Entry-exit self-service equipment information protection method based on deep federated learning

Abstract

The invention provides an entry-exit self-service equipment information protection method based on deep federated learning, and the method comprises the steps: collecting equipment related data of entry-exit self-service equipment in a protection region, extracting local features through a preset feature extraction model, carrying out the encryption aggregation of the local features through a federated learning framework, and constructing a global spatial-temporal feature matrix; constructing a space-time dual-channel neural network, taking the global space-time feature matrix as input, and outputting a dynamic risk score; a risk assessment engine based on deep reinforcement learning is constructed, an upper-layer meta-controller generates meta-strategy parameters according to historical and real-time information, and a lower-layer actuator makes a decision through a DuelingDQN algorithm; and performing threshold judgment according to the dynamic risk score, determining a hierarchical protection response level, and combining an expected accumulated reward output by a lower-layer actuator to optimize action details. According to the method, the equipment information security is guaranteed through the steps, and the accuracy and effectiveness of risk assessment and response are improved.

Description

Technical Field

[0001] The present invention relates to an information protection method for entry-exit self-service equipment based on deep federated learning, and belongs to the technical field of information data protection. Background Art

[0002] In recent years, with the continuous development of the economy and society and the deepening of reform and opening up, the flow of people between China and other countries has become increasingly frequent, and the number of people entering and leaving the country has increased significantly, with rapid entry and exit. Faced with the huge daily processing volume, traditional manual check-in counters are no longer able to meet the demand, prompting the large-scale deployment of self-service entry and exit devices. However, while these devices improve efficiency, they also expose security risks: a lack of coordinated protection mechanisms between devices makes cross-device coordinated attacks difficult to trace; sensitive data such as biometrics is vulnerable to theft during transmission; and traditional static risk rules are unable to cope with dynamically changing attack methods.

[0003] Even in the prior art, for example, the Chinese invention patent application with publication number CN119227130A discloses a method for detecting abnormal access to sensitive data based on federated learning; the method includes: implementing distributed training through federated learning, allowing model parameters to be collected and aggregated from the local devices of each participant under the premise of protecting user privacy to build a global text classification model, thereby ensuring data privacy and security; constructing a TextCNN model, which is used on the local device to train the text classification model, and can effectively capture local features in the text, improve the accuracy and generalization ability of the model on local data, and judge whether text data such as user input and program logs are abnormally accessed; compared with the static management method that obtains a token to continuously access resources, the present invention can dynamically perform abnormal access detection according to actual conditions by recognizing text, and by using federated learning, it can expand the amount of data required for model training while protecting data security, thereby improving the generalization ability of the model. The above patent is based on the fact that it relies on a single text modality (such as program logs) for modeling, and it is difficult to integrate multi-dimensional data such as network traffic statistical characteristics and biometrics. The static TextCNN model cannot capture the spatiotemporal dependencies between devices, and lacks the ability to model cross-device collaborative attacks. The response mechanism based on fixed thresholds is difficult to adapt to dynamically evolving risk scenarios, and lacks a closed-loop feedback optimization mechanism.

[0004] Therefore, there is an urgent need for an information protection method that can deeply integrate multimodal data, accurately model the spatiotemporal interaction characteristics between devices, and realize the linkage of dynamic risk assessment and intelligent response. Summary of the Invention

[0005] In order to solve the problems existing in the above-mentioned prior art, the present invention proposes an information protection method for entry and exit self-service equipment based on deep federated learning.

[0006] The technical solutions of the present invention are as follows: The present invention provides a method for protecting information of entry-exit self-service equipment based on deep federated learning, the method comprising: Obtain device-related data from all entry-exit self-service devices within the protected area, including operation logs, network traffic, and biometric data. Extract local features from the device-related data using a pre-set feature extraction model. The local features of entry-exit self-service devices are encrypted and aggregated through a federated learning framework to construct a global spatiotemporal feature matrix. Constructing a spatiotemporal dual-channel neural network, comprising a temporal channel and a spatial channel. The temporal channel uses dilated convolution to extract cross-device collaborative operation patterns, and the spatial channel uses a graph attention network to model inter-device dependencies. The global spatiotemporal feature matrix is ​​used as input to the spatiotemporal dual-channel neural network, outputting a dynamic risk score. The operation sequence and environmental context risks of the entry and exit self-service equipment are collected in real time, and the dynamic risk score is input into the pre-built risk assessment engine based on deep reinforcement learning. Combined with the real-time collected operation sequence and environmental context risks, the decision-making process is triggered, and the risk assessment engine outputs graded response protection actions.

[0007] Preferably, the preset feature extraction model includes an operation log feature extraction module, a network traffic feature extraction module, and a biometric feature extraction module, wherein: The structure of the operation log feature extraction module includes a bidirectional LSTM layer and an attention pooling layer. The operation sequence is used as the input of the operation log feature extraction module, and the operation log features are output, including the operation type, time interval and access object fields. The network traffic feature extraction module calculates network traffic features based on the traffic statistics window, including the number of packets per second, protocol type distribution and traffic entropy value; The biometric feature extraction module uses a ResNet-34 model trained with the ArcFace loss function, takes biometric data as input, and outputs biometric features, including fingerprint features and facial features; The operation log features, network traffic features and biometric features are spliced ​​together to obtain local features.

[0008] Preferably, the local features of the entry and exit self-service devices are encrypted and aggregated through the federated learning framework to construct a global spatiotemporal feature matrix, specifically: Calculate the local gradient of the preset feature extraction model corresponding to each entry and exit self-service device and encrypting the local gradient using an encryption algorithm; The entry and exit self-service device uploads the encrypted local gradient to the central server. The central server aggregates all encrypted local gradients within the protection area to obtain the global gradient, which is expressed as: ; Where, is the global gradient; The total number of self-service entry and exit devices; After encryption The local gradient of the preset feature extraction model corresponding to each entry and exit self-service device; is the modulus in the encryption algorithm; For the The local gradient of the preset feature extraction model corresponding to each entry and exit self-service device; The central server distributes the global gradient to each entry-exit self-service device. The entry-exit self-service device decrypts and updates the preset feature extraction model based on the global gradient. The decryption update is expressed as follows: ; in: ; ; Where, Extract the gradient of the model for the updated preset features; To perform decryption operations; is the Carmichael function value; To determine and The least common multiple of and Both are prime numbers in encryption algorithms; is the decryption coefficient; Transformation function for information decoding; It is the generator in the encryption algorithm; Based on the preset feature extraction model after decryption and update, feature extraction is performed on the device-related data to obtain local features, and the local features are encrypted using an encryption algorithm, which can be expressed as follows: ; ; ; Where, After encryption The first self-service entry and exit equipment dimensional local features; For the The first self-service entry and exit equipment dimensional local features; For the The first self-service entry and exit equipment Dimensional local features correspond to random numbers; is the total dimension of local features; is the set of real numbers; According to the time slice and the spatial distribution of entry and exit self-service equipment, the encrypted local features are aggregated into a global spatiotemporal feature matrix, which can be expressed as follows: ; ; Where, is the global spatiotemporal feature matrix; For the The self-service entry and exit equipment is The time slice dimensional local features; is the number of time slices.

[0009] Preferably, the time channel adopts dilated convolution to extract the cross-device collaborative operation mode specifically as follows: Slice the global spatiotemporal feature matrix by device dimension to obtain the temporal feature sequence of the entry and exit self-service equipment , expressed as: ; The first The time series feature sequence of the entry and exit self-service equipment is input into the expansion convolution layer of the time channel, which is expressed as follows: ; ; Where, For the Layer dilation convolution layer in time slice The output value of is the number of dilated convolutional layers; is the convolution kernel size; For the The weight parameters of the layer-expanded convolutional layer; For the Layer dilation: The dilation rate of the convolution layer controls the spacing of the convolution kernel elements; is the index of the convolution kernel element; For the The output value of the layer expansion convolution layer in the time slice, is the initial layer; The output of the dilated convolutional layer is processed by the LSTM network to obtain the temporal hidden state, which can be expressed as: ; Where, Time slice The hidden state of For the previous time slice The hidden state of The attention weight is calculated based on the temporal hidden state and is expressed as: ; ; Where, Time slice The attention weight of is a learnable parameter; is the transpose operation; is the hidden state weight matrix; is the context weight matrix; is the global context vector; is a nonlinear enhancement function; The weighted temporal features are obtained based on the attention weight calculation, which can be expressed as follows: .

[0010] Preferably, the spatial channel models the inter-device dependency relationship through a graph attention network: Slice the global spatiotemporal feature matrix according to the time slice dimension to obtain the global device state sequence of the time slice , expressed as: ; Taking entry-exit self-service devices as nodes and the interaction relationships between entry-exit self-service devices as edges, we construct an entry-exit self-service device association graph. The initial node feature is the global device state sequence, and the edge weight is expressed as: ; Where, For nodes and nodes The weight of the connecting edge; is the total number of communications between nodes; is the number of communications between nodes and nodes; To perform cosine similarity calculation; Dynamically calculate nodes using graph attention network and nodes The attention coefficient is expressed as follows: ; Where, is the attention coefficient of node and node; is the attention vector, which determines the strength of the association between nodes; 、 and Node 、 and Node characteristics; 、 and It is a trainable weight used to map the corresponding node feature dimension; Representation node For nodes Neighbors; For nodes The set of neighbors of By weighted aggregation of node neighborhood features, node aggregation features are generated, which can be expressed as: ; Where, For nodes Aggregation characteristics of is a nonlinear activation function, which in this embodiment is ELU (exponential linear unit) to retain negative information; is the trainable aggregation weight; Representation node For nodes Neighbors; Perform maximum pooling on the node aggregation features to obtain spatial features , expressed as: .

[0011] Preferably, the global spatiotemporal feature matrix is ​​used as the input of the spatiotemporal dual-channel neural network, and the output dynamic risk score is specifically: The time channel of the spatiotemporal dual-channel neural network is used to obtain weighted time series features, and the space channel of the spatiotemporal dual-channel neural network is used to obtain spatial features. The weighted time series features and spatial features are fused to obtain fused features, which can be expressed as follows: ; Where, It is a fusion feature; is the fusion weight; is the fusion bias; The dynamic risk score is calculated based on the fusion features and is expressed as follows: ; Where, dynamic risk scoring; is a learnable risk weight; For risk bias.

[0012] Preferably, the upper meta-controller receives the historical dynamic risk score window and the real-time operation sequence and environmental context, and generates the meta-strategy hidden state through LSTM, which is expressed as follows: ; Where, For the current moment Meta-strategy The hidden state of As of the current moment Dynamic risk scoring window; For the current moment Real-time operation sequence The sliding window mean of ; For the current moment Environmental context The sliding window mean of ; is the parameter set of LSTM; The meta-strategy hidden state is mapped to the meta-strategy parameter, which is used to regulate the action selection preference of the lower-level executor. It can be expressed as: ; Where, For the current moment Meta-strategy parameters of ; is the meta-strategy mapping function; is the parameter set of the meta-strategy mapping function; The upper meta-controller constructs the objective function to maximize long-term risk suppression and strategy stability, which can be expressed as: ; ; ; Where, is the objective function of the upper-level meta-controller; To perform the sliding window mean operation; Plan your steps for the long term; is the time slice interval; is the discount factor; is the cost weight coefficient; The moment of the upper meta-controller Instant rewards; is the variance, Constraint strategy shocks; For the moment Instant rewards; The magnitude of the reduction for the dynamic risk score; For the moment Dynamic risk scoring at the time of For the moment Dynamic risk scoring at the time of To punish the violent fluctuation coefficient of the score; For the moment Actions implementation costs; is the learning rate.

[0013] Preferably, the lower-layer executor makes action decisions in real time through the DuelingDQN algorithm, takes the state variables consisting of dynamic risk score, real-time operation sequence and environmental context as input, and outputs state value and action advantage, which can be expressed as follows: ; ; Where, To be in state Take action Expected cumulative rewards to be obtained; is the state value function; is the parameter set of the state value function; is the action advantage function; is the parameter of the action advantage function; To respond to the action set; Action Advantage Set Any action in is a state variable; The meta-strategy parameters of the upper meta-controller Adjust the expected cumulative reward and express it as follows: ; Where, is the adjusted expected cumulative reward; The immediate reward of the lower-level executor is expressed as follows: ; Where, Provide immediate rewards for lower-level executors; Reward for continuous action.

[0014] Preferably, the hierarchical protection response actions are driven collaboratively based on the dynamic risk score and the expected cumulative reward output by the lower-level actuators. Specifically: The threshold is determined based on the dynamic risk score to determine the hierarchical protection response level: when When it is determined to be a low-risk state, a three-level response is carried out, and the hierarchical protection response action is to trigger lightweight encrypted communication; when When the risk is determined to be medium, a secondary response is carried out, and the graded protection response action is to initiate a biometric authentication review; when When it is determined to be a high-risk state, a level one response is carried out, and the hierarchical protection response action is to perform fuse isolation; in, and are the preset lower and upper risk thresholds respectively; Within the hierarchical protection response levels divided by threshold judgment, the action execution details are optimized through the expected cumulative rewards output by the lower-level executor.

[0015] Preferably, the method further includes calculating the reverse impact of the dynamic risk score on the performance of the hierarchical protection response action, which is expressed as follows: ; Where, for Dynamic risk scoring at all times; For the general The global spatiotemporal feature matrix at each moment is used as the input of the spatiotemporal dual-channel neural network to output the dynamic risk score; express Momentary action The expected risk suppression amount.

[0016] The present invention has the following beneficial effects: 1. This invention is a method for protecting information on entry-exit self-service devices based on deep federated learning. It combines a preset feature extraction model with a federated learning framework to comprehensively extract the features of device-related data and achieve data aggregation while protecting data privacy. Specifically, the encryption mechanism of federated learning avoids direct sharing of device data and prevents the leakage of sensitive information. The constructed global spatiotemporal feature matrix integrates the feature information of different devices at different times, providing a more comprehensive and accurate data foundation for subsequent risk assessment.

[0017] 2. This invention provides an information protection method for self-service entry and exit equipment based on deep federated learning. It constructs a spatiotemporal dual-channel neural network for risk scoring, analyzing the risks of self-service entry and exit equipment from both temporal and spatial dimensions. The cross-device collaborative operation patterns extracted through the temporal channel can capture temporal patterns and anomalies in device operation, while the inter-device dependencies modeled through the spatial channel can identify potential connections between devices and risk propagation paths. This multi-dimensional analysis makes dynamic risk scoring more accurate and comprehensive, enabling timely identification of potential security risks associated with devices. 3. This invention provides an information protection method for self-service entry and exit equipment based on deep federated learning. A risk assessment engine based on deep reinforcement learning implements hierarchical response protection. Protection strategies are dynamically adjusted based on real-time risk conditions and environmental information. Tiered response protection actions can employ different levels of protection based on risk level. This ensures the normal operation of the equipment in low-risk conditions while enabling timely implementation of effective protection measures, such as fuse isolation, to prevent further escalation of risk in high-risk conditions. Furthermore, this invention continuously optimizes action execution details through reinforcement learning mechanisms, improving the adaptability and effectiveness of protection strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0020] It should be understood that the step numbers used herein are only for convenience of description and are not intended to limit the order in which the steps are to be executed.

[0021] It should be understood that the terms used in the present specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, the singular forms "a", "an" and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0022] The terms “include” and “comprising” indicate the presence of described features, integers, steps, operations, elements and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0023] The term "and / or" refers to and includes any and all possible combinations of one or more of the associated listed items.

[0024] Example 1: See also Figure 1 This embodiment provides a method for protecting information of entry-exit self-service equipment based on deep federated learning, the method comprising: S1. Obtain device-related data from all entry-exit self-service devices within the protected area, including operation logs, network traffic, and biometric data. Extract local features from the device-related data using a pre-set feature extraction model. S11. The operation log is used to record the operation sequence of the access subject on the self-service immigration device; the network traffic includes data transmission packets between the self-service immigration device and the external network; the biometric data includes data such as fingerprint characteristics and facial features used for identity verification; S12. Furthermore, in this embodiment, the preset feature extraction model includes an operation log feature extraction module, a network traffic feature extraction module, and a biometric feature extraction module, wherein: S121, the structure of the operation log feature extraction module includes a bidirectional LSTM layer and an attention pooling layer, takes the operation sequence as the input of the operation log feature extraction module, and outputs operation log features, including operation type, time interval, and access object fields; S122, the network traffic feature extraction module calculates network traffic features based on the traffic statistics window (data transmission packets within 1 second), including the number of data packets per second, protocol type distribution, and traffic entropy value; S123, the biometric feature extraction module uses a ResNet-34 model trained with an ArcFace loss function, takes biometric feature data as input, and outputs biometric features, including fingerprint features and facial features; S124, combining the operation log features, network traffic features, and biometric features to obtain local features; S2. Use the federated learning framework to encrypt and aggregate the local features of entry-exit self-service devices and construct a global spatiotemporal feature matrix. Specifically: S21. Calculate the local gradient of the preset feature extraction model corresponding to each entry and exit self-service device and encrypt the local gradient using an encryption algorithm, wherein the encryption algorithm includes a symmetric key encryption algorithm such as Twofish, Serpent, AES (Rijndael), Blowfish, CAST5, RC4, 3DES, and IDEA, a public key encryption and authentication protocol such as Diffie-Hellman key exchange, DSA digital signature algorithm, ElGamal encryption, elliptic curve cryptography (ECC), password authenticated key agreement (PAKE), Paillier homomorphic encryption, RSA algorithm, Cramer-Shoup encryption system, and YAK authenticated key agreement protocol, or a hybrid encryption algorithm in which a symmetric algorithm encrypts the data body and a public key algorithm encrypts the symmetric key. The corresponding encryption algorithm is selected according to the access object requirements. In this embodiment, the encryption algorithm is Paillier homomorphic encryption, specifically: S211. Initialize the Paillier key, which is expressed as: ; ; Where, is the modulus; and are all prime numbers, and ; is the generator; The public key is ; S212: Encrypt the local gradient based on the public key, which can be expressed as: ; Where, After encryption The local gradient of the preset feature extraction model corresponding to each entry and exit self-service device; For the The local gradient of the preset feature extraction model corresponding to each entry and exit self-service device; For the Each entry and exit self-service device corresponds to a random number; S213. The entry and exit self-service device uploads the encrypted local gradient to the central server. The central server aggregates all encrypted local gradients within the protection area to obtain the global gradient, which is expressed as: ; Where, is the global gradient; The total number of self-service entry and exit devices; S214. The central server distributes the global gradient to each entry-exit self-service device. The entry-exit self-service device decrypts and updates the preset feature extraction model based on the global gradient. The decryption update is expressed as: ; in: ; ; Where, Extract the gradient of the model for the updated preset features; To perform decryption operations; is the Carmichael function value; To determine and the lowest common multiple of ; is the decryption coefficient; is the information decoding conversion function. In the decryption update formula, Obtaining intermediate results ,pass The intermediate result is converted, and the conversion rule is ; The converted intermediate result and Multiply and Modulo, obtain ; and in the decryption coefficient In the formula, the intermediate result for ,pass right Perform the conversion and calculate the intermediate result after conversion and its multiplication inverse The product of Modulo, obtain ; S215. Preferably, this embodiment further includes using differential privacy technology to add Gaussian noise to the gradient when training the preset feature extraction model locally. ,in, is the standard deviation, , where Budget for privacy, is the failure probability. In this embodiment, , ; S22. Extract features from the device-related data based on the updated preset feature extraction model after decryption to obtain local features. Encrypt the local features using an encryption algorithm, which can be expressed as: ; ; ; Where, After encryption The first self-service entry and exit equipment dimensional local features; For the The first self-service entry and exit equipment dimensional local features; For the The first self-service entry and exit equipment Dimensional local features correspond to random numbers; is the total dimension of local features. Preferably, in this embodiment, the dimension of the operation log feature is 128, the dimension of the network traffic feature is 3, and the dimension of the biometric feature is 512. dimension; is the set of real numbers; S23. Aggregate the encrypted local features into a global spatiotemporal feature matrix according to the time slice and the spatial distribution of the entry and exit self-service equipment. The global spatiotemporal feature matrix is ​​a three-dimensional (time slice-equipment-feature) matrix, which is expressed as: ; ; Where, is the global spatiotemporal feature matrix; For the The self-service entry and exit equipment is The time slice dimensional local features; is the number of time slices; preferably, in this embodiment, the time slices are divided into one time slice every 5 seconds; S3. Construct a spatiotemporal dual-channel neural network, comprising a time channel and a space channel. The time channel uses dilated convolution to extract cross-device collaborative operation patterns, and the space channel uses a graph attention network to model inter-device dependencies. The global spatiotemporal feature matrix is ​​used as input to the spatiotemporal dual-channel neural network, outputting a dynamic risk score. S31, the time channel uses dilated convolution to extract the cross-device collaborative operation mode specifically as follows: Slice the global spatiotemporal feature matrix by device dimension to obtain the temporal feature sequence of the entry and exit self-service equipment , expressed as: ; The first The time series feature sequence of each entry-exit self-service device is input into the dilated convolution layer of the time channel to capture long-term temporal dependencies, which can be expressed as follows: ; ; Where, is the output of the dilated convolutional layer, i.e. Layer dilation convolution layer in time slice The output value of is the number of dilated convolutional layers; is the convolution kernel size; For the The weight parameters of the layer-expanded convolutional layer; For the Layer dilation: The dilation rate of the convolution layer controls the spacing of the convolution kernel elements; is the index of the convolution kernel element; preferably, in this embodiment, the number of the dilated convolution layers is 4, and the weight coefficients are 1, 2, 4 and 8 respectively, ; For the The output value of the layer expansion convolution layer in the time slice, is the initial layer; The output of the dilated convolutional layer is processed by the LSTM network to obtain the temporal hidden state, which can be expressed as: ; Where, Time slice The hidden state of For the previous time slice The hidden state of The attention weight is calculated based on the temporal hidden state and is expressed as: ; ; Where, Time slice The attention weight of is a learnable parameter; is the transpose operation; is the hidden state weight matrix; is the context weight matrix; is the global context vector; is a nonlinear enhancement function; The weighted temporal features are obtained based on the attention weight calculation, which can be expressed as follows: ; S32, the spatial channel uses the graph attention network to model the dependency relationship between devices: Slice the global spatiotemporal feature matrix according to the time slice dimension to obtain the global device state sequence of the time slice , expressed as: ; Taking entry-exit self-service devices as nodes and the interaction relationships between entry-exit self-service devices as edges, we construct an entry-exit self-service device association graph. The initial node feature is the global device state sequence, and the edge weight is expressed as: ; Where, For nodes and nodes The weight of the connecting edge; is the total number of communications between nodes; is the number of communications between nodes and nodes; To perform cosine similarity calculation; Dynamically calculate nodes using graph attention network and nodes The attention coefficient is expressed as follows: ; Where, is the attention coefficient of node and node; is the attention vector, which determines the strength of the association between nodes; 、 and Node 、 and Node characteristics; 、 and It is a trainable weight used to map the corresponding node feature dimension; Representation node For nodes Neighbors; For nodes The set of neighbors of By weighted aggregation of node neighborhood features, node aggregation features are generated, which can be expressed as: ; Where, For nodes Aggregation characteristics of is a nonlinear activation function, which in this embodiment is ELU (exponential linear unit) to retain negative information; is the trainable aggregation weight; Representation node For nodes Neighbors; Perform maximum pooling on the node aggregation features to obtain spatial features , expressed as: ; S33: Using the global spatiotemporal feature matrix as the input of the spatiotemporal dual-channel neural network, the output dynamic risk score is specifically: The time channel of the spatiotemporal dual-channel neural network is used to obtain weighted time series features, and the space channel of the spatiotemporal dual-channel neural network is used to obtain spatial features. The weighted time series features and spatial features are fused to obtain fused features, which can be expressed as follows: ; Where, It is a fusion feature; is the fusion weight; is the fusion bias; The dynamic risk score is calculated based on the fusion features and is expressed as follows: ; Where, dynamic risk scoring; is a learnable risk weight; for risk bias; S4: Real-time collection of entry and exit self-service device operation sequences and environmental contextual risks. Dynamic risk scores are input into a pre-built risk assessment engine based on deep reinforcement learning (DRL). Combined with the real-time collection of operation sequences and environmental contextual risks, the decision-making process is triggered, and the risk assessment engine outputs graded response protection actions. S41. The pre-built risk assessment engine based on deep reinforcement learning (DRL) includes an upper-layer meta-controller and a lower-layer executor, wherein: The upper meta-controller receives the historical dynamic risk score window and the real-time operation sequence and environment context, and generates the meta-strategy hidden state through LSTM, which is expressed as: ; Where, For the current moment Meta-strategy The hidden state of As of the current moment Dynamic risk scoring window; For the current moment Real-time operation sequence The sliding window mean of ; For the current moment Environmental context The sliding window mean of ; is the parameter set of LSTM; The meta-strategy hidden state is mapped to the meta-strategy parameter, which is used to regulate the action selection preference of the lower-level executor. It can be expressed as: ; Where, For the current moment Meta-strategy parameters of ; is the meta-strategy mapping function; is the parameter set of the meta-strategy mapping function; Preferably, in this embodiment, the upper meta-controller constructs an objective function to maximize long-term risk suppression and strategy stability, which can be expressed as: ; Where, is the objective function of the upper-level meta-controller; To perform the sliding window mean operation; Plan your steps for the long term; is the time slice interval; is the discount factor; is the cost weight coefficient; The moment of the upper meta-controller Instant rewards; is the variance, Constraint strategy shocks; Preferably, in this embodiment, the immediate reward is directly associated with the dynamic risk score change, encouraging strategies to reduce risk, which can be expressed as: ; ; Where, For the moment Instant rewards; The magnitude of the reduction for the dynamic risk score; For the moment Dynamic risk scoring at the time of For the moment Dynamic risk scoring at the time of To punish the violent fluctuation coefficient of the score; For the moment Actions implementation costs; is the learning rate; S42. The lower-layer executor makes action decisions in real time using the DuelingDQN algorithm, taking the state variables consisting of dynamic risk scores, real-time operation sequences, and environmental context as inputs, and outputs state values ​​and action advantages, which can be expressed as: ; ; Where, To be in state Take action Expected cumulative rewards to be obtained; is the state value function; is the parameter set of the state value function; is the action advantage function; is the parameter of the action advantage function; To respond to the action set; Action Advantage Set Any action in is a state variable; The meta-strategy parameters of the upper meta-controller Adjust the expected cumulative reward and express it as follows: ; Where, is the adjusted expected cumulative reward; Preferably, in this embodiment, the immediate reward of the lower-level executor focuses on short-term risk suppression and action costs, which can be expressed as follows: ; Where, Provide immediate rewards for lower-level executors; Reward for continuous action; S43. Based on the dynamic risk score and the expected cumulative reward output by the lower-level actuator, the hierarchical protection response action is driven collaboratively. Specifically: S431. Determine the hierarchical protection response level by threshold determination based on the dynamic risk score (coarse-grained decision): when When it is determined to be a low-risk state, a three-level response is carried out, and the hierarchical protection response action is to trigger lightweight encrypted communication; when When the risk is determined to be medium, a secondary response is carried out, and the graded protection response action is to initiate a biometric authentication review; when When it is determined to be a high-risk state, a level one response is carried out, and the hierarchical protection response action is to perform fuse isolation; in, and are the preset lower and upper risk thresholds respectively; S432. Within the hierarchical protection response level divided by the threshold judgment, the expected cumulative reward output by the lower-level executor is used to optimize the action execution details. For example, in the case of lightweight encrypted communication at the third level response, the encryption algorithms include AES-CTR and GCM. If the network delay is high, that is, abnormal, The low-overhead AES-CTR is preferred; biometric authentication review during secondary response, if the environment is noisy ( abnormal), Increase the weight of facial recognition; fuse isolation during the first-level response, if the dynamic risk score suddenly changes ( , is the preset dynamic risk score mutation threshold), Choose to expand quarantine to a region; Preferably, this implementation also includes calculating the reverse impact of the dynamic risk score on the execution effect of the hierarchical protection response action, which is expressed as follows: ; Where, for Dynamic risk scoring at all times; For the general The global spatiotemporal feature matrix at each moment is used as the input of the spatiotemporal dual-channel neural network to output the dynamic risk score; express Momentary action The expected risk suppression amount is learned from historical data; if the action effectively suppresses the risk, then Lower, on the contrary it may increase; S5. This embodiment also includes building a data security protection system for the entire life cycle of entry and exit self-service equipment, including: S51. Basic equipment security: Choose self-service entry and exit equipment that has passed security certification. To ensure the safety of the equipment, perform a comprehensive security assessment and vulnerability scan. The security assessment is conducted from multiple dimensions, including the hardware security of the equipment, the stability of the operating system, the compliance of software applications, etc.; while the vulnerability scan uses professional scanning tools to detect the network ports, service programs, etc. of the equipment to find possible security vulnerabilities; once potential risks are discovered, they are repaired in a timely manner. At the same time, in order to protect the data stored and transmitted by the device, it is encrypted based on the AES advanced encryption standard. As a conventional encryption standard, AES has high security and efficiency, and can effectively prevent data from being stolen or tampered with during storage and transmission. S52. Sensitive Data Access Control: After ensuring basic device security, attribute-based encryption (ABE) is implemented on sensitive data stored in self-service immigration devices using a cloud database. Attribute-based encryption (ABE) is an encryption method based on the attributes of the accessed object. Different encryption keys are generated based on the sensitivity of the data and the attributes of the accessed object. To implement this, the sensitive attributes of the data and the relevant attributes of the accessed object are first determined, and then an access control policy is generated based on these attributes.

[0025] During the encryption process, the hash value of the encrypted data is calculated. The blockchain smart contract compares the hash table. If the encrypted hash value does not exist, the encrypted data and the hash value are stored on the chain to ensure data immutability and traceability. If the encrypted hash value exists, only the hash reference is uploaded to achieve data deduplication and reduce storage space.

[0026] When an access request is triggered, the smart contract verifies the attributes of the access object. Only when the attributes of the access object meet the pre-set access control policy will access to the key material be authorized and the access object will be decrypted using the private key; otherwise, access is prohibited and the abnormal access object is marked for further analysis. S53. Real-time Security Monitoring: To promptly detect abnormal device operations, extract the operation logs of self-service entry and exit devices, including login, query, and modification operations, along with corresponding timestamps and access object IDs. Establish a real-time security monitoring system to detect abnormal behavior using pre-set criteria for abnormal operation patterns. For example, frequent login attempts within a short period of time or sensitive data queries during non-business hours may be considered abnormal operations.

[0027] A streaming computing engine is used to analyze log data in real time. Once a risk event is discovered, an early warning is triggered immediately, and the relevant information is recorded in the blockchain audit trail to facilitate subsequent auditing and tracing. S54. Data Backup and Recovery: To prevent data loss, develop a comprehensive backup plan that combines daily full backups with weekly incremental backups, storing data in a remote data center. Additionally, periodically copy the database to a removable storage device to achieve multiple data backups.

[0028] Preferably, the backup data is encrypted before storage and decrypted by the key management system during restoration to ensure the security and availability of the backup data; S55. Cloud Database Deployment and Dynamic Protection: The cloud database is deployed in the remote service center. It includes an identification query module and an interaction module, supporting multi-user command operations and data tag storage. When a self-service entry / exit device is connected, the system automatically matches the device model and dynamically adapts protection strategies based on the device's characteristics and security requirements.

[0029] The operating data of the self-service entry and exit devices will be uploaded to the cloud database in real time. Through big data comparison and analysis, the protection level will be continuously optimized, so that the protection strategy can adapt to the ever-changing security threats. S56. Dynamic Update of Protection Strategies: To maintain the effectiveness and adaptability of the protection system, an incremental update mechanism is employed. Optimally, only data changes are recorded, the hash value is recalculated, and the hash table is updated after a blockchain deduplication check. Meta-learning techniques are combined to continuously optimize the strategy model, and adversarial training generators are used to generate adversarial examples, improving the protection system's ability to resist attacks. Furthermore, data from each adjustment is recorded to provide data support for subsequent optimizations, driving the autonomous evolution of the protection strategy.

[0030] In the embodiment of the present invention, "at least one" refers to one or more, and "more" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent the existence of A alone, the existence of A and B at the same time, and the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b and c can be represented by: a, b, c, a and b, a and c, b and c, or a, b and c, where a, b, c can be single or multiple.

[0031] Those skilled in the art will appreciate that the various units and algorithm steps described in the embodiments disclosed herein can be implemented using a combination of electronic hardware, computer software, and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0032] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0033] In the several embodiments provided by the present invention, if any function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory; hereinafter referred to as: ROM), random access memory (Random Access Memory; hereinafter referred to as: RAM), magnetic disk or optical disk, and other media that can store program code.

[0034] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention's description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. The information protection method of entry and exit self-service equipment based on deep federated learning is characterized by: The method comprises: Obtain device-related data from all entry-exit self-service devices within the protected area, including operation logs, network traffic, and biometric data. Extract local features from the device-related data using a pre-set feature extraction model. The local features of entry-exit self-service devices are encrypted and aggregated through a federated learning framework to construct a global spatiotemporal feature matrix. Constructing a spatiotemporal dual-channel neural network, comprising a temporal channel and a spatial channel. The temporal channel uses dilated convolution to extract cross-device collaborative operation patterns, and the spatial channel uses a graph attention network to model inter-device dependencies. The global spatiotemporal feature matrix is ​​used as input to the spatiotemporal dual-channel neural network, outputting a dynamic risk score. The operation sequence and environmental context risks of the entry and exit self-service equipment are collected in real time, and the dynamic risk score is input into the pre-built risk assessment engine based on deep reinforcement learning. Combined with the real-time collected operation sequence and environmental context risks, the decision-making process is triggered, and the risk assessment engine outputs graded response protection actions.

2. The method for protecting information of entry-exit self-service equipment based on deep federated learning according to claim 1 is characterized in that: The preset feature extraction model includes an operation log feature extraction module, a network traffic feature extraction module, and a biometric feature extraction module, wherein: The structure of the operation log feature extraction module includes a bidirectional LSTM layer and an attention pooling layer. The operation sequence is used as the input of the operation log feature extraction module, and the operation log features are output, including the operation type, time interval and access object fields. The network traffic feature extraction module calculates network traffic features based on the traffic statistics window, including the number of packets per second, protocol type distribution and traffic entropy value; The biometric feature extraction module uses a ResNet-34 model trained with the ArcFace loss function, takes biometric data as input, and outputs biometric features, including fingerprint features and facial features; The operation log features, network traffic features and biometric features are spliced ​​together to obtain local features.

3. The method for protecting information of entry-exit self-service equipment based on deep federated learning according to claim 2 is characterized in that: The federated learning framework is used to encrypt and aggregate the local features of the entry-exit self-service devices and construct a global spatiotemporal feature matrix, specifically: Calculate the local gradient of the preset feature extraction model corresponding to each entry and exit self-service device and encrypting the local gradient using an encryption algorithm; The entry and exit self-service device uploads the encrypted local gradient to the central server. The central server aggregates all encrypted local gradients within the protection area to obtain the global gradient, which is expressed as: ； Where, is the global gradient; The total number of self-service entry and exit devices; After encryption The local gradient of the preset feature extraction model corresponding to each entry and exit self-service device; is the modulus in the encryption algorithm; For the The local gradient of the preset feature extraction model corresponding to each entry and exit self-service device; The central server distributes the global gradient to each entry-exit self-service device. The entry-exit self-service device decrypts and updates the preset feature extraction model based on the global gradient. The decryption update is expressed as follows: ； in: ； ； Where, Extract the gradient of the model for the updated preset features; To perform decryption operations; is the Carmichael function value; To determine and The least common multiple of and Both are prime numbers in encryption algorithms; is the decryption coefficient; Transformation function for information decoding; It is the generator in the encryption algorithm; Based on the preset feature extraction model after decryption and update, feature extraction is performed on the device-related data to obtain local features, and the local features are encrypted using an encryption algorithm, which can be expressed as follows: ； ； ； Where, After encryption The first self-service entry and exit equipment dimensional local features; For the The first self-service entry and exit equipment dimensional local features; For the The first self-service entry and exit equipment Dimensional local features correspond to random numbers; is the total dimension of local features; is the set of real numbers; According to the time slice and the spatial distribution of entry and exit self-service equipment, the encrypted local features are aggregated into a global spatiotemporal feature matrix, which can be expressed as follows: ； ； Where, is the global spatiotemporal feature matrix; For the The self-service entry and exit equipment is The time slice dimensional local features; is the number of time slices.

4. The method for protecting information of entry-exit self-service equipment based on deep federated learning according to claim 3 is characterized in that: The time channel uses dilated convolution to extract the cross-device collaborative operation mode specifically as follows: Slice the global spatiotemporal feature matrix by device dimension to obtain the temporal feature sequence of the entry and exit self-service equipment , expressed as: ； The first The time series feature sequence of the entry and exit self-service equipment is input into the expansion convolution layer of the time channel, which is expressed as follows: ； ； Where, For the Layer dilation convolution layer in time slice The output value of is the number of dilated convolutional layers; is the convolution kernel size; For the The weight parameters of the layer-expanded convolutional layer; For the Layer dilation: The dilation rate of the convolution layer controls the spacing of the convolution kernel elements; is the index of the convolution kernel element; For the The output value of the layer expansion convolution layer in the time slice, is the initial layer; The output of the dilated convolutional layer is processed by the LSTM network to obtain the temporal hidden state, which can be expressed as: ； Where, Time slice The hidden state of For the previous time slice The hidden state of The attention weight is calculated based on the temporal hidden state and is expressed as: ； ； Where, Time slice The attention weight of is a learnable parameter; is the transpose operation; is the hidden state weight matrix; is the context weight matrix; is the global context vector; is a nonlinear enhancement function; The weighted temporal features are obtained based on the attention weight calculation, which can be expressed as follows: 。 5. The method for protecting information of entry-exit self-service equipment based on deep federated learning according to claim 4 is characterized in that: The spatial channel models the inter-device dependencies through the graph attention network: Slice the global spatiotemporal feature matrix according to the time slice dimension to obtain the global device state sequence of the time slice , expressed as: ； Taking entry-exit self-service devices as nodes and the interaction relationships between entry-exit self-service devices as edges, we construct an entry-exit self-service device association graph. The initial node feature is the global device state sequence, and the edge weight is expressed as: ； Where, For nodes and nodes The weight of the connecting edge; is the total number of communications between nodes; is the number of communications between nodes and nodes; To perform cosine similarity calculation; Dynamically calculate nodes using graph attention network and nodes The attention coefficient is expressed as follows: ； Where, is the attention coefficient of node and node; is the attention vector, which determines the strength of the association between nodes; 、 and Node 、 and Node characteristics; 、 and It is a trainable weight used to map the corresponding node feature dimension; Representation node For nodes Neighbors; For nodes The set of neighbors of By weighted aggregation of node neighborhood features, node aggregation features are generated, which can be expressed as: ； Where, For nodes Aggregation characteristics of is a nonlinear activation function, which in this embodiment is ELU (exponential linear unit) to retain negative information; is the trainable aggregation weight; Representation node For nodes Neighbors; Perform maximum pooling on the node aggregation features to obtain spatial features , expressed as: 。 6. The method for protecting information of entry-exit self-service equipment based on deep federated learning according to claim 5 is characterized in that: The global spatiotemporal feature matrix is ​​used as the input of the spatiotemporal dual-channel neural network, and the output dynamic risk score is specifically: The time channel of the spatiotemporal dual-channel neural network is used to obtain weighted time series features, and the space channel of the spatiotemporal dual-channel neural network is used to obtain spatial features. The weighted time series features and spatial features are fused to obtain fused features, which can be expressed as follows: ； Where, It is a fusion feature; is the fusion weight; is the fusion bias; The dynamic risk score is calculated based on the fusion features and is expressed as follows: ； Where, dynamic risk scoring; is a learnable risk weight; For risk bias.

7. The method for protecting information of entry-exit self-service equipment based on deep federated learning according to claim 6 is characterized in that: The upper meta-controller receives the historical dynamic risk score window and the real-time operation sequence and environment context, and generates the meta-strategy hidden state through LSTM, which is expressed as: ； Where, For the current moment Meta-strategy The hidden state of As of the current moment Dynamic risk scoring window; For the current moment Real-time operation sequence The sliding window mean of ; For the current moment Environmental context The sliding window mean of ; is the parameter set of LSTM; The meta-strategy hidden state is mapped to the meta-strategy parameter, which is used to regulate the action selection preference of the lower-level executor. It can be expressed as: ； Where, For the current moment Meta-strategy parameters of ; is the meta-strategy mapping function; is the parameter set of the meta-strategy mapping function; The upper meta-controller constructs the objective function to maximize long-term risk suppression and strategy stability, which can be expressed as: ； ； ； Where, is the objective function of the upper-level meta-controller; To perform the sliding window mean operation; Plan your steps for the long term; is the time slice interval; is the discount factor; is the cost weight coefficient; The moment of the upper meta-controller Instant rewards; is the variance, Constraint strategy shocks; For the moment Instant rewards; The magnitude of the reduction for the dynamic risk score; For the moment Dynamic risk scoring at the time of For the moment Dynamic risk scoring at the time of To punish the violent fluctuation coefficient of the score; For the moment Actions implementation costs; is the learning rate.

8. The method for protecting information of entry-exit self-service equipment based on deep federated learning according to claim 7 is characterized in that: The lower-layer executor makes action decisions in real time through the DuelingDQN algorithm, taking the state variables consisting of dynamic risk scores, real-time operation sequences, and environmental context as input, and outputs state values ​​and action advantages, which can be expressed as follows: ； ； Where, To be in state Take action Expected cumulative rewards to be obtained; is the state value function; is the parameter set of the state value function; is the action advantage function; is the parameter of the action advantage function; To respond to the action set; Action Advantage Set Any action in is a state variable; The meta-strategy parameters of the upper meta-controller Adjust the expected cumulative reward and express it as follows: ； Where, is the adjusted expected cumulative reward; The immediate reward of the lower-level executor is expressed as follows: ； Where, Provide immediate rewards for lower-level executors; Reward for continuous action.

9. The method for protecting information of entry-exit self-service equipment based on deep federated learning according to claim 8 is characterized in that: Based on the dynamic risk score and the expected cumulative reward output by the lower-level actuator, the hierarchical protection response actions are driven collaboratively. Specifically: The threshold is determined based on the dynamic risk score to determine the hierarchical protection response level: when When it is determined to be a low-risk state, a three-level response is carried out, and the hierarchical protection response action is to trigger lightweight encrypted communication; when When the risk is determined to be medium, a secondary response is carried out, and the graded protection response action is to initiate a biometric authentication review; when When it is determined to be a high-risk state, a level one response is carried out, and the hierarchical protection response action is to perform fuse isolation; in, and are the preset lower and upper risk thresholds respectively; Within the hierarchical protection response levels divided by threshold judgment, the action execution details are optimized through the expected cumulative rewards output by the lower-level executor.

10. The method for protecting information of entry-exit self-service equipment based on deep federated learning according to claim 9 is characterized in that: The method further includes calculating the reverse impact of the dynamic risk score on the performance of the hierarchical protection response action, which is expressed as follows: ； Where, for Dynamic risk scoring at all times; For the general The global spatiotemporal feature matrix at each moment is used as the input of the spatiotemporal dual-channel neural network to output the dynamic risk score; express Momentary action The expected risk suppression amount.

Images