A method and system for ship identification using multi-channel communication

Abstract

This invention relates to the field of communication data processing, specifically to a method and system for ship identification in multi-channel communication. It includes: acquiring historical multi-channel communication data; determining batch camouflage verification data based on camouflage verification data; acquiring a batch ensemble classifier and a local ensemble classifier; determining first derived data and second derived data based on the batch camouflage verification data and normal verification data; determining the batch camouflage verification data with updated weights based on the first and second derived data; determining an overall ensemble classifier; acquiring target multi-channel communication data; determining target batch data from the batch camouflage verification data with updated weights based on the target feature vector; and determining the detection result corresponding to the target multi-channel communication data based on the target batch data and the overall ensemble classifier. This invention can improve the accuracy of ship identification.

Description

Technical Field

[0001] This invention relates to the field of communication data processing, and specifically to a method and system for ship identification in multi-channel communication. Background Technology

[0002] In ship identification, existing methods mainly rely on multi-channel information fusion for authentication. By integrating the inherent attributes (including information authority, update frequency, and transmission reliability) of different channels (such as AIS, S-AIS, LRIT, and 4G / 5G) and their output content (such as the consistency of location, trajectory, and static information), the credibility weight of each channel data is dynamically evaluated, and then the authentication result is obtained through weighted fusion.

[0003] However, this method faces significant challenges in practical applications. First, due to the high scenario dependence of channel performance (such as high accuracy of near-shore AIS and wide coverage of offshore S-AIS), existing static weight models are difficult to adaptively adjust, resulting in poor environmental adaptability. Second, as fraud methods continue to evolve, attackers have acquired the ability to simultaneously forge data across multiple channels, generating highly consistent fake data and rendering the consistency verification mechanism on which the system depends ineffective.

[0004] Therefore, based on the problems existing in the above-mentioned prior art, traditional static models are difficult to adapt to continuously changing fraud strategies, resulting in a decline in recognition performance. Summary of the Invention

[0005] This invention provides a ship identification method and system using multi-channel communication to solve existing problems.

[0006] The ship identification method using multi-channel communication of the present invention adopts the following technical solution:

[0007] One embodiment of the present invention provides a ship identification method for multi-channel communication, the method comprising the following steps:

[0008] Acquire historical multi-channel communication data, which includes normal verification data and spoofed verification data. The amount of normal verification data is greater than that of spoofed verification data, and the spoofed verification data is obtained by sorting the data according to the capture time of each spoofed verification data.

[0009] Determine the batch spoofing verification data based on the spoofing verification data;

[0010] The AdaBoost ensemble classifier is trained using batch spoofing verification data and normal verification data to obtain batch ensemble classifier and local ensemble classifier.

[0011] Based on the batch spoofing verification data and normal verification data, determine the first and second derived data;

[0012] Based on the first and second derived data, the data weights of each batch in the batch spoofing verification data are calculated using the batch ensemble classifier and the local ensemble classifier to obtain the batch spoofing verification data with updated weights.

[0013] The AdaBoost ensemble classifier is trained using normal validation data and batch spoofed validation data with updated weights to obtain the overall ensemble classifier.

[0014] Acquire target multi-channel communication data and calculate the target feature vector corresponding to the target multi-channel communication data. Determine the target batch data from the batch camouflage verification data after updating the weights based on the target feature vector.

[0015] Based on the target batch data and the overall integrated classifier, determine the detection results corresponding to the target multi-channel communication data.

[0016] Optionally, batch spoofing verification data is determined based on the spoofing verification data, specifically including:

[0017] Extract the feature vector of each data point in the fake verification data, and calculate the cosine similarity of the feature vectors of two adjacent fake verification data points in turn to obtain a similarity sequence.

[0018] Find the local minimum of the similarity sequence, and divide the similarity sequence at the position of the element of the local minimum to obtain similarity sequence segments;

[0019] Each element in the similarity sequence segment is replaced with two corresponding fake verification data, and each similarity sequence segment is deduplicated to obtain batch fake verification data. The number of batches in the batch fake verification data is the same as the number of segments in the similarity sequence segment.

[0020] Optionally, the AdaBoost ensemble classifier is trained based on batch spoofing verification data and normal verification data to obtain a batch ensemble classifier and a local ensemble classifier, specifically including:

[0021] Using normal verification data as comparison data, the smote sampling increment method is used to increment the data of each batch in the batch spoofing verification data to obtain batch increment spoofing verification data. The number of batches in the batch increment spoofing verification data is the same as the number of batches in the batch spoofing verification data, and the data volume of each batch in the batch increment spoofing verification data is the same as the data volume of the normal verification data.

[0022] Normal validation data is used as correct data, and each batch of data in the batch incremental spoofing validation data is used as the first abnormal data to train the AdaBoost ensemble classifier, so as to obtain the batch ensemble classifier corresponding to each batch of data in the batch incremental spoofing validation data.

[0023] The AdaBoost ensemble classifier is trained using batch spoofed verification data and normal verification data to obtain a local ensemble classifier.

[0024] Optionally, the AdaBoost ensemble classifier is trained based on batch spoofing verification data and normal verification data to obtain a local ensemble classifier, specifically including:

[0025] By sequentially grouping adjacent batches in the batch camouflage verification data, local data is obtained.

[0026] Using normal validation data as comparison data, the smote sampling increment method is used to increment the data of each group in the local data to obtain local increment data. The number of groups in the local increment data is equal to the number of batches of batch fake validation data minus 1, and the amount of data in each group in the local increment data is the same as the amount of data in the normal validation data.

[0027] Using normal validation data as correct data and each group of data in the local incremental data as the second abnormal data, the AdaBoost ensemble classifier is trained to obtain the local ensemble classifier corresponding to each group of data in the local incremental data.

[0028] Optionally, based on batch spoofing verification data and normal verification data, first derived data and second derived data are determined, specifically including:

[0029] The first derived data for each batch of batch spoofing verification data is obtained by subtracting the corresponding batch data from the data of each batch in the batch incremental spoofing verification data.

[0030] The incremental data of the Ath group in the local data is obtained by subtracting the data of the Ath group in the local incremental data. The adjacent batches in the batch spoofing verification data contained in the Ath group in the local data are the Xth batch and the (X+1)th batch, respectively.

[0031] The incremental data of group A is divided into the second derived data of batch X and the second derived data of batch X+1. The ratio of the data volume of the second derived data of batch X to the data volume of the second derived data of batch X+1 is equal to the ratio of the data volume of batch X to the data volume of batch X+1 in the batch spoofing verification data.

[0032] Optionally, based on the first derived data and the second derived data, a batch ensemble classifier and a local ensemble classifier are used to calculate the data weights of each batch in the batch spoofing verification data, resulting in batch spoofing verification data with updated weights, specifically including:

[0033] Input the data of the (M+1)th batch in the batch spoofing verification data into the batch ensemble classifier corresponding to the Mth batch, and output the classification data corresponding to the data of the (M+1)th batch through the batch ensemble classifier corresponding to the Mth batch;

[0034] Determine the correctly classified data from the classification data corresponding to the (M+1)th batch of data;

[0035] The ratio of the amount of data in the (M+1)th batch to the amount of correctly classified data is defined as the difference between the (M+1)th batch and the Mth batch.

[0036] Obtain the difference between the second batch and the first batch in the Nth local ensemble classifier;

[0037] The first batch and the second batch are determined from the local data corresponding to the Nth local ensemble classifier, wherein the capture time of the first batch is earlier than that of the second batch.

[0038] Determine the first batch misclassification rate of the first batch in the Nth local ensemble classifier and the second batch misclassification rate of the second batch in the Nth local ensemble classifier, respectively;

[0039] Based on the difference between the second batch and the first batch in the Nth local ensemble classifier, the misclassification rate of the first batch in the Nth local ensemble classifier, and the misclassification rate of the second batch in the Nth local ensemble classifier, the weights of the first batch data and the weights of the second batch data in the Nth local ensemble classifier are determined respectively.

[0040] Based on the weights of the first and second batches of data in each local ensemble classifier, the data weights of each batch of data in the batch masquerading verification data are determined, resulting in the batch masquerading verification data with updated weights.

[0041] Optionally, the AdaBoost ensemble classifier is trained using normal validation data and batch masquerading validation data with updated weights to obtain the overall ensemble classifier, specifically including:

[0042] Normal verification data is used as correct data. The batch fake verification data is incrementally processed using the smote sampling incremental method to obtain the overall incremental data. The amount of data in the overall incremental data is the same as the amount of data in the normal verification data.

[0043] The normal validation data is used as the correct data, the overall incremental data is used as the third abnormal data, and the data weights of each batch of the batch fake validation data after the weights are updated are used as the data weights of each batch in the third abnormal data. The AdaBoost ensemble classifier is trained to obtain the overall ensemble classifier.

[0044] Optionally, the target batch data is determined from the weighted batch camouflage verification data based on the target feature vector, specifically including:

[0045] Determine the representative data for each batch in the batch spoofing verification data after updating the weights;

[0046] Calculate the cosine similarity between the target feature vector and the representative data of each batch;

[0047] The representative data with the highest cosine similarity is selected as the target batch data.

[0048] Optionally, based on the target batch data and the overall ensemble classifier, the detection results corresponding to the target multi-channel communication data are determined, specifically including:

[0049] Subtract the batch spoofing verification data from the overall incremental data to obtain the incremental data of the batch spoofing verification data.

[0050] The incremental data of the batch spoofing verification data is divided into the third derived data of each batch in the batch spoofing verification data. The ratio of the data volume of the third derived data of each batch is equal to the ratio of the data volume of each batch in the batch spoofing verification data.

[0051] Obtain the difference data between the first derived data of each batch and the third derived data of the corresponding batch, and obtain the derived difference data of each batch;

[0052] The derived difference data of the Zth batch are input into each base classifier in the overall ensemble classifier, and the misclassification rate corresponding to the derived difference data of the Zth batch is output by each base classifier.

[0053] The base classifier with the smallest misclassification rate among the misclassification rates of the derived difference data of the Zth batch is determined as the recognition base classifier for the data of the Zth batch.

[0054] Obtain the latest target batch data corresponding to the target batch data, and obtain the spoofing success rate of the target batch data and the spoofing success rate of the latest target batch data;

[0055] When the spoofing success rate of the latest target batch data is less than the spoofing success rate of the target batch data, the target feature vector is input into the overall ensemble classifier, and the detection result of the target multi-channel communication data is output by the overall ensemble classifier.

[0056] When the spoofing success rate of the latest target batch data is greater than or equal to the spoofing success rate of the target batch data, the weights of the recognition base classifier corresponding to the target batch data are updated to obtain the final overall ensemble classifier.

[0057] The target feature vector is input into the final ensemble classifier, which then outputs the detection results of the target multi-channel communication data.

[0058] This invention proposes a multi-channel communication ship identification system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the multi-channel communication ship identification method described above.

[0059] The beneficial effects of the technical solution of the present invention are:

[0060] In this embodiment of the invention, historical camouflage data is clustered chronologically and divided into multiple categories based on feature similarity. This process can systematically identify typical camouflage patterns at different development stages. The clustered categories serve as prior knowledge for differential processing in subsequent weighted training and ensemble classification processes, thereby enhancing the model's ability to identify known and potential novel camouflage strategies. The performance of the dedicated classifier reflects its specific recognition capability, while the misclassification rate of the local classifier reflects the data's generalization ability across batch scenarios, thus comprehensively measuring the effectiveness of camouflage verification under different data distributions. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 A flowchart illustrating a ship identification method using multi-channel communication according to an embodiment of the present invention;

[0063] Figure 2 A diagram illustrating the partitioning of verification data to disguise its purpose;

[0064] Figure 3 A schematic diagram of derived data;

[0065] Figure 4 This is a structural diagram of a multi-channel communication ship identification system provided in one embodiment of the present invention. Detailed Implementation

[0066] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a multi-channel communication ship identification method proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0067] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0068] The following description, in conjunction with the accompanying drawings, details a specific scheme for a multi-channel communication ship identification method provided by the present invention.

[0069] This invention provides a ship identification method and system using multi-channel communication. Please refer to [link / reference]. Figure 1 The diagram illustrates a flowchart of a ship identification method for multi-channel communication according to an embodiment of the present invention, the method comprising the following steps:

[0070] S101. Obtain historical multi-channel communication data, which includes normal verification data and spoofed verification data. The amount of normal verification data is greater than that of spoofed verification data, and the spoofed verification data is obtained by sorting the data according to the capture time of each spoofed verification data.

[0071] S102. Determine the batch camouflage verification data based on the camouflage verification data.

[0072] In this embodiment, determining batch camouflage verification data based on camouflage verification data specifically includes:

[0073] Extract the feature vector of each data point in the fake verification data, and calculate the cosine similarity of the feature vectors of two adjacent fake verification data points in turn to obtain a similarity sequence.

[0074] Find the local minimum of the similarity sequence, and divide the similarity sequence at the position of the element of the local minimum to obtain similarity sequence segments;

[0075] Each element in the similarity sequence segment is replaced with two corresponding fake verification data, and each similarity sequence segment is deduplicated to obtain batch fake verification data. The number of batches in the batch fake verification data is the same as the number of segments in the similarity sequence segment.

[0076] For example, feature vectors characterizing abnormal ship identities are extracted from spoofing verification data such as AIS, S-AIS, and LRIT. A feature vector X is constructed for each sample (i.e., a specific ship) at the capture time, and a corresponding label Y is assigned. Here, Y1=0 indicates that the feature at Y1 is abnormal. 0 represents abnormality, and 1 represents normality. The label data originates from historically confirmed deception events or manual verification records. Features included in the label data may include:

[0077] I. Spatial consistency characteristics.

[0078] Location difference: The maximum Euclidean distance difference between reported locations on different channels;

[0079] Heading difference: The maximum difference in heading reported by different channels;

[0080] Speed ​​difference: The maximum difference in speed reported by different channels;

[0081] II. Characteristics of Time-Based Behavior.

[0082] Message update frequency: Is the recent AIS message update frequency abnormal, such as being significantly lower than the frequency it should be in motion?

[0083] Historical trajectory consistency: The residuals between the predicted position, heading, and speed based on historical trajectories and the current observations can be calculated using methods such as Kalman filtering;

[0084] III. Characteristics of Message Content

[0085] Validity of maritime mobile services: Does the MMSI number conform to the country code and vessel type code format specified by the ITU?

[0086] Changes in static information: Whether static information such as ship name, call sign, and destination has undergone abnormal changes in a short period of time;

[0087] IV. Channel-specific characteristics.

[0088] Signal strength: VHF signal reception strength; an abnormally high signal strength may indicate the presence of a spoofing device nearby.

[0089] Communication delay: Is there any abnormality in the response delay of the satellite channel, such as a timeout in response to a challenge code?

[0090] V. Contextual Features.

[0091] High-risk area identification: Whether it is located in a historically high-incidence area of ​​fraud or a sensitive area, such as a no-navigation zone or border waters;

[0092] Expected navigation indicator: Whether it is in the expected navigation plan, and if so, whether it has been notified in advance.

[0093] For the spoofed verification data, the feature vector of each spoofed verification data point at the capture time is obtained, and then the cosine similarity of two adjacent spoofed verification data points is obtained to form a similarity sequence. If two adjacent spoofed verification data points belong to the same spoofing mode, the cosine similarity of the corresponding two vectors is relatively high.

[0094] By finding multiple local minima in the similarity sequence, the similarity sequence is divided into multiple sequence segments, that is, the spoofing verification data is divided into multiple data segments. Spoofing verification data within the same data segment are similar and likely belong to the same spoofing pattern. For example... Figure 2 As shown, Figure 2 This diagram illustrates the partitioning of the spoofing verification data. The large rectangle formed by the four smaller rectangles represents all the spoofing verification data, and each smaller rectangle represents a batch of spoofing verification data. The data in the adjacent batches of each rectangle are recorded in the left and right rectangles.

[0095] The spoofing verification data in each data segment is recorded as the same batch of spoofing verification data. The data referred to as batches below are all the spoofing verification data of this batch.

[0096] S103. Train the AdaBoost ensemble classifier based on batch spoofing verification data and normal verification data to obtain the batch ensemble classifier and the local ensemble classifier.

[0097] In this embodiment, the AdaBoost ensemble classifier is trained based on batch spoofing verification data and normal verification data to obtain a batch ensemble classifier and a local ensemble classifier, specifically including:

[0098] Using normal verification data as comparison data, the smote sampling increment method is used to increment the data of each batch in the batch spoofing verification data to obtain batch increment spoofing verification data. The number of batches in the batch increment spoofing verification data is the same as the number of batches in the batch spoofing verification data, and the data volume of each batch in the batch increment spoofing verification data is the same as the data volume of the normal verification data.

[0099] Normal validation data is used as correct data, and each batch of data in the batch incremental spoofing validation data is used as the first abnormal data to train the AdaBoost ensemble classifier, so as to obtain the batch ensemble classifier corresponding to each batch of data in the batch incremental spoofing validation data.

[0100] The AdaBoost ensemble classifier is trained using batch spoofed verification data and normal verification data to obtain a local ensemble classifier.

[0101] The AdaBoost ensemble classifier is trained using batch spoofing verification data and normal verification data to obtain a local ensemble classifier, specifically including:

[0102] By sequentially grouping adjacent batches in the batch camouflage verification data, local data is obtained.

[0103] Using normal validation data as comparison data, the smote sampling increment method is used to increment the data of each group in the local data to obtain local increment data. The number of groups in the local increment data is equal to the number of batches of batch fake validation data minus 1, and the amount of data in each group in the local increment data is the same as the amount of data in the normal validation data.

[0104] Using normal validation data as correct data and each group of data in the local incremental data as the second abnormal data, the AdaBoost ensemble classifier is trained to obtain the local ensemble classifier corresponding to each group of data in the local incremental data.

[0105] For example, considering that the amount of normal verification data is large while the amount of fake verification data is small, the fake verification data is first incremented by the smote sampling increment method to obtain batch increment fake verification data.

[0106] Normal data and batch incremental spoofing verification data for each batch (here, the spoofing verification data for that batch after incremental processing) are used as training data. The corresponding AdaBoost ensemble classifier is obtained through training and is denoted as the batch ensemble classifier. The number of batch ensemble classifiers is the same as the number of batches of batch spoofing verification data.

[0107] Each batch ensemble classifier is trained using normal data and a batch of fake verification data, and has a high recognition ability for that batch of data.

[0108] By combining data from two adjacent batches to form local data, and using normal data and each local incremental data (here, local incremental data refers to local data that has already undergone incrementing) as training data, an ensemble classifier is trained, denoted as the local ensemble classifier.

[0109] Each local ensemble classifier's training data includes spoofed verification data from two adjacent batches.

[0110] In this embodiment, a dedicated ensemble classifier is first trained for each batch to achieve optimal detection for that batch of data; secondly, a local ensemble classifier is trained for each pair of adjacent batches. The performance of the dedicated classifier reflects its specific recognition capability, while the misclassification rate of the local classifier reflects the generalization capability of the data in cross-batch scenarios, thus comprehensively measuring the effectiveness of spoofing verification under different data distributions.

[0111] S104. Based on the batch spoofing verification data and normal verification data, determine the first derived data and the second derived data.

[0112] In this embodiment, the first derived data and the second derived data are determined based on the batch spoofing verification data and the normal verification data, specifically including:

[0113] The first derived data for each batch of batch spoofing verification data is obtained by subtracting the corresponding batch data from the data of each batch in the batch incremental spoofing verification data.

[0114] The incremental data of the Ath group in the local data is obtained by subtracting the data of the Ath group in the local incremental data. The adjacent batches in the batch spoofing verification data contained in the Ath group in the local data are the Xth batch and the (X+1)th batch, respectively.

[0115] The incremental data of group A is divided into the second derived data of batch X and the second derived data of batch X+1. The ratio of the data volume of the second derived data of batch X to the data volume of the second derived data of batch X+1 is equal to the ratio of the data volume of batch X to the data volume of batch X+1 in the batch spoofing verification data.

[0116] For example, when training a batch ensemble classifier, in order to obtain a data volume that is not much different from the normal data, the incremental data has the largest data volume, which is denoted as the first derived data of the batch data.

[0117] When training a local ensemble classifier, both batches of data are incremented. Therefore, the amount of incremental data in each batch is less than the amount of data in the first derived data of each batch, and is referred to as the second derived data of each batch.

[0118] like Figure 3 As shown, Figure 3This is a schematic diagram of the derived data. The left side shows the normal validation data and three batches of data, A, B, and C. The width of the rectangles indicates the amount of data, with the batch data clearly having a smaller amount of data. The middle side shows that when training the batch ensemble classifier, the data of batch A reaches a similar amount to the normal validation data through incremental changes. The right side shows that when training the local ensemble classifier, the data of batches A and B, through incremental changes, have a much larger amount of data than the normal validation data.

[0119] S105. Based on the first derived data and the second derived data, use the batch ensemble classifier and the local ensemble classifier to calculate the data weight of each batch in the batch masquerading verification data, and obtain the batch masquerading verification data with updated weights.

[0120] In this embodiment, based on the first derived data and the second derived data, a batch ensemble classifier and a local ensemble classifier are used to calculate the data weight of each batch in the batch spoofing verification data, resulting in batch spoofing verification data with updated weights. Specifically, this includes:

[0121] The data of the (M+1)th batch in the batch spoofing verification data is input into the batch ensemble classifier corresponding to the Mth batch, and the classification data corresponding to the data of the (M+1)th batch is output through the batch ensemble classifier corresponding to the Mth batch.

[0122] Determine the correctly classified data from the classification data corresponding to the (M+1)th batch of data;

[0123] The ratio of the data volume of the (M+1)th batch to the data volume of the correctly classified data is determined as the difference between the (M+1)th batch and the Mth batch.

[0124] Obtain the difference between the second batch and the first batch in the Nth local ensemble classifier;

[0125] The first batch and the second batch are determined from the local data corresponding to the Nth local ensemble classifier, wherein the capture time of the first batch is earlier than that of the second batch.

[0126] Determine the first batch misclassification rate of the first batch in the Nth local ensemble classifier and the second batch misclassification rate of the second batch in the Nth local ensemble classifier, respectively;

[0127] Based on the difference between the second batch and the first batch in the Nth local ensemble classifier, the misclassification rate of the first batch in the Nth local ensemble classifier, and the misclassification rate of the second batch in the Nth local ensemble classifier, the first batch data weights corresponding to the first batch data in the Nth local ensemble classifier and the second batch data weights corresponding to the second batch data in the Nth local ensemble classifier are determined respectively.

[0128] Based on the weights of the first and second batches of data in each local ensemble classifier, the data weights of each batch of data in the batch spoofing verification data are determined, resulting in the batch spoofing verification data with updated weights.

[0129] Specifically, the misclassification rates of the first batch in the first batch of the Nth local ensemble classifier and the second batch in the second batch of the Nth local ensemble classifier are determined, including:

[0130] Obtain the first batch of first-generation first-generation data and the first batch of second-generation second-generation data from the first-generation first-generation data and the second-generation second-generation data, respectively;

[0131] Obtain the second batch of first-derived data and the second batch of second-derived data from the first-derived data and the second-derived data, respectively;

[0132] Obtain the first batch of difference data between the first batch of first-derived data and the first batch of second-derived data;

[0133] Obtain the second batch of difference data between the first batch of first-derived data and the first batch of second-derived data;

[0134] The first batch of difference data and the second batch of difference data are input into the Nth local ensemble classifier to obtain the first batch of output data and the second batch of output data, respectively.

[0135] Determine the correct data from the first batch of output data;

[0136] Determine the correct data for the second batch from the output data of the second batch;

[0137] The ratio of the amount of correct data in the first batch to the amount of difference data in the first batch is determined as the first batch misclassification rate in the Nth local ensemble classifier.

[0138] The ratio of the amount of correct data in the second batch to the amount of discrepancy data in the second batch is determined as the second batch misclassification rate in the Nth local ensemble classifier.

[0139] Specifically, based on the difference between the second batch and the first batch in the Nth local ensemble classifier, the misclassification rate of the first batch in the Nth local ensemble classifier, and the misclassification rate of the second batch in the Nth local ensemble classifier, the first batch data weights corresponding to the first batch data in the Nth local ensemble classifier and the second batch data weights corresponding to the second batch data in the Nth local ensemble classifier are determined respectively, including:

[0140] When the first batch misclassification rate in the Nth local ensemble classifier is greater than the second batch misclassification rate, and when the first batch misclassification rate in the (N-1)th local ensemble classifier is greater than the second batch misclassification rate, calculate the misclassification rate difference value of the Nth local ensemble classifier. The misclassification rate difference value of the Nth local ensemble classifier is obtained by taking the absolute value of the difference between the first batch misclassification rate and the second batch misclassification rate in the Nth local ensemble classifier.

[0141] Calculate the normalized difference value of the Nth local ensemble classifier. The normalized difference value of the Nth local ensemble classifier is obtained by taking the negative difference of the second batch compared with the first batch in the Nth local ensemble classifier and using it as the exponent of the natural base.

[0142] The misclassification rate difference value of the Nth local ensemble classifier is normalized and multiplied by the difference normalization value of the Nth local ensemble classifier. The result of the multiplication is determined as the first batch data weight corresponding to the first batch of data in the Nth local ensemble classifier.

[0143] When the misclassification rate of the second batch in the Nth local ensemble classifier is greater than or equal to the misclassification rate of the first batch, and when the misclassification rate of the first batch in the (N+1)th local ensemble classifier is less than or equal to the misclassification rate of the second batch, the misclassification rate difference value of the Nth local ensemble classifier is normalized and multiplied by the normalized difference value of the Nth local ensemble classifier. The result of the multiplication is determined as the weight of the second batch of data corresponding to the second batch of data in the Nth local ensemble classifier.

[0144] When the misclassification rate of the second batch in the Nth local ensemble classifier is greater than or equal to the misclassification rate of the first batch, and when the misclassification rate of the first batch in the N+1th local ensemble classifier is greater than the misclassification rate of the second batch, obtain the difference between the second batch and the first batch in the N+1th local ensemble classifier.

[0145] Calculate the misclassification rate difference value of the N+1th local ensemble classifier. The misclassification rate difference value of the N+1th local ensemble classifier is obtained by taking the absolute value of the difference between the first batch misclassification rate and the second batch misclassification rate in the N+1th local ensemble classifier.

[0146] Calculate the normalized difference value of the N+1th local ensemble classifier. The normalized difference value of the N+1th local ensemble classifier is obtained by taking the negative difference of the second batch compared with the first batch in the N+1th local ensemble classifier and using it as the exponent of the natural base.

[0147] The misclassification rate difference value of the N+1th local ensemble classifier is normalized and multiplied by the difference normalization value of the N+1th local ensemble classifier. The result of the multiplication is determined as the first batch data weight corresponding to the first batch of data in the N+1th local ensemble classifier.

[0148] The weights of the second batch of data in the Nth local ensemble classifier and the weights of the first batch of data in the (N+1)th local ensemble classifier are summed to determine the weights of the second batch of data in the Nth local ensemble classifier.

[0149] For example, for each batch ensemble classifier, the data from the next batch can be input into the batch ensemble classifier corresponding to the previous batch. The output of the previous batch ensemble classifier shows the number, n, of data from the next batch that is identified as normal (actually disguised verification data). The ratio of n to the number of input data is used as the difference between the next batch and the previous batch. The greater the difference, the greater the difference between the disguised patterns of the adjacent batch and the batch data in the training data corresponding to the batch ensemble classifier, indicating that the batch ensemble classifier failed to identify the adjacent batch of data well.

[0150] For any one of the two batches corresponding to each local ensemble classifier, first obtain the difference data between the first and second derived data of that batch. Use this difference data as the input to the local ensemble classifier to obtain the output. The misclassification rate (the ratio of correctly classified data in the difference data to the input data) can be obtained from the output. This allows us to obtain the misclassification rate for each of the two batches corresponding to each local ensemble classifier.

[0151] If the misclassification rates of both batches of data are smaller, but the difference in recognition ability between the two batches of data by the local ensemble classifiers is greater, it indicates that the more difficult-to-identify batch of data has a higher disguise ability. When building the overall ensemble classifier, it is necessary to assign a larger initial weight to this batch of data, rather than assigning the same weight to all training data as in the conventional method of building an ensemble classifier. This will increase the overall ensemble classifier's attention to this batch of data, and ultimately enable the trained overall ensemble classifier to better identify data that uses the same disguise pattern as this batch of data.

[0152] For each local ensemble classifier, the batch with the larger misclassification rate is selected as the batch whose initial weight needs to be increased.

[0153] Let c represent the difference between the previous and subsequent batches for each local ensemble classifier, and let j represent the normalized value of the absolute value of the difference in misclassification rates between the two local ensemble classifiers. The results are used as weighting coefficients for batches of data with larger misclassification rates.

[0154] If a batch of data is a batch of data that needs to have its initial weight increased in both local ensemble classifiers, then the weight coefficients in these two local ensemble classifiers are added together to obtain the weight coefficient of that batch of data. Thus, the batch spoofing verification data with updated weights can be obtained.

[0155] S106. Train the AdaBoost ensemble classifier using normal validation data and batch spoofed validation data with updated weights to obtain the overall ensemble classifier.

[0156] In this embodiment, the AdaBoost ensemble classifier is trained based on normal verification data and batch spoofing verification data with updated weights to obtain the overall ensemble classifier, specifically including:

[0157] Normal verification data is used as correct data. The batch fake verification data is incrementally processed using the smote sampling incremental method to obtain the overall incremental data. The amount of data in the overall incremental data is the same as the amount of data in the normal verification data.

[0158] The normal validation data is used as the correct data, the overall incremental data is used as the third abnormal data, and the data weights of each batch of the batch fake validation data after the weights are updated are used as the data weights of each batch in the third abnormal data. The AdaBoost ensemble classifier is trained to obtain the overall ensemble classifier.

[0159] For example, when training the overall ensemble classifier, all batches of incremental data are involved, and the amount of incremental data in each batch is less than that of the first two classifiers.

[0160] S107. Obtain the target multi-channel communication data and calculate the target feature vector corresponding to the target multi-channel communication data. Determine the target batch data from the batch camouflage verification data after updating the weight based on the target feature vector.

[0161] In this embodiment, the target batch data is determined from the weighted batch spoofing verification data based on the target feature vector, specifically including:

[0162] Determine the representative data for each batch in the batch spoofing verification data after updating the weights;

[0163] Calculate the cosine similarity between the target feature vector and the representative data of each batch;

[0164] The representative data with the highest cosine similarity is selected as the target batch data.

[0165] The representative data for each batch in the batch spoofing verification data after the weights have been determined includes:

[0166] Calculate the cosine similarity between the Bth data item in the Ath batch of the batch spoofing verification data after the update weights and the other data items in the Ath batch, and sum them.

[0167] The data in batch A that has the highest sum of cosine similarity to the other data in batch A is determined as the representative data of batch A.

[0168] For example, for the target multi-channel communication data of the ship that needs to be verified, which is currently received, the target feature vector corresponding to the target multi-channel communication data is first calculated. The similarity between each data point in each batch and other data points is calculated, and the data with the highest similarity is used as the representative data of that batch.

[0169] Calculate the cosine similarity between the vector and the representative data of each batch, and take the batch data with the highest similarity as the target batch data.

[0170] S108. Based on the target batch data and the overall integrated classifier, determine the detection results corresponding to the target multi-channel communication data.

[0171] Based on the target batch data and the overall ensemble classifier, the detection results corresponding to the target multi-channel communication data are determined, specifically including:

[0172] Subtract the batch spoofing verification data from the overall incremental data to obtain the incremental data of the batch spoofing verification data.

[0173] The incremental data of the batch spoofing verification data is divided into the third derived data of each batch in the batch spoofing verification data. The ratio of the data volume of the third derived data of each batch is equal to the ratio of the data volume of each batch in the batch spoofing verification data.

[0174] Obtain the difference data between the first derived data of each batch and the third derived data of the corresponding batch, and obtain the derived difference data of each batch;

[0175] The derived difference data of the Zth batch are input into each base classifier in the overall ensemble classifier, and the misclassification rate corresponding to the derived difference data of the Zth batch is output by each base classifier.

[0176] The base classifier with the smallest misclassification rate among the misclassification rates of the derived difference data of the Zth batch is determined as the recognition base classifier for the data of the Zth batch.

[0177] Obtain the latest target batch data corresponding to the target batch data, and obtain the spoofing success rate of the target batch data and the spoofing success rate of the latest target batch data;

[0178] When the spoofing success rate of the latest target batch data is less than the spoofing success rate of the target batch data, the target feature vector is input into the overall ensemble classifier, and the detection result of the target multi-channel communication data is output by the overall ensemble classifier.

[0179] When the spoofing success rate of the latest target batch data is greater than or equal to the spoofing success rate of the target batch data, the weights of the recognition base classifier corresponding to the target batch data are updated to obtain the final overall ensemble classifier.

[0180] The target feature vector is input into the final ensemble classifier, which then outputs the detection results of the target multi-channel communication data.

[0181] Obtain the spoofing success rate of the target batch data and the spoofing success rate of the latest target batch data, specifically including:

[0182] The target batch data and the latest target batch data are sent to the historical classifier respectively, and the historical classifier outputs the detection results corresponding to the target batch data and the detection results corresponding to the latest target batch data respectively.

[0183] The ratio of the amount of data that passed detection in the detection results corresponding to the target batch data to the amount of data in the target batch data is calculated, and the target batch data is determined.

[0184] The ratio of the amount of data that passed detection in the detection results corresponding to the latest target batch data to the amount of data in the latest target batch data is calculated and determined as the spoofing success rate of the latest target batch data.

[0185] Update the weights of the base classifier corresponding to the target batch of data, specifically including:

[0186] Update the weights of the base classifier corresponding to the target batch data to: use the initial weights plus the spoofing success rate of the latest target batch data.

[0187] For example, when training the overall ensemble classifier, all batches of incremental data are involved. The amount of incremental data in each batch is less than that of the first two classifiers, and is referred to as the third derived data of each batch.

[0188] Before training the overall ensemble classifier, the weights for each training data point need to be set. In the previous batch ensemble classifier and local ensemble classifier, the initial weights for each training data point were the same. For each batch of data, the weights calculated in S105 are used as the initial weights for each data point in that batch, and then the overall ensemble classifier is trained.

[0189] The difference between the first and third derived data in each batch is used as input to each base classifier in the overall ensemble classifier to obtain the misclassification rate. The base classifier with the lowest misclassification rate is then used as the recognition base classifier for the corresponding batch of data.

[0190] The latest target batch data is statistically analyzed and compared with the spoofing success rate of this batch data when constructing the overall ensemble classifier. If the spoofing success rate of the latest target batch data is less than that of the target batch data, the vector is directly input into the overall ensemble classifier to obtain the output. If the spoofing success rate of the latest target batch data is greater than or equal to that of the target batch data, the weights of the corresponding base classifier for that target batch data in the overall ensemble classifier are updated to: the original weights + the spoofing success rate of the latest target batch data. This increases the overall ensemble classifier's ability to recognize this type of spoofing pattern. The weights of other base classifiers are reduced proportionally, with the sum of all base classifier weights equal to 1. Using the adjusted overall ensemble classifier, this vector is used as input to obtain the output. An output of 0 indicates normal operation and successful ship identification; an output of 1 indicates an anomaly and failed ship identification.

[0191] This invention is now complete.

[0192] In summary, in this embodiment of the invention, historical camouflage data is clustered chronologically and divided into multiple categories based on feature similarity. This process can systematically identify typical camouflage patterns at different development stages. The clustered categories serve as prior knowledge for differential processing in subsequent weighted training and ensemble classification processes, thereby enhancing the model's ability to identify known and potential novel camouflage strategies. The performance of the dedicated classifier reflects its specific recognition capability, while the misclassification rate of the local classifier reflects the generalization ability of the data across batch scenarios, thus comprehensively measuring the effectiveness of camouflage verification under different data distributions.

[0193] This invention also proposes a multi-channel communication ship identification system; please refer to [link / reference]. Figure 4 The diagram shows a structural diagram of a multi-channel communication ship identification system provided by an embodiment of the present invention. The system includes: a data acquisition module 101, a data processing module 102, and an identification output module 103.

[0194] The data acquisition module 101 is used to acquire historical multi-channel communication data, which includes normal verification data and fake verification data. The amount of normal verification data is greater than that of fake verification data, and the fake verification data is obtained by sorting the data according to the capture time of each fake verification data.

[0195] The data processing module 102 is used to: determine batch spoofing verification data based on the spoofing verification data; train an AdaBoost ensemble classifier based on the batch spoofing verification data and normal verification data to obtain a batch ensemble classifier and a local ensemble classifier; determine first derived data and second derived data based on the batch spoofing verification data and normal verification data; calculate the data weight of each batch of data in the batch spoofing verification data using the batch ensemble classifier and the local ensemble classifier based on the first derived data and the second derived data to obtain the batch spoofing verification data with updated weights; train an AdaBoost ensemble classifier based on the normal verification data and the batch spoofing verification data with updated weights to obtain an overall ensemble classifier; acquire target multi-channel communication data and calculate the target feature vector corresponding to the target multi-channel communication data; and determine the target batch data from the batch spoofing verification data with updated weights based on the target feature vector.

[0196] The identification output module 103 is used to determine the detection result corresponding to the target multi-channel communication data based on the target batch data and the overall integrated classifier.

[0197] It should be noted that the system provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer device can be divided into different functional modules to complete all or part of the functions described above. In addition, the multi-channel communication ship identification system and the multi-channel communication ship identification method provided in the above embodiments belong to the same concept, and their specific implementation process can be found in the method embodiments, which will not be repeated here.

[0198] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0199] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0200] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method of ship identification for multi-channel communication, characterized by, The method comprises the following steps: obtaining historical multi-channel communication data, wherein the historical multi-channel communication data comprises normal verification data and camouflage verification data, the data amount of the normal verification data is greater than that of the camouflage verification data, and the camouflage verification data is arranged in order of capture time of each piece of camouflage verification data; dividing the camouflage verification data into multiple data segments based on similarity changes between feature vectors of adjacent camouflage verification data, determining the camouflage verification data in each data segment as camouflage verification data of a batch, and obtaining batch camouflage verification data; constructing batch incremental camouflage verification data of each batch of batch camouflage verification data based on the normal verification data, taking the normal verification data as correct data, and taking each batch incremental camouflage verification data as first abnormal data to train an adaboost integrated classifier, thereby obtaining a batch integrated classifier corresponding to each batch; determining adjacent batch camouflage verification data of each batch as a group of local data based on the normal verification data, constructing local incremental data of each group of local data, taking the normal verification data as correct data, and taking each group of local incremental data as second abnormal data to train an adaboost integrated classifier, thereby obtaining a local integrated classifier corresponding to each group; determining first derivative data in each batch incremental camouflage verification data that exceeds original batch incremental camouflage verification data of the batch, and determining second derivative data in each group of local incremental data that exceeds original local data of the group; determining a difference degree between adjacent batches of batch camouflage verification data based on the batch integrated classifier, determining a misclassification rate of each batch in the two batches corresponding to the batch camouflage verification data based on the local integrated classifier, calculating data weights of data in each batch of batch camouflage verification data according to the difference degree and the misclassification rate of each batch, and obtaining batch camouflage verification data after updating weights; training the adaboost integrated classifier based on the normal verification data and the batch camouflage verification data after updating weights, and obtaining an overall integrated classifier; obtaining target multi-channel communication data, calculating a target feature vector corresponding to the target multi-channel communication data, and determining target batch data from the batch camouflage verification data after updating weights according to the target feature vector; determining a detection result corresponding to the target multi-channel communication data based on the target batch data and the overall integrated classifier.

2. The method of claim 1, wherein, The method for dividing the camouflage verification data into multiple data segments based on similarity changes between feature vectors of adjacent camouflage verification data, and determining the camouflage verification data in each data segment as camouflage verification data of a batch to obtain batch camouflage verification data comprises the following steps: extracting feature vectors of each piece of data in the camouflage verification data, and sequentially calculating cosine similarities of feature vectors of adjacent two pieces of camouflage verification data to obtain a similarity sequence; obtaining local minimum values of the similarity sequence, and dividing the similarity sequence at the element positions of the local minimum values to obtain similarity sequence segments. Replace elements in each similarity sequence segment with two pieces of camouflage verification data corresponding to each element, and de-duplicate each similarity sequence segment to obtain batch camouflage verification data, wherein the number of batches in the batch camouflage verification data is the same as the number of segments in the similarity sequence segment.

3. The method of claim 2, wherein, The batch incremental camouflage verification data of each batch is constructed based on the normal verification data, the normal verification data is taken as correct data, and each batch incremental camouflage verification data is taken as first abnormal data to train the adaboost integrated classifier, to obtain the batch integrated classifier corresponding to each batch, specifically including: The normal verification data is taken as comparative data, and the smote sampling incremental method is used to increment the data in each batch of the batch camouflage verification data, to obtain batch incremental camouflage verification data, wherein the number of batches of the batch incremental camouflage verification data is the same as the number of batches in the batch camouflage verification data, and the data amount in each batch of the batch incremental camouflage verification data is the same as the data amount of the normal verification data. The normal verification data is taken as correct data, and the data in each batch of the batch incremental camouflage verification data is taken as first abnormal data to train the adaboost integrated classifier, to obtain the batch integrated classifier corresponding to the data in each batch of the batch incremental camouflage verification data.

4. The method of claim 3, wherein, The batch camouflage verification data of adjacent batches is determined as a group of local data, the local incremental data of each group of local data is constructed based on the normal verification data, the normal verification data is taken as correct data, and each group of local incremental data is taken as second abnormal data to train the adaboost integrated classifier, to obtain the local integrated classifier corresponding to each group, specifically including: The data of adjacent batches in the batch camouflage verification data is sequentially determined as a group of data to obtain local data. The normal verification data is taken as comparative data, and the smote sampling incremental method is used to increment the data in each group of the local data, to obtain local incremental data, wherein the number of groups of the local incremental data is equal to the number of batches of the batch camouflage verification data minus 1, and the data amount in each group of the local incremental data is the same as the data amount of the normal verification data. The normal verification data is taken as correct data, and the data in each group of the local incremental data is taken as second abnormal data to train the adaboost integrated classifier, to obtain the local integrated classifier corresponding to the data in each group of the local incremental data.

5. The method of claim 4, wherein, The part of the batch incremental camouflage verification data of each batch that exceeds the original batch incremental camouflage verification data is determined as first derivative data, and the part of each group of local incremental data that exceeds the original local data is determined as second derivative data, specifically including: subtracting data of each batch in the batch camouflage verification data from data of a corresponding batch in the batch camouflage verification data to obtain first derivative data of each batch in the batch camouflage verification data; subtracting data of the A th group in the local data from data of the A th group in the local incremental data to obtain incremental data of the A th group in the local data, wherein the data of the A th group in the local data contains adjacent batches in the batch camouflage verification data as the X th batch and the X+1 th batch respectively; dividing the incremental data of the A th group into second derivative data of the X th batch and second derivative data of the X+1 th batch respectively, wherein a ratio of a data amount of the second derivative data of the X th batch to a data amount of the second derivative data of the X+1 th batch is equal to a ratio of a data amount of the X th batch to a data amount of the X+1 th batch in the batch camouflage verification data.

6. The method of claim 5, wherein, The batch-based integrated classifier determines the difference between adjacent batches in the batch camouflage verification data, and based on the local integrated classifier, determines the misclassification rate of each batch in the corresponding two batches, calculates the data weight of the data of each batch in the camouflage verification data according to the difference and the misclassification rate of each batch, and obtains the batch camouflage verification data after updating the weight, specifically including: inputting data of the M+1 th batch in the batch camouflage verification data into the batch integrated classifier corresponding to the M th batch, and outputting classification data corresponding to the data of the M+1 th batch through the batch integrated classifier corresponding to the M th batch; determining classification correct data from the classification data corresponding to the data of the M+1 th batch; determining the difference of the M+1 th batch compared with the M th batch as a ratio of a data amount of the data of the M+1 th batch to a data amount of the classification correct data; obtaining the difference of the second batch compared with the first batch in the N th local integrated classifier; determining the first batch and the second batch from the local data corresponding to the N th local integrated classifier, wherein the capture time of the first batch is earlier than the time of the second batch; determining a first batch misclassification rate of the first batch in the N th local integrated classifier and a second batch misclassification rate of the second batch in the N th local integrated classifier respectively; determining a first batch data weight corresponding to the data of the first batch in the N th local integrated classifier and a second batch data weight corresponding to the data of the second batch in the N th local integrated classifier according to the difference of the second batch compared with the first batch in the N th local integrated classifier, the first batch misclassification rate of the first batch in the N th local integrated classifier and the second batch misclassification rate of the second batch in the N th local integrated classifier; determining the data weight of the data of each batch in the batch camouflage verification data according to the first batch data weight and the second batch data weight in each local integrated classifier, and obtaining the batch camouflage verification data after updating the weight.

7. The method of claim 6, wherein, The adaboost integrated classifier is trained according to the normal verification data and the batch camouflage verification data after the weight is updated, and an overall integrated classifier is obtained, and the method specifically comprises the following steps: The normal verification data is used as correct data, and the smote sampling increment method is used to increment the batch camouflage verification data, and overall increment data is obtained, wherein the data amount of the overall increment data is the same as that of the normal verification data; The normal verification data is used as correct data, the overall increment data is used as third abnormal data, the data weight of each batch data of the batch camouflage verification data after the weight is updated is used as the data weight of each batch data in the third abnormal data, and the adaboost integrated classifier is trained to obtain an overall integrated classifier.

8. The method of claim 7, wherein, The target batch data is determined from the batch camouflage verification data after the weight is updated according to the target feature vector, and the method specifically comprises the following steps: The representative data of each batch in the batch camouflage verification data after the weight is updated is determined; The cosine similarity between the target feature vector and the representative data of each batch is calculated; The representative data with the maximum cosine similarity is determined as the target batch data.

9. The method of claim 8, wherein, The detection result corresponding to the target multi-channel communication data is determined according to the target batch data and the overall integrated classifier, and the method specifically comprises the following steps: The overall increment data is obtained by subtracting the batch camouflage verification data from the overall increment data; The batch camouflage verification data is divided into third derivative data of each batch, and the ratio of the data amount of the third derivative data of each batch is equal to the ratio of the data amount of each batch in the batch camouflage verification data; The difference data between the first derivative data of each batch and the third derivative data corresponding to each batch is obtained, and the derivative difference data of each batch is obtained; The derivative difference data of the Zth batch is input into each base classifier in the overall integrated classifier, and the misclassification rate corresponding to the derivative difference data of the Zth batch is output by each base classifier; The base classifier corresponding to the minimum misclassification rate in the misclassification rate corresponding to the derivative difference data of the Zth batch is determined as the recognition base classifier of the data of the Zth batch; The latest target batch data corresponding to the target batch data is obtained, and the camouflage success rate of the target batch data and the camouflage success rate of the latest target batch data are obtained; When the camouflage success rate of the latest target batch data is less than the camouflage success rate of the target batch data, the target feature vector is input into the overall integrated classifier, and the detection result of the target multi-channel communication data is output by the overall integrated classifier; When the camouflage success rate of the latest target batch data is greater than or equal to the camouflage success rate of the target batch data, the weight of the recognition base classifier corresponding to the target batch data is updated, and a final overall integrated classifier is obtained. The target feature vector is input into the final overall integrated classifier, and a detection result of the target multi-channel communication data is output by the final overall integrated classifier.

10. A ship identification system for multi-channel communication, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The computer program, when executed by a processor, implements the steps of the method for ship identification according to any one of claims 1-9.

Images