A network information security system and method fusing identity authentication

Abstract

A network information security system and method integrating identity authentication, belonging to the field of Internet of Things (IoT) security technology, aims to solve the problem of not being able to identify operator changes during active sessions in real time. The network information security method integrating identity authentication includes: acquiring capacitor discharge sequence data and spatial acceleration sequence data of the target device during touch operation; inputting the capacitor discharge sequence data and spatial acceleration sequence data into an identity authentication network model; generating interaction feature values; generating a deviation gradient between the interaction feature values ​​and preset legal values ​​based on a customized identity deviation loss function; adjusting the time weight allocation coefficients in the identity authentication network model based on the deviation gradient and executing backpropagation logic; generating a target connection identity risk value; and cutting off the network data packet transmission and reception link of the target device when the target connection identity risk value is not lower than a disconnection threshold. This invention improves the uniqueness of terminal network access identity authentication.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) security technology, specifically to a network information security system and method that integrates identity authentication. Background Technology

[0002] In network information security application scenarios, maintaining the legitimacy of the terminal operator's identity during an active session is the foundation for ensuring that sensitive data is not accessed without authorization. The relevant network information security processing logic usually verifies the static password or biometric fingerprint at the initial login node, and then issues a continuously valid network session credential to maintain a long connection, or forcibly interrupts the operation when the preset time period arrives and pops up a window requiring the operator to re-enter the secondary verification password for identity verification.

[0003] When a terminal is in an unlocked active session state and is taken over or replaced by an unauthorized person, the pre-established network session credentials are still valid, and the periodic secondary verification mechanism will block the normal continuous business operation flow, making it impossible to detect the security risk at the moment the operator changes, thus causing unauthorized leakage of core business data.

[0004] To address the above issues, a network information security system and method integrating identity authentication is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a network information security system and method that integrates identity authentication. By using this invention, the problem of not being able to identify operator changes during active sessions in real time, as mentioned above, is solved.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a network information security method integrating identity authentication, comprising: Acquire the capacitor discharge sequence data and spatial acceleration sequence data of the target device during touch operation. The capacitor discharge sequence data is a set of capacitor discharge change values ​​recorded by the screen touch-sensitive unit of the target device in continuous time segments. The spatial acceleration sequence data is the three-dimensional axial acceleration value recorded by the built-in sensor in the corresponding continuous time segments. The capacitor discharge sequence data and the spatial acceleration sequence data are input into the identity verification network model, which is configured with time weight allocation coefficients and a customized identity bias loss function. In the processing layer of the authentication network model, element-wise fusion processing is performed based on the capacitor discharge sequence data and the spatial acceleration sequence data to generate interactive feature values; Based on the customized identity deviation loss function, a deviation gradient between the interaction feature values ​​and the preset legal values ​​is generated; Based on the deviation gradient and by performing backpropagation logic, the time weight allocation coefficients in the authentication network model are adjusted. The interaction feature values ​​are weighted and aggregated using the adjusted time weight allocation coefficients to generate target connection identity risk values. When the target connection identity risk value is not lower than the disconnection threshold, the network data packet transmission and reception link of the target device is cut off.

[0007] Further, acquiring the capacitor discharge sequence data of the target device during touch operation includes: Read the values ​​of the row and column electrode intersection nodes in the screen touch-sensitive unit; The values ​​of the cross nodes are extracted according to the length of the continuous time sequence segments, and a node capacitance matrix is ​​constructed as the capacitor discharge sequence data. Determine the discrete variance of adjacent time step values ​​in the node capacitance matrix; When the discrete variance value exceeds a preset jitter threshold, it is determined that high-frequency interference exists; When the high-frequency interference is determined to exist, the decay learning rate of the current time segment is determined based on the ratio between the preset base learning rate value and the preset penalty constant.

[0008] Furthermore, before inputting the capacitor discharge sequence data and the spatial acceleration sequence data into the authentication network model, the method further includes: Monitor the pressing contact area value of the screen touch-sensitive unit; A trigger signal is generated when the pressed contact area value is greater than zero. When the trigger signal is valid, the time window is divided as the continuous timing segment.

[0009] Further, in the processing layer of the authentication network model, element-wise fusion processing is performed based on the capacitor discharge sequence data and the spatial acceleration sequence data to generate interactive feature values, including: The capacitor discharge sequence data is reorganized into a first one-dimensional vector; The spatial acceleration sequence data is reorganized into a second one-dimensional vector, where the total number of elements in the first one-dimensional vector is equal to the total number of elements in the second one-dimensional vector. Element-by-element fusion processing is performed on the corresponding element positions of the first one-dimensional vector and the second one-dimensional vector to generate the interactive feature values.

[0010] Further, generating the deviation gradient between the interaction feature values ​​and the preset legal values ​​based on the customized identity deviation loss function includes: The customized identity deviation loss function consists of a basic absolute error metric and a device ambient temperature compensation constant. The device ambient temperature compensation constant is determined by adjusting a preset temperature drift ratio based on the difference between the current battery motherboard temperature of the target device and the standard room temperature. Based on the interaction feature values ​​and the pre-recorded historical valid values, a difference measurement is performed to generate a target difference; The absolute value of the target difference is obtained as the basic absolute error metric.

[0011] Furthermore, the device ambient temperature compensation constant is determined based on the difference between the current battery motherboard temperature of the target device and the standard room temperature, after adjustment by a preset temperature drift ratio value, including: Obtain the current charging current value of the target device; When the current charging current value is greater than zero, an updated drift ratio value is generated based on the ratio between the preset temperature drift ratio value and the current charging current value. Based on the difference between the current battery motherboard temperature and the standard room temperature, the device ambient temperature compensation constant is generated after the drift ratio is adjusted by updating.

[0012] Further, adjusting the time weight allocation coefficients in the authentication network model based on the bias gradient and performing backpropagation logic includes: Read the historical gradient values ​​stored in the previous time series segment; Based on the historical gradient values ​​and the preset attenuation coefficient, an inertial gradient value is generated; The deviation gradient and the inertial gradient are fused to generate the final update gradient; The time weight allocation coefficients are updated based on the final update gradient.

[0013] Further, the step of using the adjusted time weight allocation coefficient to perform weighted aggregation processing on the interaction feature values ​​to generate target connection identity risk values ​​includes: Compare the time weight allocation coefficient with the preset retention threshold; When the time weight allocation coefficient is not higher than the preset retention threshold, the corresponding interaction feature value is reset to zero; The reset interaction feature values ​​are weighted and aggregated with the corresponding time weight allocation coefficients to generate the target connection identity risk value.

[0014] Furthermore, before cutting off the network data packet transmission and reception link of the target device when the target connection identity risk value is not lower than the disconnection threshold, the method further includes: Extract the target transmission port number of the network data packet to be sent; Based on the pre-set port security level correspondence table, query the security level identifier corresponding to the target transmission port number; Based on the penalty coefficient corresponding to the security level identifier, the disconnection threshold is reduced proportionally; When the target connection identity risk value is not lower than the reduced disconnection threshold, a link blocking command is generated; When the link blocking command is valid, the gateway media access control address stored in the internal register of the target device's network interface controller is erased; Reset the cached data pointer in the data queue to zero address.

[0015] A network information security system integrating identity authentication includes: The data acquisition module acquires the capacitor discharge sequence data and spatial acceleration sequence data of the target device during touch operation. The capacitor discharge sequence data is a set of capacitor discharge change values ​​recorded by the screen touch-sensitive unit of the target device in continuous time segments. The spatial acceleration sequence data is the three-dimensional axial acceleration value recorded by the built-in sensor in the corresponding continuous time segments. The model configuration module inputs the capacitor discharge sequence data and the spatial acceleration sequence data into the identity verification network model, which is configured with time weight allocation coefficients and a customized identity bias loss function. The interaction feature calculation module performs element-wise fusion processing on the capacitor discharge sequence data and the spatial acceleration sequence data in the processing layer of the identity verification network model to generate interaction feature values. The deviation gradient generation module generates the deviation gradient between the interaction feature values ​​and the preset legal values ​​based on the customized identity deviation loss function. The weight adaptive adjustment module adjusts the time weight allocation coefficients in the authentication network model based on the deviation gradient and by executing backpropagation logic. The risk value calculation module uses the adjusted time weight allocation coefficient to perform weighted aggregation processing on the interaction feature values ​​to generate the target connection identity risk value. The network link control module disconnects the network data packet transmission and reception link of the target device when the target connection identity risk value is not lower than the disconnection threshold.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention integrates the inherent characteristics of the device's capacitive touch hardware with the user's device posture and operation behavior characteristics to perform identity verification. Unlike the single static authentication mode of account password and fixed hardware identifier, it can continuously verify the operator's identity during the active network session of the device. It can promptly identify operator changes and unauthorized takeover behavior during the session, avoid the leakage risks caused by static login authentication and periodic secondary verification, and ensure the continuous security of terminal network sessions. 2. This invention identifies high-frequency interference in touch control through discrete variance, and combines it with device temperature drift compensation and charging condition adaptive correction mechanisms to offset feature data deviations caused by complex operating conditions, reduce false judgments in identity verification, and improve the stability and adaptability of identity authentication in different devices and usage scenarios. 3. This invention sets up a gradient inertial iteration mechanism and a dynamic time weight adjustment strategy, which can smooth data fluctuations caused by single operation anomalies and environmental disturbances. At the same time, it supports the periodic update of the standard features of legitimate users, which can adapt to subtle changes in users' long-term operating habits. This ensures the smooth operation of model iteration and reduces the probability of authentication misjudgment caused by normal operation iteration. It adopts a lightweight embedded neural network model, which completes data collection, feature fusion, risk calculation and identity verification on the terminal locally, without relying on cloud computing power, thus avoiding the latency and data leakage risks caused by cloud data transmission. Attached Figure Description

[0017] Figure 1 This is a flowchart outlining the steps of the present invention. Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: This application provides a network information security method that integrates identity authentication. Please refer to [link / reference]. Figure 1 The network information security method integrating identity authentication provided in this application includes the following steps: Step 101: Acquire the capacitor discharge sequence data and spatial acceleration sequence data of the target device during touch operation. Specifically, the target device is a mobile terminal, IoT industrial control terminal, or other device equipped with a touch screen and a three-axis attitude sensing module. Two types of core temporal characteristic data are collected synchronously throughout the entire process of the device performing network access touch operation. The capacitor discharge sequence data is collected through the matrix-type touch-sensitive unit of the device's touch screen. The touch-sensitive unit consists of a sensing array composed of multiple rows and columns of transparent electrodes, which can capture the charge fluctuations caused by changes in touch pressure, touch area, and touch position in real time. According to the system's preset fixed-duration continuous time sequence, the real-time capacitor discharge values ​​of each electrode intersection node are collected frame by frame and integrated into a complete set of capacitor discharge change values, fully preserving the inherent hardware characteristics of the user's touch operation. The spatial acceleration sequence data is collected through the device's built-in three-axis attitude sensing module. The axial acceleration sensor collects data synchronously, with the sensor sampling sequence perfectly aligned with the capacitance data acquisition sequence. Within each identical time segment, the acceleration changes along the X, Y, and Z axes of the device are collected in real time. After continuous sampling, the time sequence is sorted to form a set of three-dimensional axial acceleration values, accurately recording human behavior characteristics such as device tilt angle, shaking amplitude, and operating posture during user touch operation. The continuous time segment mentioned here refers to a sampling time window with a preset fixed duration, continuous timing, and interconnectedness, used to ensure that the timing dimension of the two types of feature data is consistent and the number of sampling frames is aligned. For example, the duration of a single time window can be preset to 20ms. During the entire process of the user clicking the device network access confirmation button, capacitance data and acceleration data are sampled synchronously every 20ms, continuously iterating throughout the process until the touch operation ends, ensuring that the sampling period of each batch of feature data is completely consistent. Step 102: Input the capacitor discharge sequence data and spatial acceleration sequence data into the authentication network model. Specifically, this application uses a lightweight embedded neural network model as the authentication network model. The model has small parameter size and low power consumption, and can be directly deployed on the local controller of various terminal devices without relying on cloud computing, thus avoiding the risk of cloud data transmission delay and leakage. The model is pre-configured with two types of core iterable parameter systems: time weight allocation coefficients corresponding to the feature weights of each time series segment, and a customized identity deviation loss function adapted to device temperature drift, touch interference, and attitude disturbance. After completing invalid value removal, time series alignment, and preliminary noise reduction preprocessing of the two types of time series data, the structured capacitor discharge sequence data and spatial acceleration sequence data are synchronized. The data is input to the input end of the identity verification network model to complete model initialization and feature data loading, providing effective data support for subsequent feature fusion, bias calculation, and weight iteration. The time weight allocation coefficient mentioned here refers to an adjustable parameter that quantifies the degree of influence of different time-series touch features on the identity verification result. The customized identity bias loss function refers to an error calculation function used to accurately quantify the degree of deviation between real-time operation features and the standard identity features of legitimate users. For example, in the data preprocessing stage, abnormal capacitance values ​​caused by voltage fluctuations at the moment of touch and invalid data with zero acceleration in the device's static state can be removed. At the same time, the sampling frame number and time-series node of the two types of data are matched one by one. After the preprocessing is completed, the normalized structured data is input into the model input end in batches to start the local model calculation process. Step 103: In the processing layer of the identity verification network model, element-wise fusion processing is performed based on the capacitor discharge sequence data and the spatial acceleration sequence data to generate interactive feature values. Specifically, the processing layer of the identity verification network model has a built-in dimension normalization and feature fusion module. First, the two types of time series data with different dimensional structures are preprocessed. Through matrix flattening, dimension completion, and linear scale normalization algorithms, the two-dimensional capacitor discharge sequence data and the three-dimensional spatial acceleration sequence data are uniformly transformed into standardized feature vectors with completely consistent dimensional dimensions and numerical scales, eliminating the computational bias caused by dimensional differences. After dimension normalization, an element-wise point-to-point fusion operation is adopted to combine the touch capacitive features representing the inherent attributes of the device hardware with the user-specific features. By deeply coupling the posture and acceleration features of the attributes, we can mine the correlation and interaction information between touch force, touch range, device tilt angle, and shaking state, and generate interaction feature values ​​with dual uniqueness of user and device. The interaction feature values ​​mentioned here refer to highly recognizable exclusive identity representation values ​​that integrate hardware touch features and human operation behavior features. For example, the original capacitance data is a two-dimensional time series matrix, and the acceleration data is a three-dimensional spatial time series matrix. By performing dimension completion operation, the two types of data are unified into a 1×N dimension feature vector. Then, the capacitance feature values ​​and acceleration feature values ​​under the same time series node are matched and coupled one by one. For example, the corresponding features of the user lightly touching the screen and the device tilting slightly are fused to generate exclusive interaction features and distinguish the operating habits of different users. Step 104: Based on the customized identity deviation loss function, generate the deviation gradient between the interaction feature values ​​and the preset legal values. Specifically, the preset legal values ​​are device-specific standard identity benchmark features, obtained by legitimate users collecting massive amounts of feature data through multiple normal touch network access operations under interference-free, normal temperature conditions. These benchmark values ​​are accumulated through model training and possess the characteristics of being device-specific, user-specific, and unreplicable. The customized identity deviation loss function built into the model is called to compare the real-time generated interaction feature values ​​with the locally stored preset legal values ​​dimension by dimension, quantifying the overall error deviation value between the real-time features and the benchmark features. Further, using the time weight allocation coefficient as the independent variable and the total error value as the base value... The dependent variable is calculated by first-order differentiation to solve for the instantaneous rate of change of the error with time-series weights, thus obtaining the corresponding deviation gradient. The deviation gradient here refers to a dynamic value that characterizes the degree of deviation between the current touch operation's identity features and the legitimate baseline features. The larger the absolute value of the gradient, the higher the risk of identity forgery and abnormal operation. For example, the model pre-stores the standard interaction feature values ​​of legitimate users entering the network normally. When the interaction features of this touch operation deviate from the standard values ​​in the dimensions of touch force and device tilt angle, the overall error of all dimensions is first counted, and then the rate of change of the error with time weights is calculated to obtain the corresponding deviation gradient. If the gradient value is large, it indicates that the characteristics of this operation are significantly different from those of legitimate users. Step 105: Based on the deviation gradient and executing backpropagation logic, adjust the time weight allocation coefficients in the authentication network model. Specifically, a lightweight backpropagation iterative logic adapted to the terminal is adopted, eliminating the need for complex matrix operations. It adaptively fine-tunes the time weight allocation coefficients corresponding to each time segment based on the positive and negative directions and magnitude of the deviation gradient. For time segments with large absolute deviation gradient values ​​and low feature matching, it is determined that there is touch interference or suspected counterfeiting operation in that time period, and the corresponding time weight is actively reduced to weaken the verification impact of abnormal and invalid features. For time segments with small absolute deviation gradient values, high feature matching, and stable operation... For each time segment, if it is determined to be a legitimate and normalized operation, the corresponding time weight is proactively increased to strengthen the dominant role of valid and legitimate features. At the same time, boundary constraints are applied to the updated weight values ​​to limit them to a reasonable range, avoiding extreme weight values ​​that could cause model failure and completing adaptive optimization iteration of model parameters. For example, if the gradient values ​​of the first three time segments of the touch screen are too large and the feature matching degree is low, the model automatically decreases the time weight of that period. If the gradient values ​​of the middle six time segments are close to zero and the operation features are stable, the model slightly increases the corresponding weights. Finally, all weight values ​​are constrained to a reasonable range of 0.01-0.99. Step 106: The adjusted time weight allocation coefficients are used to perform weighted aggregation on the interaction feature values ​​to generate the target connection identity risk value. Specifically, the adaptively iteratively optimized time weight allocation coefficients for each time segment are used as weighting factors to perform segment-by-segment weighted calculations on the interaction feature values ​​corresponding to each time segment. After filtering out invalid noise features with excessively low weight values, all valid weighted feature values ​​are globally summed, and then normalized to map the values ​​to a fixed risk range, ultimately obtaining the quantified target connection identity risk value. The target connection identity risk value mentioned here refers to the core indicator that quantifies the credibility of the current device's network access operation. The value is positively correlated with the risk of illegal network access; the higher the value, the greater the risk of identity forgery and malicious access. For example, the optimized weights for each time period are multiplied one by one with the corresponding interaction feature values. Invalid calculation results with weights below 0.05 are removed, and all remaining product results are summed to obtain the total feature value. Then, the total feature value is converted into a risk value in the 0-100 range through normalization, making it easier to intuitively determine the risk level. Step 107: When the target connection identity risk value is not lower than the disconnection threshold, disconnect the network data packet transmission and reception link of the target device. Specifically, the disconnection threshold is a standard risk judgment benchmark trained and calibrated by a large number of legitimate and illegitimate network access samples, which can be adapted to the normal working conditions of most terminals entering the network. Compare the real-time calculated target connection identity risk value with the preset disconnection threshold. When the risk value is greater than or equal to the disconnection threshold, accurately determine that the current device has the risk of identity forgery, illegal network access, or malicious attack. Immediately trigger the terminal's underlying network protection mechanism to forcibly disconnect the device's uplink and downlink network data packet transmission and reception links, block the data transmission channels between the device and the local area network and the external network, and prohibit the device from accessing the network, accessing internal network resources, and transmitting business data. If the risk value is lower than the disconnection threshold, the identity is determined to be legitimate, and the device is allowed to enter the network normally. For example, the preset normal network access disconnection threshold is 60. When the risk value generated in this verification is 65, it is determined to be an abnormal network access, and the device's network data transmission and reception are immediately blocked. If the risk value is 32, the identity is determined to be legitimate, and the device is allowed to enter the network.

[0020] In one implementation, the authentication network model can adaptively adjust the number of network layers and parameter precision according to the device's computing power resources. High-computing-power devices use a complete network structure to improve verification accuracy, while low-computing-power IoT terminals use a lightweight pruning structure to ensure real-time detection. For example, high-computing-power terminals such as smartphones and tablets use a complete model structure to refine the feature verification dimensions; while low-computing-power industrial IoT terminals prune redundant model parameters, simplify the calculation process, and ensure that the verification speed is adapted to the device performance. In another implementation, the preset legal values ​​support periodic dynamic updates, which can adaptively learn subtle changes in users' long-term operating habits and avoid misjudgments caused by the iteration of users' normal operating habits. For example, if the touch pressure decreases slightly or the device holding angle changes slightly during long-term use, the model can update the standard legal values ​​monthly to adapt to changes in users' normal operating habits. Through the above technical solutions, a two-dimensional identity verification system based on hardware touch features and user behavior and posture features is constructed to replace the traditional authentication method of single account password and fixed hardware identifier. Relying on customized loss function and adaptive weight iteration mechanism, a security assessment of network access identity risk is achieved, thereby improving the security and environmental adaptability of terminal network access. In some embodiments, acquiring capacitor discharge sequence data of the target device during touch operation includes: Step 201: Read the values ​​of the row and column electrode intersection nodes in the touch-sensitive unit of the screen; specifically, traverse all row and column electrode intersection sensing nodes of the matrix touch-sensitive array of the device's touch screen in real time, collect real-time capacitance discharge values ​​of each node, completely capture the touch charge change information of the entire screen area, retain the original feature data corresponding to the touch position, force, and range without blind spots, and ensure the integrity and accuracy of the original data by making the values ​​of each node independent and not interfering with each other; for example, the device's touch screen has a built-in 16×32 electrode sensing array. During the touch network operation, traverse 512 electrode intersection nodes row by row and column by column, and collect the real-time capacitance discharge value of each node simultaneously, completely covering the charge data of the touch area and the non-touch area; Step 202: Extract the values ​​of cross nodes according to the length of the continuous time sequence segments, and construct a node capacitance matrix as the capacitor discharge sequence data. Specifically, using a preset fixed-length time sequence segment as the extraction window, extract the capacitance values ​​of all electrode cross nodes in each time sequence window frame by frame. Construct a two-dimensional structured node capacitance matrix with the time sequence frame number as the matrix row and the electrode node number as the matrix column. Each row in the matrix corresponds to the global touch capacitance data of a time sequence segment, and each column corresponds to the time sequence change data of a single electrode node. This time sequence matrix is ​​determined as the capacitor discharge sequence data, realizing the structured regularization of the disordered original capacitance data. For example, if the preset time sequence window is 20ms, a single network access touch contains 10 time sequence segments, and the corresponding matrix generates 10 rows of data, each row containing the capacitance values ​​of 512 electrode nodes, finally forming a 10×512 two-dimensional node capacitance matrix as the capacitor discharge sequence data of this sampling. Step 203: Determine the discrete variance of adjacent time step values ​​in the node capacitance matrix. Specifically, calculate the difference between the capacitance values ​​of the same node at two adjacent time steps in the node capacitance matrix to generate a time-series fluctuation difference sequence. Calculate the average value of the difference sequence, then calculate the squared difference between each difference and the average value, and sum them to obtain the discrete variance value characterizing the severity of touch data fluctuations, accurately quantifying the abnormal fluctuation amplitude of capacitance data caused by external interference. The specific calculation process is as follows: First, select the same electrode node, and calculate the difference between the values ​​of the 1st and 2nd time steps, and the difference between the values ​​of the 2nd and 3rd time steps in sequence, and so on. Similarly, all fluctuation differences at the node are obtained, forming a difference sequence. Then, all differences are summed and divided by the total number of differences to obtain the average difference. Finally, the difference between each difference and the average is squared, all squared results are summed, and then divided by the total number to obtain the discrete variance of the node. The discrete variance of the overall matrix is ​​the average of the variances of all nodes. For example, if the difference sequence of 5 time steps of a certain electrode node is 0.2, 0.5, 0.1, 0.4, 0.3, the average value is calculated to be 0.3. Then, the squared differences between each value and the average value are calculated in turn, summed, and averaged to obtain the discrete variance of the node. Step 204: When the discrete variance value exceeds the preset jitter threshold, high-frequency interference is determined to exist. Specifically, the preset jitter threshold is a fluctuation benchmark value calibrated based on the normal touch operation conditions of the device, which can be adaptively fine-tuned according to the device model and usage scenario. When the discrete variance value is greater than this threshold, it proves that there is severe fluctuation in the capacitance data within the current time segment, which is determined to be high-frequency noise interference caused by environmental touch interference, screen jitter, or electromagnetic interference. If the discrete variance value is less than or equal to the threshold, the data is determined to be stable and without high-frequency interference. For example, the preset jitter threshold for a normal device is 0.08. If the discrete variance of the sampled data is 0.15, which is greater than the threshold, high-frequency interference such as screen jitter or electromagnetic interference is determined to exist. If the discrete variance is 0.05, the data is determined to be stable and without interference. Step 205: When high-frequency interference is detected, the decayed learning rate for the current time segment is determined based on the ratio between the preset base learning rate and the preset penalty constant. Specifically, the device pre-stores a fixed base learning rate and an interference penalty constant. After high-frequency interference is detected, the model iteration learning rate for the current time segment is dynamically decayed by dividing the base learning rate by the penalty constant. This reduces the negative impact of interference data on model weight updates and avoids abnormal data causing model parameter distortion and decreased verification accuracy. During interference-free periods, the original base learning rate is retained to ensure normal model iteration and optimization. The specific calculation process is as follows: the fixed base learning rate and the preset penalty constant value stored locally on the device are retrieved. The base learning rate value is used as the dividend, and the penalty constant is used as the divisor. The result is the decayed learning rate for the current interference period. The more severe the interference, the larger the penalty constant, and the smaller the calculated decayed learning rate value, resulting in a smoother model iteration speed. For example, if the preset base learning rate is 0.01, and the penalty constant is 5 after high-frequency interference is detected, the decayed learning rate is 0.002 obtained through division, which is used for model iteration in the current time segment.

[0021] In one implementation, the preset penalty constant can be adaptively fine-tuned according to the severity of the interference. The larger the discrete variance, the larger the penalty constant value, the more obvious the learning rate decay, and the stronger the anti-interference effect. For example, when the discrete variance is 0.1, the penalty constant is 3, and when the discrete variance is 0.2, the penalty constant is 6, corresponding to a gradual increase in the learning rate decay.

[0022] The above technical solution accurately identifies high-frequency interference noise from touch screens by using discrete variance and suppresses the negative impact of interference data on model iteration by dynamically decaying the learning rate. This solves the problems of feature data distortion and model verification misjudgment caused by device touch screen jitter and environmental interference, and improves the effectiveness of capacitance feature acquisition and model stability under complex working conditions.

[0023] In some embodiments, before inputting the capacitor discharge sequence data and spatial acceleration sequence data into the authentication network model, the method further includes: Step 301: Monitor the pressing contact area value of the screen touch-sensitive unit; Specifically, the device background polls the sensing parameters of the touch screen touch-sensitive unit in real time and continuously monitors the screen pressing contact area value; When there is no touch operation, the screen has no pressure sensing, and the contact area value is always zero; When there is a valid touch pressing action, the contact area value is greater than zero, and the value changes dynamically with the touch pressing range, which can accurately distinguish between the idle state and the operating state of the device; For example, when the device is in standby and the screen is not touched, the contact area is always 0; When the user clicks the network access touch button, the screen senses the pressing area, and the contact area value becomes the corresponding value in the range of 20-50 square millimeters. Step 302: Generate a trigger signal when the pressed contact area value is greater than zero. Specifically, the monitored contact area value is compared with the zero reference value in real time. Once a contact area value greater than zero is detected, it is immediately determined that the device has a valid touch operation, and a high-level valid trigger signal is generated simultaneously to start the subsequent data acquisition and timing segmentation process. When the contact area value returns to zero, the trigger signal is canceled and data acquisition is paused. For example, if the background detects that the contact area value changes from 0 to 35 square millimeters, a trigger signal is immediately generated to activate the data acquisition module. When the user releases their finger, the contact area returns to zero, the trigger signal becomes invalid, and all data sampling operations are paused. Step 303: When the trigger signal is valid, divide the time window into continuous time segments. Specifically, within the valid touch period when the trigger signal is effective, a fixed-duration sliding time window mechanism is used to automatically divide the time segments into continuous, equal-length, and connected time segments. Data sampling and time segmentation are only performed on the valid touch period, completely eliminating invalid blank data in idle periods without operation. The sliding time window can be dynamically updated in real time to ensure that each time segment contains valid touch features, reduce invalid model calculations, and improve data processing efficiency and feature data validity. For example, if the trigger signal is effective for 200ms and the preset single time window is 20ms, the system automatically divides this period into 10 continuous time windows, and collects feature data only for these 10 windows, without collecting invalid data in idle periods before and after touch. The above technical solution triggers data acquisition by the touch contact area, and collects feature data only for effective operation periods by dividing the time sequence into segments, thereby reducing interference from invalid data during idle periods. In some embodiments, in the processing layer of the authentication network model, element-wise fusion processing is performed based on capacitor discharge sequence data and spatial acceleration sequence data to generate interactive feature values, including: Step 401: Reorganize the capacitor discharge sequence data into a first one-dimensional vector. Specifically, the original capacitor discharge sequence data is a two-dimensional node capacitor matrix of "time frame number × number of electrode nodes". Through matrix flattening operation, all elements of the two-dimensional matrix are unfolded in sequence according to time and node order, and reorganized into a first one-dimensional feature vector with regular dimensions. This completely preserves all the original feature information of the capacitive touch, with no data loss and no feature disorder. For example, if the original capacitor matrix is ​​a two-dimensional structure of 10×512, all 5120 values ​​are arranged in sequence according to time sequence and electrode node numbering order, and reorganized into a first one-dimensional feature vector of 1×5120. Step 402: Reorganize the spatial acceleration sequence data into a second one-dimensional vector. The total number of elements in the first one-dimensional vector is equal to the total number of elements in the second one-dimensional vector. Specifically, the original spatial acceleration sequence data is a three-dimensional matrix of "time frame number × three-dimensional axis × number of sampling points". Through matrix flattening, dimension completion, and scale normalization preprocessing, the three-dimensional heterogeneous data is normalized into a one-dimensional feature vector, which is defined as the second one-dimensional vector. Through dimension adaptation operation, the total number of elements in the first one-dimensional vector and the second one-dimensional vector is forcibly unified, so as to achieve complete dimension alignment of the two sets of feature data, laying the foundation for subsequent element-by-element fusion operation. For example, after flattening the original three-dimensional acceleration matrix, a 1×4800 vector is obtained. Through dimension completion operation, 80 normalized zero values ​​are added, and finally a 1×5120 second one-dimensional vector is generated, which is completely consistent with the dimension of the first one-dimensional vector. Step 403: Perform element-by-element fusion processing on the corresponding elements of the first and second one-dimensional vectors to generate interactive feature values. Specifically, iterate through all corresponding elements of the two one-dimensional vectors, and use a multiplication and superposition fusion algorithm to couple the capacitance and acceleration feature values ​​of each set of positions to achieve point-to-point deep fusion of hardware touch features and operation posture features. After completing the fusion operation of all vector elements bit by bit, integrate them globally in chronological order to generate a set of interactive feature values ​​with unique identifiers, fully explore the correlation and interaction information of the two types of features, and avoid the limitations of single feature verification. The specific calculation process is as follows: extract the values ​​of the same position in the two vectors in sequence, multiply the two sets of values ​​to obtain the single-position fusion value, and after completing all position operations, integrate and summarize all fusion values ​​in chronological order to form a complete set of interactive feature values. For example, the capacitance feature value of the 100th position is 0.6, and the corresponding acceleration feature value is 0.4. Multiplying the two results in 0.24 is used as the fusion feature of that position. After bit-by-bit operation, integrate all results to obtain the final set of interactive features.

[0024] Through the above technical solutions, the dimensional unification and element-by-element deep fusion of heterogeneous temporal features are achieved, solving the problem that data with different structural features cannot be accurately coupled. The generated interactive features have both device hardware uniqueness and user behavior uniqueness, improving the recognizability and imitation of identity features. In some embodiments, a deviation gradient between interaction feature values ​​and preset legal values ​​is generated based on a customized identity deviation loss function, including: Step 501: Configure a customized identity deviation loss function. The loss function consists of a basic absolute error metric and a device ambient temperature compensation constant. The device ambient temperature compensation constant is determined based on the difference between the current battery motherboard temperature of the target device and the standard room temperature, adjusted by a preset temperature drift ratio. Specifically, the standard room temperature is preset to 25℃, which is the standard operating temperature of the device's touch sensor and capacitive sensing unit. Long-term operation and charging of the device will cause the motherboard to heat up, resulting in temperature drift and causing systematic deviations in capacitance and acceleration sensor values. This loss function uses the temperature compensation constant to offset such systematic errors and improve the accuracy of error calculation. The preset temperature drift ratio is a fixed parameter calibrated at the device's factory, representing the characteristic deviation ratio corresponding to each 1℃ change in temperature. For example, if the device's factory-calibrated temperature drift ratio is 0.02 / ℃, when the device's motherboard temperature is 30℃, it is 5℃ higher than the standard room temperature. The corresponding temperature compensation value can be calculated based on this ratio to correct the characteristic error caused by temperature drift. Step 502: Perform a difference measurement based on the interaction feature values ​​and the pre-recorded historical valid values ​​to generate a target difference. Specifically, retrieve the historical standard valid values ​​of valid users pre-stored in the device's local non-volatile storage unit, compare the real-time generated interaction feature values ​​with the historical valid values ​​dimension by dimension, calculate the feature difference for each dimension, and integrate all dimension differences to obtain the overall target difference, accurately representing the overall deviation between the real-time features and the valid benchmark features. The specific calculation process is as follows: compare the values ​​of each dimension of the real-time interaction features with the standard valid features one by one, calculate the value difference of each dimension, and sum up the differences of all dimensions to obtain the overall target difference. For example, the interaction features include 10 verification dimensions such as touch force, device tilt angle, and touch position. Calculate the value difference of each of the 10 dimensions, and sum them up to obtain the total target difference for this verification. Step 503: Obtain the absolute value of the target difference as the basic absolute error metric. Specifically, take the absolute value of the overall target difference obtained by solving, eliminate the interference of positive and negative directions of the deviation, retain only the error amplitude information, and determine the absolute value as the basic absolute error metric to characterize the basic deviation of the identity feature, as the core calculation basis of the loss function. For example, if the calculated target difference is -8.6, the absolute value is 8.6, which is the basic absolute error metric, objectively reflecting the basic deviation of this feature verification.

[0025] The above technical solution constructs a two-dimensional loss function that includes basic error and temperature compensation. This function can accurately quantify the basic deviation of identity features and compensate for the systematic feature error caused by equipment temperature drift. This solves the problem of misjudgment in identity verification under equipment heating conditions and improves the accuracy of deviation calculation under complex temperature conditions.

[0026] In some embodiments, the device ambient temperature compensation constant is determined based on the difference between the current battery motherboard temperature of the target device and the standard room temperature, after adjustment by a preset temperature drift ratio value, including: Step 601: Obtain the current charging current value of the target device; specifically, the battery charging current parameters are collected in real time through the device power management module to accurately obtain the current charging current value of the device, which is used to determine the real-time working condition of the device and distinguish between normal standby operation and high-temperature charging operation; for example, when the device is in standby and not charging, the charging current value is 0; when the device is wired charging, the charging current value is 2A; when wirelessly charging, the charging current value is 1A. The device's working status can be accurately distinguished by the current value. Step 602: When the current charging current is greater than zero, an updated drift ratio is generated based on the ratio between the preset temperature drift ratio and the current charging current. Specifically, when the charging current is greater than zero, the device is determined to be in charging mode. At this time, the device heats up faster and the temperature drift is greater, and the fixed reference ratio cannot adapt to the characteristics of the operating condition. The preset baseline temperature drift ratio is dynamically corrected by the ratio of the real-time charging current to the device's rated charging current. The larger the current, the larger the updated drift ratio, accurately matching the error characteristics of the high-temperature charging condition. If the charging current is zero, it is considered a normal operating condition, and the base temperature drift ratio is used. The specific calculation process is as follows: divide the current real-time charging current value of the equipment by the rated charging current value specified by the equipment at the factory to obtain the current correction coefficient. Then multiply the correction coefficient by the base temperature drift ratio value to obtain the updated drift ratio value under the charging condition. For example, if the rated charging current of the equipment is 3A, the current real-time charging current is 2A, the current correction coefficient is 0.6, the base drift ratio is 0.02 / ℃, and after multiplication, the updated drift ratio is 0.012 / ℃. Step 603: Based on the difference between the current battery motherboard temperature and the standard room temperature, after updating the drift ratio value, a device ambient temperature compensation constant term is generated. Specifically, the operating temperature of the device's battery motherboard is collected in real time, and the positive and negative temperature difference between the current temperature and the standard room temperature of 25℃ is calculated. The temperature difference value is multiplied by the drift ratio value of the corresponding operating condition. The basic drift ratio value is used for normal operating conditions, and the updated drift ratio value is used for charging operating conditions. Finally, an accurate device ambient temperature compensation constant term is calculated to achieve adaptive temperature error compensation under all operating conditions. The specific calculation process is as follows: subtract the standard room temperature of 25℃ from the current motherboard temperature of the device to obtain the temperature difference value. Multiply the temperature difference value by the drift ratio value of the corresponding operating condition, and the calculation result is the temperature compensation constant term. For example, under the charging condition, the motherboard temperature is 32℃, the temperature difference is 7℃, the updated drift ratio is 0.012 / ℃, and the two are multiplied to obtain the temperature compensation constant term of 0.084. By using the above technical solutions, the model can distinguish between normal operating conditions and high-temperature charging conditions, dynamically and adaptively adjust the temperature drift ratio, accurately adapt to the temperature deviation characteristics of devices under different operating conditions, completely solve the problem of misjudgment and missed judgment of identity verification caused by feature drift in high-temperature charging scenarios, and improve the model's adaptability to all scenarios.

[0027] In some embodiments, adjusting the time weight allocation coefficients in the authentication network model based on the bias gradient and performing backpropagation logic includes: Step 701: Read the historical gradient values ​​stored in the previous time series segment; specifically, after the model completes the weight iteration of each time series segment, it will cache the current gradient value in the local storage unit; when performing the weight update of the current time series segment, retrieve the cached historical gradient values ​​of the previous time step to retain the effective feature information of the model's historical iterations; for example, when performing the weight update of the 5th time series segment, retrieve the gradient values ​​of the 4th time series segment cached locally as the historical reference data for this iteration; Step 702: Generate inertial gradient values ​​based on historical gradient values ​​and a preset decay coefficient. Specifically, the preset gradient decay coefficient has a value range of (0,1) to weaken invalid historical gradients and retain effective iterative trends. Multiply the historical gradient values ​​by the decay coefficient to calculate the inertial gradient values ​​with iterative inertial constraint effects, avoiding sudden changes in model iteration caused by single instantaneous abnormal data. The specific calculation process is as follows: retrieve the preset fixed decay coefficient, and directly multiply the cached historical gradient values ​​by the decay coefficient. The result is the inertial gradient value. For example, if the preset decay coefficient is 0.8 and the historical gradient value is 0.25, multiplying the two gives an inertial gradient value of 0.2. Step 703: Fuse the deviation gradient and inertial gradient values ​​to generate the final update gradient. Specifically, the deviation gradient of the current time segment solved in real time is superimposed and fused with the calculated inertial gradient value. This takes into account both real-time error characteristics and historical iteration trends to generate a stable and accurate final update gradient, smoothing out gradient anomalies caused by single environmental interference and touch fluctuations. The specific calculation process is as follows: directly add the real-time deviation gradient value and the inertial gradient value of the current time segment, and the sum is the final update gradient. For example, if the real-time deviation gradient is 0.15 and the inertial gradient is 0.2, the final update gradient after fusion is 0.35. Step 704: Update the time weight allocation coefficient based on the final update gradient. Specifically, the weight correction amount is obtained by multiplying the final update gradient by the preset learning rate. The original time weight allocation coefficient is then superimposed with the weight correction amount to complete the iterative weight update. Simultaneously, boundary constraints are applied to the updated weight values ​​to stabilize them within the reasonable range of [0.01, 0.99], preventing model failure and parameter oscillations caused by weights reaching zero or full values. The specific calculation process is as follows: First, multiply the final update gradient by the preset learning rate to obtain the weight correction value. Then, add the original time weight allocation coefficient to the correction value to obtain the updated weight value. Finally, determine whether the value is within the range of 0.01-0.99. If it exceeds the range, it is forcibly corrected to the interval threshold. For example, if the preset learning rate is 0.01, the final update gradient is 0.35, the weight correction amount is 0.0035, the original weight is 0.12, and the updated weight is 0.1235, which is within the reasonable range and takes effect directly.

[0028] By introducing the above technical solution, a gradient inertia mechanism is introduced to integrate historical iteration information and real-time error information, effectively smoothing out the interference of single abnormal data and avoiding model weight oscillations. In some embodiments, the interaction feature values ​​are weighted and aggregated using the adjusted time weight allocation coefficient to generate a target connection identity risk value, including: Step 801: Compare the time weight allocation coefficients with the preset retention threshold; specifically, the device pre-stores a fixed weight retention threshold to filter out invalid features with low confidence and strong interference; compare the updated time weight allocation coefficients of each time series segment with the retention threshold one by one to distinguish between valid feature periods and invalid noise feature periods; for example, if the preset weight retention threshold is 0.05, compare the updated weight values ​​of the 10 time series segments with 0.05 respectively to filter out invalid periods below the threshold; Step 802: When the time weight allocation coefficient is not higher than the preset retention threshold, the corresponding interaction feature value is reset to zero. Specifically, if the time weight allocation coefficient of a certain time segment is less than or equal to the retention threshold, it is determined that the touch features of that time segment are severely affected by environmental interference and operation fluctuations, have extremely low credibility, and have no effective identity verification value. The interaction feature value corresponding to that time segment is directly set to zero to completely eliminate noise feature interference. For example, the weight values ​​of the 2nd and 7th time segments are 0.03 and 0.04, respectively, which are both lower than the retention threshold of 0.05. The interaction feature values ​​corresponding to the two time segments are directly reset to 0 to eliminate invalid noise data. Step 803: Perform weighted aggregation processing on the reset interaction feature values ​​and the corresponding time weight allocation coefficients to generate target connection identity risk values. Specifically, the effective interaction feature values ​​after zero-denoising are multiplied segment by segment by the corresponding optimized time weight allocation coefficients to complete the weighted operation. The weighted feature values ​​of all time segments are accumulated to obtain the total feature value. Then, the total value is mapped to a fixed risk range through normalization calibration to generate standardized and quantifiable target connection identity risk values. The specific calculation process is as follows: the interaction feature value after processing each time segment is multiplied by the corresponding weight to obtain the single-time period weighted feature value. The weighted feature values ​​of all time periods are added together to obtain the total feature value. Finally, the total feature value is converted into a standardized risk value in the range of 0-100 through a normalization algorithm. For example, the weighted feature values ​​of 8 effective time periods are accumulated to obtain a total feature value of 28.5. After normalization calibration, the final risk value is 28.5.

[0029] By using the above technical solution, invalid noise features are accurately screened and forced to zero through weight thresholds, and only highly credible valid identity features are retained to participate in risk calculation. This effectively filters feature noise caused by environmental interference and operational fluctuations, thereby improving the accuracy of target risk value calculation and identity risk judgment.

[0030] In some embodiments, before disconnecting the network packet transmission and reception link of the target device when the target connection identity risk value is not lower than the disconnection threshold, the method further includes: Step 901: Extract the target transmission port number of the network data packet to be sent; specifically, capture the header information of the network data packet to be transmitted by the device in real time, parse and extract the target transmission port number corresponding to the data packet, and accurately locate the service port accessed by the device in the current network; for example, when the device accesses the intranet file server, the target port is 445 after parsing the data packet header; when accessing ordinary web page services, the corresponding port is 80. Step 902: Based on the pre-set port security level mapping table, query the security level identifier corresponding to the target transmission port number. Specifically, the device has a pre-stored port security level mapping table, which classifies different service ports such as data transmission, device control, privacy interaction, and normal access into high, medium, and low security levels. Each security level corresponds to a unique security level identifier and penalty coefficient. Quickly match and query the corresponding security level by port number to determine the security protection level of the current service. For example, the device control port 445 is preset to be a high security level with a penalty coefficient of 0.3; the normal access port 80 is preset to be a low security level with a penalty coefficient of 0.1. Step 903: Based on the penalty coefficient corresponding to the security level identifier, proportionally reduce the disconnection threshold. Specifically, high-security ports correspond to a larger penalty coefficient and stricter risk control standards; low-security ports correspond to a smaller penalty coefficient and more lenient risk control standards. The original disconnection threshold is dynamically lowered using the correction formula: Corrected disconnection threshold = Original disconnection threshold × (1 - Penalty coefficient), achieving differentiated protection with stricter risk control for high-security ports and more lenient risk control for low-security ports. The specific calculation process is as follows: retrieve the system's original fixed disconnection threshold, combine it with the penalty coefficient corresponding to the current business port, subtract the penalty coefficient from 1 to obtain the correction ratio, and then multiply the original threshold by the correction ratio to obtain the corrected disconnection threshold adapted to the current business. For example, if the original disconnection threshold is 60, the penalty coefficient for high-security business is 0.3, the correction ratio is 0.7, and the corrected threshold is 42; if the penalty coefficient for low-security business is 0.1, the correction ratio is 0.9, and the corrected threshold is 54. Step 904: When the target connection identity risk value is not lower than the reduced disconnection threshold, a link blocking instruction is generated. Specifically, the real-time risk value is compared with the dynamically corrected differentiated disconnection threshold. If the risk value meets the standard, it is judged as a high-risk illegal access behavior, and a low-level link blocking instruction is immediately generated. For example, the correction threshold for high-security services is 42, and the current risk value is 45, which meets the judgment condition, so a link blocking instruction is generated immediately. The correction threshold for low-security services is 54, and a risk value of 45 will not trigger a blocking instruction. Step 905: When the link blocking command is effective, erase the gateway media access control address stored in the internal register of the target device's network interface controller; specifically, after the blocking command takes effect, directly operate the underlying register of the device's network interface controller to clear the internally stored gateway MAC binding information, release the inherent binding association between the device and the gateway, and prevent the device from retrying to access the network based on residual binding information; for example, if the device was originally bound to the MAC address of the local area network gateway, after the blocking command takes effect, directly clear the binding data in the register, and the device cannot reconnect to the network through residual binding information; Step 906: Reset the cached data pointers in the data queue to be sent to zero address; specifically, set all data pointers in the device's network cache queue to zero, completely clear the cached malicious data packets and residual transmission data, clear the device's local network transmission cache, and completely block the malicious data transmission path; for example, the malicious data packets to be transmitted and the half-transmission service data remaining in the device's cache queue are all cleared when the data pointers are set to zero, and no residual data can continue to be transmitted.

[0031] The above technical solutions enable differentiated dynamic risk control based on port security levels, tighten protection standards for high-security business ports, and enhance the protection capabilities of core businesses. At the same time, the deep protection mechanism of erasing underlying gateway information and clearing cache avoids the problem of incomplete blocking of conventional links. Example 2: This application provides a network information security system that integrates identity authentication. Please refer to [link / reference]. Figure 2 The system specifically includes: a data acquisition module, a model configuration module, an interactive feature calculation module, a deviation gradient generation module, a weight adaptive adjustment module, a risk numerical calculation module, and a network link management module. The data acquisition module is used to acquire in real time the capacitor discharge sequence data and spatial acceleration sequence data of the target device during touch operation. It synchronously acquires time-series feature data through the touch screen touch-sensitive unit and the triaxial sensor, and completes time sequence alignment, invalid data removal, high-frequency interference noise reduction and dimensional preprocessing to output structured two-dimensional time-series feature data, providing the raw data source for model calculation. At the same time, it supports effective touch trigger monitoring and dynamic segmentation of time sequence to avoid invalid data acquisition and improve data processing efficiency. For example, the module can automatically identify idle touch operation of the device, stop invalid data acquisition, and only perform sampling and preprocessing operations during the effective network access touch period, reducing system resource consumption. The model configuration module is used to complete the local deployment of the lightweight identity verification network model on the terminal. It configures core parameters such as the initial time weight allocation coefficient, customized identity bias loss function, learning rate, gradient decay coefficient, and various preset thresholds to achieve model initialization and feature data input adaptation, ensuring that the model can run stably in low computing power and low power consumption scenarios on the terminal. For example, for IoT low computing power terminals, this module can automatically prune redundant model parameters, simplify computation configuration, adapt to terminal hardware performance, and ensure normal model operation. The interaction feature calculation module performs dimensional normalization, vector unification, and element-wise deep fusion operations on the input two-dimensional temporal feature data. It removes the dimensional differences of heterogeneous data, mines the correlation and interaction information between hardware touch features and user operation behavior features, and generates highly identifiable and unique identity interaction feature values, completing the deep purification and coupling of the original features. For example, this module can unify the dimensions of two-dimensional capacitance data and three-dimensional acceleration data and fuse them point by point to mine the exclusive correlation features between user touch force and device grip posture, forming a unique identity identifier. The deviation gradient generation module is used to perform error compensation calculations by combining a customized identity deviation loss function with device temperature and charging conditions. It quantifies the deviation magnitude between real-time interaction features and legitimate benchmark features, and solves for accurate deviation gradient values ​​through differentiation. This accurately reflects the degree of abnormal deviation of the current identity operation and provides a quantitative basis for model weight iteration. For example, under high temperature charging conditions, this module can automatically add temperature compensation values ​​to correct the feature deviation calculation results and ensure the accuracy of the gradient values. The adaptive weight adjustment module retrieves historical gradient information to generate an inertial gradient, fuses it with real-time deviation gradients to obtain the final updated gradient, and adaptively adjusts the time weight allocation coefficients of each time segment based on lightweight backpropagation logic. At the same time, it completes weight boundary constraints and anomaly correction to ensure smooth and oscillating model iteration and continuously optimize identity verification accuracy. For example, when a single touch fluctuation causes an anomaly in the real-time gradient, this module can smoothly correct it through historical gradients to avoid weight fluctuations and maintain the stability of model iteration. The risk value calculation module is used to filter out invalid noise features by pre-set weight retention thresholds, perform weighted aggregation and normalization calibration on valid interaction features and optimized weights, and calculate the quantified target connection identity risk value to objectively characterize the security risk level of the device's network access identity. For example, this module can automatically filter invalid features caused by environmental interference and calculate the risk value only through valid features to ensure the accuracy of risk judgment. The network link management module is used to dynamically adjust the risk disconnection threshold based on the port security level, complete differentiated risk judgment, generate blocking instructions when illegal network access risk is detected, and perform deep protection operations such as network link disconnection, gateway MAC address erasure, and cache data clearing to intercept malicious network access behavior and ensure terminal network access security. For example, for high-security business ports such as device management, this module automatically tightens risk control standards, prioritizes blocking abnormal access, and protects the security of core business. Through the above technical solutions, this application constructs a dual-dimensional fusion verification system that integrates the inherent characteristics of device hardware touch control with the characteristics of user operation posture and behavior. By using a customized loss function with temperature compensation, it reduces the verification misjudgment problems caused by device temperature drift and charging condition interference. Through a gradient inertia adaptive weight iteration mechanism, it dynamically optimizes model parameters to adapt to subtle changes in user operation habits and complex environmental disturbances. Through invalid feature screening and noise reduction, and multi-condition adaptive correction, it improves the accuracy of identity risk assessment. It can run locally in real time on various low-computing-power terminal devices, improves the security and environmental adaptability of IoT terminal network access, and enhances the intelligent level of network information security protection.

[0032] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0033] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A network information security method integrating identity authentication, characterized in that, The method includes: Acquire the capacitor discharge sequence data and spatial acceleration sequence data of the target device during touch operation. The capacitor discharge sequence data is a set of capacitor discharge change values ​​recorded by the screen touch-sensitive unit of the target device in continuous time segments. The spatial acceleration sequence data is the three-dimensional axial acceleration value recorded by the built-in sensor in the corresponding continuous time segments. The capacitor discharge sequence data and the spatial acceleration sequence data are input into the identity verification network model, which is configured with time weight allocation coefficients and a customized identity bias loss function. In the processing layer of the authentication network model, element-wise fusion processing is performed based on the capacitor discharge sequence data and the spatial acceleration sequence data to generate interactive feature values; Based on the customized identity deviation loss function, a deviation gradient between the interaction feature values ​​and the preset legal values ​​is generated; Based on the deviation gradient and by performing backpropagation logic, the time weight allocation coefficients in the authentication network model are adjusted. The interaction feature values ​​are weighted and aggregated using the adjusted time weight allocation coefficients to generate target connection identity risk values. When the target connection identity risk value is not lower than the disconnection threshold, the network data packet transmission and reception link of the target device is cut off.

2. The network information security method integrating identity authentication according to claim 1, characterized in that, The acquisition of capacitor discharge sequence data of the target device during touch operation includes: Read the values ​​of the row and column electrode intersection nodes in the screen touch-sensitive unit; The values ​​of the cross nodes are extracted according to the length of the continuous time sequence segments, and a node capacitance matrix is ​​constructed as the capacitor discharge sequence data. Determine the discrete variance of adjacent time step values ​​in the node capacitance matrix; When the discrete variance value exceeds a preset jitter threshold, it is determined that high-frequency interference exists; When the high-frequency interference is determined to exist, the decay learning rate of the current time segment is determined based on the ratio between the preset base learning rate value and the preset penalty constant.

3. The network information security method integrating identity authentication according to claim 1, characterized in that, Before inputting the capacitor discharge sequence data and the spatial acceleration sequence data into the authentication network model, the method further includes: Monitor the pressing contact area value of the screen touch-sensitive unit; A trigger signal is generated when the pressed contact area value is greater than zero. When the trigger signal is valid, the time window is divided as the continuous timing segment.

4. The network information security method integrating identity authentication according to claim 1, characterized in that, In the processing layer of the authentication network model, element-wise fusion processing is performed based on the capacitor discharge sequence data and the spatial acceleration sequence data to generate interactive feature values, including: The capacitor discharge sequence data is reorganized into a first one-dimensional vector; The spatial acceleration sequence data is reorganized into a second one-dimensional vector, where the total number of elements in the first one-dimensional vector is equal to the total number of elements in the second one-dimensional vector. Element-by-element fusion processing is performed on the corresponding element positions of the first one-dimensional vector and the second one-dimensional vector to generate the interactive feature values.

5. A network information security method integrating identity authentication according to claim 1, characterized in that, The step of generating the deviation gradient between the interaction feature values ​​and the preset legal values ​​based on the customized identity deviation loss function includes: The customized identity deviation loss function consists of a basic absolute error metric and a device ambient temperature compensation constant. The device ambient temperature compensation constant is determined by adjusting a preset temperature drift ratio based on the difference between the current battery motherboard temperature of the target device and the standard room temperature. Based on the interaction feature values ​​and the pre-recorded historical valid values, a difference measurement is performed to generate a target difference; The absolute value of the target difference is obtained as the basic absolute error metric.

6. A network information security method integrating identity authentication according to claim 5, characterized in that, The device ambient temperature compensation constant is determined based on the difference between the current battery motherboard temperature of the target device and the standard room temperature, after adjustment by a preset temperature drift ratio value, and includes: Obtain the current charging current value of the target device; When the current charging current value is greater than zero, an updated drift ratio value is generated based on the ratio between the preset temperature drift ratio value and the current charging current value. Based on the difference between the current battery motherboard temperature and the standard room temperature, the device ambient temperature compensation constant is generated after the drift ratio is adjusted by updating.

7. A network information security method integrating identity authentication according to claim 1, characterized in that, The step of adjusting the time weight allocation coefficients in the authentication network model based on the deviation gradient and performing backpropagation logic includes: Read the historical gradient values ​​stored in the previous time series segment; Based on the historical gradient values ​​and the preset attenuation coefficient, an inertial gradient value is generated; The deviation gradient and the inertial gradient are fused to generate the final update gradient; The time weight allocation coefficients are updated based on the final update gradient.

8. A network information security method integrating identity authentication according to claim 1, characterized in that, The step of using the adjusted time weight allocation coefficient to perform weighted aggregation processing on the interaction feature values ​​to generate target connection identity risk values ​​includes: Compare the time weight allocation coefficient with the preset retention threshold; When the time weight allocation coefficient is not higher than the preset retention threshold, the corresponding interaction feature value is reset to zero; The reset interaction feature values ​​are weighted and aggregated with the corresponding time weight allocation coefficients to generate the target connection identity risk value.

9. A network information security method integrating identity authentication according to claim 1, characterized in that, Before disconnecting the network data packet transmission and reception link of the target device when the target connection identity risk value is not lower than the disconnection threshold, the method further includes: Extract the target transmission port number of the network data packet to be sent; Based on the pre-set port security level correspondence table, query the security level identifier corresponding to the target transmission port number; Based on the penalty coefficient corresponding to the security level identifier, the disconnection threshold is reduced proportionally; When the target connection identity risk value is not lower than the reduced disconnection threshold, a link blocking command is generated; When the link blocking command is valid, the gateway media access control address stored in the internal register of the target device's network interface controller is erased; Reset the cached data pointer in the data queue to zero address.

10. A network information security system integrating identity authentication, applied to the network information security method integrating identity authentication as described in any one of claims 1 to 9, characterized in that, The system includes: The data acquisition module acquires the capacitor discharge sequence data and spatial acceleration sequence data of the target device during touch operation. The capacitor discharge sequence data is a set of capacitor discharge change values ​​recorded by the screen touch-sensitive unit of the target device in continuous time segments. The spatial acceleration sequence data is the three-dimensional axial acceleration value recorded by the built-in sensor in the corresponding continuous time segments. The model configuration module inputs the capacitor discharge sequence data and the spatial acceleration sequence data into the identity verification network model, which is configured with time weight allocation coefficients and a customized identity bias loss function. The interaction feature calculation module performs element-wise fusion processing on the capacitor discharge sequence data and the spatial acceleration sequence data in the processing layer of the identity verification network model to generate interaction feature values. The deviation gradient generation module generates the deviation gradient between the interaction feature values ​​and the preset legal values ​​based on the customized identity deviation loss function. The weight adaptive adjustment module adjusts the time weight allocation coefficients in the authentication network model based on the deviation gradient and by executing backpropagation logic. The risk value calculation module uses the adjusted time weight allocation coefficient to perform weighted aggregation processing on the interaction feature values ​​to generate the target connection identity risk value. The network link control module disconnects the network data packet transmission and reception link of the target device when the target connection identity risk value is not lower than the disconnection threshold.

Images