Internet of Things equipment binding method and system based on AI behavior data

Abstract

The invention discloses an Internet of Things equipment binding method and system based on AI behavior data, and relates to the technical field of Internet of Things, and the method comprises the steps: defining Internet of Things equipment as nodes, defining a space-time interaction feature set as node attributes, obtaining the weights of edges between the nodes through calculating the behavior relevance, completing the construction of a behavior relevance topological graph, and obtaining a behavior relevance topological graph; and verifying the behavior consistency of the new equipment by using the trained GNN through the behavior association topological graph, obtaining the legality probability of the new equipment, setting a trigger threshold value, collecting implicit bioelectrical impedance data of the new equipment when the legality probability of the new equipment is smaller than the trigger threshold value, performing cross-modal feature fusion, generating a user behavior identification vector, and performing user behavior identification on the user behavior identification vector. Performing final verification to obtain a final verification result of the new equipment; according to the method, the mining of the deep behavior relationship is realized, and the limitation of a static topological structure on the mining of a dynamic interaction mode is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Internet of Things, and in particular to an Internet of Things device binding method and system based on AI behavior data. BACKGROUND

[0002] As a core link in the fields of smart home, industrial Internet of Things and smart city, the Internet of Things device binding technology aims to ensure that only legitimate devices can access the network and complete binding by verifying the interaction behavior between users and devices. In recent years, with the wide application of Internet of Things devices, the binding technology has gradually evolved from traditional static authentication based on password or device identification to dynamic verification method based on user behavior analysis. Early methods rely on single modal data, and gradually develop to use multi-modal behavior data for comprehensive verification. Graph neural networks are introduced into device behavior correlation analysis due to their advantages in modeling complex relationships, capturing the spatio-temporal interaction characteristics between devices by constructing behavior correlation topology graph. In addition, cross-modal feature fusion technology combines biological signals and behavior data, which significantly improves the verification accuracy. Generative adversarial networks are also applied to adversarial behavior detection, which trains the discriminator by generating fake behavior feature vectors to enhance the recognition ability of abnormal behavior. These technological advances have promoted the development of Internet of Things device binding from static security to dynamic and intelligent direction, adapting to diverse application scenarios.

[0003] The existing Internet of Things device binding method has significant limitations in behavior consistency verification and adversarial behavior detection. For example, traditional methods based on single modal data cannot fully capture the multi-dimensional characteristics of user-device interaction, resulting in insufficient discrimination of behavior verification and being easily bypassed by imitation attacks. Moreover, although the existing behavior correlation analysis based on graph neural networks can model the relationship between devices, the evaluation of the behavior consistency of new devices relies on static topology structure, lacking in-depth mining of dynamic interaction patterns, which limits the verification accuracy. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides an Internet of Things device binding method based on AI behavior data to solve the problem of being unable to fully capture the multi-dimensional characteristics of user-device interaction and relying on static topology structure for the behavior consistency evaluation of new devices.

[0006] To solve the above technical problems, the present application provides the following technical solutions: In a first aspect, the present application provides an Internet of Things device binding method based on AI behavior data, which comprises, Collecting multi-modal behavior data, pre-processing and feature extraction to obtain a set of spatio-temporal interaction features, the multi-modal behavior data including initial interaction data of a new device and interaction data of a bound device; The Internet of Things device is defined as a node, the space-time interaction feature set is defined as a node attribute, the weight of the edge between the nodes is obtained by calculating the behavior correlation, and the construction of the behavior correlation topology graph is completed; Through the behavior correlation topology graph, the trained GNN is used to verify the behavior consistency of the new device, and the legality probability of the new device is obtained; A trigger threshold is set, and when the legality probability of the new device is less than the trigger threshold, the implicit bioelectrical impedance data of the new device is collected, the cross-modal feature fusion is performed, the user behavior identification vector is generated, and the final verification is performed to obtain the final verification result of the new device; According to the user behavior identification vector, a false binding request is detected and stripped by using a reverse mechanism, and the Internet of Things device binding is finally completed.

[0007] As a preferred scheme of the Internet of Things device binding method based on AI behavior data, wherein: the space-time interaction feature set is obtained by performing time sequence alignment and noise removal on the multi-modal behavior data collected by the bound device and the new device, performing feature extraction and standardization processing, and then arranging the space-time interaction feature set.

[0008] As a preferred scheme of the Internet of Things device binding method based on AI behavior data, wherein: the Internet of Things device is defined as a node, the space-time interaction feature set is defined as a node attribute, the weight of the edge between the nodes is obtained by calculating the behavior correlation, and the construction of the behavior correlation topology graph is completed, and the specific steps include, Each Internet of Things device is represented as a node of the behavior correlation topology graph, and the space-time interaction feature set is used as a node attribute; The behavior similarity between devices is quantified by calculating the Pearson correlation coefficient, and the spatial distance between each pair of Internet of Things devices in the behavior correlation topology graph is calculated using Euclidean distance; The edge weight of the behavior correlation topology graph is calculated according to the behavior similarity and the spatial distance of each pair of nodes of the behavior correlation topology graph, and the construction of the behavior correlation topology graph is completed.

[0009] As a preferred scheme of the Internet of Things device binding method based on AI behavior data, wherein: through the space-time interaction feature set and the behavior correlation topology graph, the trained GNN is used to verify the behavior consistency of the new device, and the legality probability of the new device is obtained, and the specific steps include, A historical behavior correlation topology graph is constructed based on historical behavior data, a graph neural network is trained, the weight matrix and the bias matrix of the graph neural network are optimized by supervised learning, and a trained graph neural network is obtained; The behavior association topology graph is input into the trained graph neural network. The graph neural network aggregates the behavioral relationships between devices through graph convolution, evaluates the behavioral consistency of the new device with the bound devices, and outputs the legitimacy probability of the new device.

[0010] As a preferred solution of the method for binding IoT devices based on AI behavior data described in the present invention, wherein: the trigger threshold is set, and when the legitimacy probability of the new device is less than the trigger threshold, the implicit bioelectrical impedance data of the new device is collected, cross-modal feature fusion is performed, and a user behavior identification vector is generated, which specifically includes the following steps: Based on the confidence requirements, a trigger threshold is set. When the legitimacy probability of a new device is less than the trigger threshold, implicit bioelectrical impedance data and touch behavior data are collected. Training a multimodal neural network based on historical real-world behavior data, inputting historical implicit bioelectrical impedance data and historical touch behavior data into the multimodal neural network, and optimizing the attention weight matrix and attention bias vector of the multimodal neural network using supervised learning to obtain a trained multimodal neural network; wherein the historical real-world behavior data includes historical implicit bioelectrical impedance data and historical touch behavior data; The implicit bioelectrical impedance data and touch behavior data are input into the trained multimodal neural network. The multimodal neural network assigns weights through the attention mechanism to obtain the user behavior identification vector.

[0011] As a preferred solution of the method for binding IoT devices based on AI behavior data of the present invention, the method of obtaining the final verification result of the new device specifically includes the following steps: Set a similarity threshold based on the historical user identification vector and calculate the cosine similarity between the user behavior identification vector and the historical user identification vector; When the cosine similarity is higher than the similarity threshold, it is marked as legal, otherwise it is marked as illegal.

[0012] As a preferred solution of the IoT device binding method based on AI behavior data described in the present invention, the final completion of IoT device binding refers to taking the user behavior identification vector as input, and detecting the abnormality of the user behavior identification vector through the trained adversarial generative network. If false behavior is detected, the binding is rejected and the abnormal characteristics are recorded. If the detection passes and the verification result is legal, the binding is completed.

[0013] In a second aspect, the present invention provides an IoT device binding system based on AI behavioral data, comprising: an acquisition module that collects multimodal behavioral data, performs preprocessing and feature extraction, and obtains a spatiotemporal interaction feature set; A construction module defines the Internet of Things device as a node, defines the spatio-temporal interaction feature set as a node attribute, obtains the weight of the edge between nodes by calculating the behavior correlation, and completes the construction of the behavior correlation topology graph; A judgment module verifies the behavior consistency of the new device by using the trained GNN through the behavior correlation topology graph, and obtains the legitimacy probability of the new device. A verification module sets a trigger threshold, collects the implicit bioelectrical impedance data of the new device when the legitimacy probability of the new device is less than the trigger threshold, performs cross-modal feature fusion to generate a user behavior identification vector, and performs final verification to obtain the final verification result of the new device. A binding module detects and peels off a false binding request by using a reverse mechanism according to the user behavior identification vector, and finally completes the binding of the Internet of Things device.

[0014] In a third aspect, the present application provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and wherein the computer program is executed by the processor to implement any step of the AI behavior data-based Internet of Things device binding method according to the first aspect of the present application.

[0015] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement any step of the AI behavior data-based Internet of Things device binding method according to the first aspect of the present application.

[0016] The present application has the following advantages: by collecting multi-modal behavior data and generating a spatio-temporal interaction feature set, multi-dimensional behavior representation of user-device interaction is realized, providing high-resolution input for behavior consistency analysis. Compared with traditional single-modal data, multi-modal data enriches the behavior dimension and enhances the behavior discrimination, and then the implicit bioelectrical impedance feature vector and the touch feature vector are fused through the attention mechanism of the multi-modal neural network to generate a high-discrimination user behavior identification vector, further improving the verification accuracy. In addition, by constructing a dynamic behavior correlation topology graph, dynamic modeling of the spatio-temporal interaction relationship between devices is realized, and by using the trained graph neural network, behavior consistency analysis is performed based on the dynamic behavior correlation topology graph, deep behavior relationship is mined, and the limitation of static topology structure on dynamic interaction mode mining is solved. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creating any creative labor.

[0018] Fig. 1 Flow chart of the method for binding the Internet of Things device based on AI behavior data.

[0019] Fig. 2 Schematic diagram of the system for binding the Internet of Things device based on AI behavior data.

[0020] Fig. 3 Schematic diagram of the process for constructing the behavior association topology graph.

[0021] Fig. 4 Schematic diagram of the cross-modal feature fusion. DETAILED DESCRIPTION

[0022] In order to make the above objectives, features and advantages of the present application more apparent, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0023] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application, but the present application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the concept of the present application, so the present application is not limited to the specific embodiments disclosed below.

[0024] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the present application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment that excludes other embodiments.

[0025] Reference Figs. 1-4 For one embodiment of the present application, the embodiment provides a method for binding the Internet of Things device based on AI behavior data, comprising the following steps: S1, collecting multi-modal behavior data, pre-processing and feature extraction, and obtaining a set of space-time interaction features; Specifically, the following steps are included, S1.1, collecting multi-modal behavior data through the user's bound device and the new device, and pre-processing, specifically: the smart phone collects the sliding trajectory, operation timestamp and geographic position through the touch screen, the sliding trajectory includes x, y coordinate sequence, sliding speed and pressure; the smart watch collects heart rate data, acceleration data and timestamp, wherein the acceleration data reflects the wrist movement; the smart light records the switching time and brightness adjustment log; the new device collects initial interaction data, including voiceprint features of voice commands and touch operation data, wherein the touch operation data includes click position and click pressure, for example, the new device is a smart speaker.

[0026] Further explanation, by collecting multi-modal behavior data, covering various interaction modes and biological signals of users and devices, greatly enriches the dimension of behavior information All the collected data are encrypted and transmitted to the cloud server through the MQTT protocol for preprocessing. The preprocessing content includes time series alignment and noise removal. Time series alignment is to unify multi-device data to the millisecond time axis based on the linear weighting of adjacent timestamps through linear interpolation, ensuring that the data of different devices are aligned in the time dimension. Noise removal uses median filtering to process sliding trajectory data, eliminates invalid points, and applies low-pass filtering to heart rate data and voiceprint data to remove high-frequency noise. After preprocessing, the data are formatted into a unified time series dataset, including the sliding trajectory, operation timestamp, and geographic location of the smartphone, the heart rate and acceleration data of the smartwatch, the switch time and brightness adjustment log of the smart lamp, and the voiceprint features and touch operation data of the new device smart speaker.

[0027] S1.2, according to the preprocessed unified time series dataset, extract features and generate a spatio-temporal interaction feature set for each device, specifically: for the sliding trajectory of the smartphone, calculate the curvature, sliding speed average, and pressure average through the second derivative of the coordinate sequence; for the heart rate data of the smartwatch, extract the main frequency and heart rate variance of the heart rate data through fast Fourier transform, and for the acceleration data of the smartwatch, extract the acceleration average and standard deviation; for the switch time and brightness adjustment log of the smart lamp, calculate the switch time interval and brightness change frequency; for the voiceprint features of the smart speaker, convert the audio signal to the frequency domain through fast Fourier transform, apply the Mel filter bank to extract the log energy, calculate the discrete cosine transform, and generate the Mel frequency cepstral coefficient; for the touch operation data, calculate the click pressure average and click interval.

[0028] After standardizing all the extracted features, integrate them into a spatio-temporal interaction feature set, which contains time features, space features, and interaction features, covering the behavior characteristics of smartphones, smartwatches, smart lamps, and smart speakers.

[0029] Further explanation, the spatio-temporal interaction feature set comprehensively captures user behavior patterns and device usage characteristics, providing high-resolution input data for subsequent AI-based behavior consistency verification, significantly improving the accuracy and reliability of the binding process.

[0030] S2, define the Internet of Things devices as nodes and the spatio-temporal interaction feature set as node attributes, calculate the weight of the edges between nodes by calculating the behavior correlation, and complete the construction of the behavior correlation topology graph. Specifically, the following steps are included, S2.1, define each Internet of Things device as a node of a behavior association topology graph, including the bound smart phone, smart watch, smart light and new device smart speaker, the node attribute directly adopts the features in the space-time interaction feature set; the new device is a temporary node, and the node attribute also adopts the feature vector in the space-time interaction feature set, which is generated based on the initial interaction data.

[0031] S2.2, for each pair of nodes, the behavior similarity of the space-time interaction feature set of each pair of nodes is calculated using the Pearson correlation coefficient, and the expression is: ; Wherein, represents and The Pearson correlation coefficient between them is used to measure the behavior similarity between the space-time interaction feature sets of and , represents the space-time interaction feature set of device , represents the space-time interaction feature set of device , represents the mean of the space-time interaction feature set of device , represents the mean of the space-time interaction feature set of device .

[0032] According to the operation time stamp in the space-time interaction feature set of each node, the interaction time difference between device and device is calculated, which is obtained by comparing the operation time stamps of smart phones, smart watches, smart lights and smart speakers, such as sliding time, heart rate recording time, switch time, voice command time.

[0033] According to the geographical position in the space-time interaction feature set of each node, the spatial distance between device and device is calculated. Here, the Euclidean distance formula is used to solve, and the expression is: ; Wherein, represents the spatial distance between device and device , unit: meter, represents the coordinate value of the geographical position of device on the x-axis, represents the coordinate value of the geographical position of device on the x-axis, represents the coordinate value of the geographical position of device on the y-axis, representing device a coordinate value of the geographical position of the representing device on the y-axis; For each pair of nodes, the edge weight of the behavior correlation topology graph is calculated according to the behavior similarity, interaction time difference and spatial distance of each pair of nodes using the NumPy library, ensuring that recent and close-range interactions have higher weights, and the expression for calculation is: wherein, represents the edge weight between the device and the device in the behavior correlation topology graph, represents the weight coefficient of the behavior similarity term, represents the weight coefficient of the time decay term, represents the interaction time difference between the device and the device , represents the time decay parameter, represents the exponential function, represents the weight coefficient of the spatial decay term, represents the spatial decay parameter.

[0034] S2.3, using the NetworkX library to generate a directed behavior correlation topology graph, specifically: the node set of the behavior correlation topology graph includes smart phones, smart watches, smart lights and smart speakers (temporary nodes), the node attribute is the spatiotemporal interaction feature set, an edge is established for each pair of nodes based on the edge weight, and the behavior correlation topology graph is generated, representing the spatiotemporal behavior consistency between devices. The behavior correlation topology graph is stored in the Neo4j database, including node attributes and edge weights, supporting efficient query and dynamic update.

[0035] Further explanation, by generating the behavior correlation topology graph, integrating node attributes and edge weights, dynamically capturing the spatiotemporal behavior consistency between devices, through multi-dimensional behavior features (time, space, interaction) and weighted topology structure, the accuracy of behavior consistency analysis in the binding process is significantly improved, and the calculation of edge weight also considers behavior similarity, time difference and spatial distance, enhancing the accuracy of the behavior correlation topology graph for complex scenarios (such as multi-device collaborative operation), ensuring that the binding process can accurately distinguish between legal and abnormal devices.

[0036] S3, through the behavior correlation topology graph, using the trained GNN to verify the behavior consistency of the new device, obtaining the legality probability of the new device; Specifically, the steps include, ​S3.1. Based on historical behavior data, a historical behavior association topology graph is constructed and a graph neural network is trained. The historical behavior data already contains the spatiotemporal interaction feature sets and labels of bound devices (such as smartphones, smart watches, and smart lights). The labels are divided into legal and illegal. Legal represents real user devices, and illegal represents abnormal or unauthorized devices. Based on historical behavior data, a historical behavior association topology graph is constructed. The nodes are bound devices, and the node attributes are the spatiotemporal interaction feature sets in the historical behavior data.

[0037] GNN (Graph Neural Network) takes the historical behavior association topology graph as input and adopts a two-layer graph convolution structure, which is expressed as: ; in, Representation device In the graph neural network The node representation of the layer is in the form of a vector, ReLU activation function is applied to the convolution output of the graph neural network to perform nonlinear transformation. Represents the weight matrix, which is the trainable parameter of the graph neural network. Represents the mean aggregation function, used to aggregate devices All neighbor nodes of The layer representation, Representation device The node representation in the lth layer of the graph neural network is in the form of a vector, Representation device The set of neighbor nodes in the behavior association topology graph, Represents the bias matrix, which is a trainable parameter of the graph neural network. Representation device The node representation in the lth layer of the graph neural network is in the form of a vector.

[0038] For the weight matrix and the bias matrix , and the training process optimizes the graph neural network weight matrix in a supervised learning manner and the bias matrix , using the labels of historical behavior data, the goal is to maximize the behavioral consistency of legal device nodes and minimize the connection weight of illegal nodes. The training is performed on the cloud server, using the TensorFlow framework, with a batch size of 32 and a learning rate of 0.01. The training is repeated for 100 epochs and the trained weight matrix is ​​converted to and the bias matrix Save it and get the trained GNN.

[0039] S3.2, input the behavior correlation topology graph into the trained GNN, aggregate neighbor node information through two layers of graph convolution, the first layer of convolution aggregates the spatio-temporal interaction feature set of direct neighbors to generate an intermediate representation; the second layer of convolution further aggregates to generate the final node representation, reflecting the deep behavior consistency between devices. Forward propagation is performed on the new device node (smart speaker) to output a legality probability (range 0-1) reflecting its behavior consistency with the bound devices.

[0040] Further explanation, through the graph neural network to capture the complex spatio-temporal behavior relationship between devices, significantly improve the accuracy of new device behavior consistency verification, ensure that only high probability legal device directly through the binding, enhance the security of the binding process, and the historical behavior correlation topology graph provides rich training data, real-time behavior correlation topology graph captures the dynamic characteristics of the current device interaction, the combination of the two makes the graph neural network adapt to the change of behavior mode in different scenes, enhances the accuracy of verification, through the synergistic effect of historical and real-time data, effectively deal with complex or abnormal behavior, reduce the misjudgment rate.

[0041] S4, set a trigger threshold, when the legality probability of the new device is less than the trigger threshold, collect the implicit bioimpedance data of the new device, perform cross-modal feature fusion to generate a user behavior identification vector, and perform final verification to obtain the final verification result of the new device; Specifically includes the following steps, S4.1, based on the high confidence requirement of behavior consistency verification, set the trigger threshold to 0.9, if the legality probability is higher than the trigger threshold, consider that the new device behavior is consistent, the trigger signal is "no further verification", if the legality probability is lower than 0.9, the trigger signal is "bioimpedance verification is needed", to start the subsequent cross-modal fusion.

[0042] S4.2, when the trigger signal is "bioimpedance verification is needed", first collect the implicit bioimpedance data and touch behavior data through the capacitive touch sensor on the new device, the specific process is: the capacitive touch sensor measures the impedance fluctuation when the user's finger touches with alternating current signal, and collects touch behavior data at the same time, the touch behavior data includes sliding trajectory, click pressure and timestamp.

[0043] The impedance data is subjected to fast Fourier transform to extract the frequency spectrum features to generate an implicit bioimpedance feature vector, and the touch behavior data is subjected to statistical feature extraction to generate a touch feature vector, the statistical features including sliding speed mean value, trajectory curvature, and click interval.

[0044] S4.3, training the multi-modal neural network based on historical real behavior data, the historical real behavior data including historical implicit bioelectrical impedance feature vectors and historical touch feature vectors and labels, and splicing the historical implicit bioelectrical impedance feature vectors and the historical touch feature vectors into input feature vectors. The structure of the multi-modal neural network includes an input layer, an attention mechanism layer, three layers of fully connected hidden layers, and an output layer, the input layer is used to receive the input feature vectors; the attention mechanism layer allocates weights to each dimension of the input feature vectors, enhancing key features; the three layers of fully connected hidden layers each have 256 nodes; and the output layer generates a user behavior identification vector.

[0045] The core of cross-modal fusion is to use the attention mechanism layer of the multi-modal neural network to allocate weights to each dimension of the input feature vectors, and the expression is: ; Among them, represents the input feature vector processed by the attention mechanism of the multi-modal neural network, represents the input feature vector, represents a Softmax function, represents an attention weight matrix, represents an attention bias vector.

[0046] The training process is to optimize the attention weight matrix and the attention bias vector of the multi-modal neural network in a supervised learning manner, using the labels of the historical real behavior data, the goal is to minimize the classification loss, the batch size is 32, the learning rate is 0.001, and the training is performed for 150 epochs. The trained attention weight matrix and attention bias vector are saved to complete the training of the multi-modal neural network.

[0047] The real-time implicit bioelectrical impedance feature vector and the touch feature vector are spliced to obtain a real-time input feature vector, which is input into the trained multi-modal neural network. The input layer of the trained multi-modal neural network receives the real-time input feature vector, and through the attention mechanism layer with the optimized attention weight matrix and attention bias vector, the contribution of key features of the input feature vector is enhanced, and then input into the three layers of fully connected hidden layers for nonlinear transformation and deep feature extraction. The low-dimensional features are mapped to a high-dimensional space and gradually reduced in dimension to capture the complex interaction patterns between the implicit bioelectrical impedance feature vectors and the touch feature vectors, generate more discriminative feature representations, and finally generate a user behavior identification vector through the output layer using linear transformation, reflecting the comprehensive behavior characteristics of the new device.

[0048] S4.4, comparing the obtained user behavior identification vector with the historical user identification vector stored in the cloud to generate a final verification result of the new device. The similarity comparison adopts a cosine similarity method to measure, and the expression is: ; wherein, represents the cosine similarity between the user behavior identification vector and the historical user identification vector stored in the cloud, represents the user behavior identification vector, represents the historical user identification vector.

[0049] The similarity threshold is set based on the statistical characteristics of the historical user identification vector. If the cosine similarity is higher than the similarity threshold, it is considered that the behavior of the new device is highly consistent with the historical user behavior, and the final verification result of the new device is marked as "legal"; otherwise, it is marked as "illegal", for example, the similarity threshold is set to 0.95. Only when the cosine similarity is higher than 0.95, it is considered that the behavior of the new device is highly consistent with the historical user behavior.

[0050] Further, the implicit bioimpedance feature vector and the touch feature vector are used to realize robust behavior verification, which significantly reduces the risk of unauthorized device binding and enhances security, and the attention mechanism layer of the multi-modal neural network is used to fuse the implicit bioimpedance feature vector and the touch feature vector, thereby generating a highly differentiated user behavior identifier vector. This cross-modal method improves the robustness to behavior changes or attacks and ensures reliable verification in different IoT scenarios.

[0051] S5, according to the user behavior identification vector and the final verification result, using the reverse mechanism to detect and strip the false binding request, finally completing the Internet of Things device binding.

[0052] Specifically, the following steps are included, S5.1, based on the historical real behavior data and the user behavior identification vector and the final verification result, training the generative adversarial network, the generative adversarial network includes a generator and a discriminator, the generator takes random noise as input, generates a fake behavior feature vector through a three-layer fully connected network, and the fake behavior feature vector matches the user behavior identification vector in format; the discriminator takes the user behavior identification vector as input, and outputs an abnormal probability through a four-layer fully connected network, the range is 0-1.

[0053] The training process uses the historical real behavior data and the fake behavior feature vector generated by the generator to optimize the loss function of the generative adversarial network, and the training is performed on the cloud server using the TensorFlow framework, the batch size is 64, the learning rate is 0.001, and the training is performed for 200 epochs. The trained generator and discriminator parameters are stored in the PostgreSQL database.

[0054] S5.2, the discriminator takes the user behavior identification vector as input, calculates the anomaly probability through a four-layer fully connected network, the first three layers use the ReLU activation function, and the last layer uses the Sigmoid function, the anomaly probability represents the possibility of the user behavior identification vector being a fake behavior, and also reflects the deviation degree of the user behavior identification vector from the historical real behavior data.

[0055] By setting the anomaly threshold value by experience, if the anomaly probability is higher than the anomaly threshold value, for example, the anomaly threshold value is 0.3, analyze the abnormal characteristics, for example, the micro-jitter frequency of the sliding track, calculate the proportion of high-frequency Fourier components through fast Fourier transform, reflect the unnatural fluctuation in the fake behavior, and extract the activation value of the second layer of the discriminator through the reverse mechanism, build the abnormal feature space, if the anomaly probability is lower than the anomaly threshold value, directly mark the user behavior identification vector as no need for further analysis, skip the reverse mechanism processing, and enter the subsequent decision.

[0056] Example: based on the output of the second layer of the discriminator, combined with the micro-jitter frequency of the sliding track, an abnormal feature vector is generated, including the activation value and the micro-jitter frequency feature, and the abnormal feature vector is organized into an abnormal feature space in the cloud server.

[0057] Based on the output of the abnormal feature space, K-means clustering is used to separate real and fake behaviors, generate the new device's adversarial detection results and blacklist database records, specifically: if the anomaly probability is higher than 0.3, the K-means clustering sets the cluster number k to 2, based on the Euclidean distance, using the Scikit-learn library, the iteration number is 100, the projection of the user behavior identification vector in the abnormal feature space is divided into real class and fake class, if the user behavior identification vector is classified as fake class, reject the new device's binding request, and record the abnormal feature to the blacklist database, if the user behavior identification vector is classified as real class, confirm that the new device passes the adversarial detection, and allow binding.

[0058] Further explanation, through the adversarial generation network and the reverse mechanism, the fake behavior (such as the simulated sliding track) is accurately detected, the risk of false binding request passing the verification is significantly reduced, the security of the binding process is enhanced, and it is suitable for complex Internet of Things scenarios.

[0059] The embodiment also provides an Internet of Things device binding system based on AI behavior data, comprising: The acquisition module acquires multi-modal behavior data, performs preprocessing and feature extraction, and obtains a set of spatio-temporal interaction features; The construction module defines the Internet of Things device as a node, defines the set of spatio-temporal interaction features as a node attribute, calculates the weight of the edge between the nodes through the behavior correlation, and completes the construction of the behavior correlation topology graph; The judgment module verifies the behavior consistency of the new device by using the trained GNN through the behavior correlation topology graph, and obtains the legitimacy probability of the new device. The verification module sets a trigger threshold, collects the implicit bioelectrical impedance data of the new device when the legitimacy probability of the new device is less than the trigger threshold, performs cross-modal feature fusion, generates a user behavior identification vector, and performs final verification to obtain the final verification result of the new device. The binding module detects and peels off a false binding request by using a reverse mechanism according to the user behavior identification vector, and finally completes the binding of the Internet of Things device.

[0060] The embodiment also provides a computer device suitable for the AI behavior data-based Internet of Things device binding method, which comprises a memory and a processor.

[0061] The computer device can be a terminal, and the computer device comprises a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is configured to perform wired or wireless communication with an external terminal. The wireless communication can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device. In addition, the input device can also be an external keyboard, touchpad or mouse, etc.

[0062] The embodiment also provides a storage medium having a computer program stored thereon, the program being executed by a processor to implement the method for binding an Internet of Things device based on AI behavior data as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic storage, a flash memory, a magnetic disk, or an optical disk.

[0063] To sum up, the present application: by collecting multi-modal behavior data and generating a set of space-time interaction features, multi-dimensional behavior representation of user-device interaction is realized, providing high-resolution input for behavior consistency analysis. Compared with traditional single-modal data, multi-modal data enriches the behavior dimension and enhances the behavior distinguishability, and then the high-distinguishability user behavior identification vector is generated by fusing the implicit bioelectrical impedance feature vector and the touch feature vector through the attention mechanism of the multi-modal neural network, further improving the verification accuracy; in addition, by constructing a dynamic behavior association topology graph, dynamic modeling of the space-time interaction relationship between devices is realized, and the trained graph neural network is used for behavior consistency analysis based on the dynamic behavior association topology graph, deep behavior relationship mining is realized, and the limitation of static topology structure on dynamic interaction mode mining is solved.

[0064] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A method for binding IoT devices based on AI behavior data, characterized by: include, Collecting multimodal behavior data, performing preprocessing and feature extraction, and obtaining a spatiotemporal interaction feature set. The multimodal behavior data includes initial interaction data of new devices and interaction data of bound devices. IoT devices are defined as nodes, and spatiotemporal interaction feature sets are defined as node attributes. The weights of the edges between nodes are obtained by calculating the behavioral correlation, completing the construction of the behavioral correlation topology graph. Through the behavior association topology graph, the trained GNN is used to verify the behavioral consistency of the new device and obtain the legitimacy probability of the new device; Set a trigger threshold. When the legitimacy probability of a new device is less than the trigger threshold, collect the implicit bioelectrical impedance data of the new device, perform cross-modal feature fusion, generate a user behavior identification vector, and perform final verification to obtain the final verification result of the new device. Based on the user behavior identification vector, the reverse mechanism is used to detect and remove false binding requests, and finally complete the IoT device binding.

2. The method for binding IoT devices based on AI behavior data according to claim 1, wherein: Obtaining the spatiotemporal interaction feature set refers to performing time series alignment and noise removal on the multimodal behavior data collected by the bound device and the new device, performing feature extraction, and performing standardization processing to organize them into the spatiotemporal interaction feature set.

3. The method for binding IoT devices based on AI behavior data according to claim 2, wherein: The IoT devices are defined as nodes, and the spatiotemporal interaction feature set is defined as node attributes. The weights of the edges between nodes are obtained by calculating the behavioral correlation, and the construction of the behavioral correlation topology graph is completed. The specific steps include: Each IoT device is represented as a node in the behavior association topology graph, and the spatiotemporal interaction feature set is used as the node attribute; The behavioral similarity between devices is quantified by calculating the Pearson correlation coefficient, and the spatial distance between each pair of IoT devices in the behavioral correlation topology graph is calculated using Euclidean method; For each pair of nodes in the behavior association topology graph, the edge weight of the behavior association topology graph is calculated based on the behavior similarity and spatial distance to complete the construction of the behavior association topology graph.

4. The method for binding IoT devices based on AI behavior data according to claim 3, wherein: Through the spatiotemporal interaction feature set and the behavior association topology graph, the trained GNN is used to verify the behavioral consistency of the new device and obtain the legitimacy probability of the new device. The specific steps include: Based on historical behavior data, a historical behavior association topology graph is constructed, and a graph neural network is trained. The weight matrix and bias matrix of the graph neural network are optimized through supervised learning to obtain a trained graph neural network. The behavior association topology graph is input into the trained graph neural network. The graph neural network aggregates the behavioral relationships between devices through graph convolution, evaluates the behavioral consistency of the new device with the bound devices, and outputs the legitimacy probability of the new device.

5. The method for binding IoT devices based on AI behavior data according to claim 4, characterized in that: The trigger threshold is set. When the legitimacy probability of a new device is less than the trigger threshold, implicit bioelectrical impedance data of the new device is collected, cross-modal feature fusion is performed, and a user behavior identification vector is generated. Specifically, the following steps are included: Based on the confidence requirements, a trigger threshold is set. When the legitimacy probability of a new device is less than the trigger threshold, implicit bioelectrical impedance data and touch behavior data are collected. Training a multimodal neural network based on historical real-world behavior data, inputting historical implicit bioelectrical impedance data and historical touch behavior data into the multimodal neural network, and optimizing the attention weight matrix and attention bias vector of the multimodal neural network using supervised learning to obtain a trained multimodal neural network; wherein the historical real-world behavior data includes historical implicit bioelectrical impedance data and historical touch behavior data; The implicit bioelectrical impedance data and touch behavior data are input into the trained multimodal neural network. The multimodal neural network assigns weights through the attention mechanism to obtain the user behavior identification vector.

6. The method for binding IoT devices based on AI behavior data according to claim 5, characterized in that: The final verification result of the new device is obtained, which specifically includes the following steps: Set a similarity threshold based on the historical user identification vector and calculate the cosine similarity between the user behavior identification vector and the historical user identification vector; When the cosine similarity is higher than the similarity threshold, it is marked as legal, otherwise it is marked as illegal.

7. The method for binding IoT devices based on AI behavior data according to claim 6, characterized in that: The final completion of IoT device binding refers to taking the user behavior identification vector as input and detecting the abnormality of the user behavior identification vector through the trained adversarial network. If it is detected as false behavior, the binding is rejected and the abnormal characteristics are recorded. If it passes the detection and the verification result is legal, the binding is completed.

8. An IoT device binding system based on AI behavior data, based on the IoT device binding method based on AI behavior data according to any one of claims 1 to 7, characterized in that: include, The acquisition module collects multimodal behavior data, performs preprocessing and feature extraction, and obtains a spatiotemporal interaction feature set; In the construction module, IoT devices are defined as nodes, spatiotemporal interaction feature sets are defined as node attributes, and the weights of the edges between nodes are obtained by calculating the behavioral correlation, completing the construction of the behavioral correlation topology graph. The judgment module verifies the behavioral consistency of the new device using the trained GNN through the behavior association topology graph and obtains the legitimacy probability of the new device; The verification module sets a trigger threshold. When the legitimacy probability of a new device is less than the trigger threshold, it collects the implicit bioelectrical impedance data of the new device, performs cross-modal feature fusion, generates a user behavior identification vector, and performs final verification to obtain the final verification result of the new device. The binding module uses a reverse mechanism to detect and remove false binding requests based on the user behavior identification vector, and finally completes the binding of IoT devices.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for binding an IoT device based on AI behavior data according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for binding an IoT device based on AI behavior data according to any one of claims 1 to 7 are implemented.

Images