traffic detection executors and set to execution state; the network traffic device traffic detection task is fragmented and distributed to the selected executors for execution; S53. After each round of execution, the system sends the execution results of each abnormal traffic detection execution entity to the adjudication mechanism. The adjudication mechanism evaluates the credibility of the execution results and makes multi-mode adjudication, and sends the adjudication results to the output module and the negative feedback mechanism. S54. After each round of adjudication, the negative feedback mechanism will generate feedback information based on the received adjudication results to guide the next round of task scheduling. The negative feedback mechanism will remove, clean up, and reorganize the executors with credibility below the threshold during the adjudication phase from the execution mechanism, and select the same number of ready executors from the executor pool to execute the next round of traffic detection tasks. If no executor falls below the credibility threshold during the adjudication phase, the negative feedback mechanism will calculate the cumulative execution time for each executor. When the execution time of execution body i Reaching the system's set execution and launch time When this happens, the negative feedback mechanism will take executor i offline to the executor pool, put it into the ready state, clear the execution information, and select a new executor to replace it in performing the task.

Images