Laboratory data security protection system fusing federated learning and reinforcement learning

Abstract

The invention relates to the technical field of network security and distributed artificial intelligence, and discloses a federated learning and reinforcement learning fused laboratory data security protection system, which comprises an edge intelligent agent, a federated collaborative server and an encrypted communication network, the edge intelligent agent dynamically adjusts a strategy network structure according to data complexity by using a manifold topology resonance module, generates a virtual sample to assist training through an adversarial course enhancement module, and separates a local private gradient from a shared gradient by using a gradient orthogonal decomposition module; the federated collaborative server is responsible for gradient weighted aggregation and global attention parameter distribution based on node reputation. According to the method, sample scarcity is made up through geometric manifold adaptation heterogeneous data features, orthogonal projection cutting gradient privacy association and adversarial generation, high-privacy cooperative training in a distributed environment is realized, the problems of poor model adaptability and privacy leakage in laboratory data protection are solved, and the security and adaptability of a system are improved.

Description

Technical Field

[0001] This invention relates to the fields of network security and distributed artificial intelligence technology, specifically a laboratory data security protection system that integrates federated learning and reinforcement learning. Background Technology

[0002] Laboratory data security protection systems are mainly used to ensure the confidentiality and integrity of high-value experimental data generated by scientific research instruments. Federated learning is a distributed machine learning framework that allows participants to collaboratively train a global model without sharing local raw data. Reinforcement learning is a computational method that optimizes decision-making strategies through trial and error between agents and the environment. The distributed security architecture that combines federated learning and reinforcement learning aims to achieve dynamic defense against heterogeneous network attacks and privacy protection through collaborative computing of each node.

[0003] Existing distributed security technologies typically deploy a uniformly structured neural network model across various edge computing nodes. These edge nodes independently calculate model gradients or parameter updates based on locally collected traffic data, and then transmit the results to a central server via an encrypted communication channel for aggregation. To mitigate privacy risks, some solutions employ encryption protocols at the transport layer and introduce noise mechanisms into the aggregation algorithm. At the defense strategy execution level, the system utilizes a policy network to receive local network state input, iteratively optimizes algorithms to adjust actions for handling abnormal traffic, and finally receives global parameters distributed by the central server to synchronize the defense strategy.

[0004] However, due to the fundamental differences in the types of instruments and experimental projects equipped in different laboratories, the communication data generated by each node exhibits statistical heterogeneity in the feature space distribution. This causes existing fixed-structure models to be unable to adaptively adapt to this differentiated data geometric complexity, leading to underfitting or overfitting problems in local models. At the same time, since simple parameter encryption does not mathematically sever the correlation between gradients and original data, attackers can still reconstruct the original experimental samples by analyzing the uploaded gradients. Furthermore, due to the lack of a reputation evaluation mechanism for participating nodes and the scarcity of local real attack samples, the global model is highly susceptible to poisoning by malicious nodes and fails, resulting in insufficient generalization ability of reinforcement learning policy networks when facing new threats. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a laboratory data security protection system that integrates federated learning and reinforcement learning. It aims to solve the problems that existing fixed model structures cannot adapt to the characteristics of heterogeneous laboratory data, gradient transmission is prone to leaking the privacy of original data, and the model adaptability is poor due to the scarcity of samples and malicious nodes.

[0006] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a laboratory data security protection system that integrates federated learning and reinforcement learning. This system can achieve access control and anomaly detection of laboratory data by collaboratively training security policies through distributed edge computing nodes without directly transmitting the original experimental data.

[0007] The system includes: an encrypted communication network, an edge intelligent agent, and a federated collaborative server.

[0008] The encrypted communication network establishes an encrypted transmission tunnel between the edge intelligent agent and the federated collaborative server, maintaining a persistent secure session connection for laboratory data security.

[0009] The edge intelligent agent is deployed on the computing nodes of the laboratory data processing unit to perform local reinforcement learning training tasks. During the data processing phase, the edge intelligent agent receives global attention distribution parameters through an encrypted communication network, uses these parameters to determine the weighted regions for feature extraction, and performs weighted processing on the laboratory data to generate feature fingerprint vectors.

[0010] Meanwhile, the agent generates network structure adjustment instructions based on the distribution characteristics of the data (such as the local topological dimension calculated using manifold learning), dynamically changes the target layer depth and width of the reinforcement learning policy network, and freezes the propagation of redundant neurons through a mask matrix, so that the network capacity adaptively matches the geometric complexity of the current data.

[0011] During the model training and update phases, edge intelligent agents generate virtual samples to assist in the training of reinforcement learning policy networks.

[0012] Specifically, the agent uses the feature fingerprint vector as a seed, combines it with random noise to generate counterfactual samples, and constrains the direction of variation of these samples relative to the real samples to be consistent with the direction of the global gradient residual vector, thereby covering the discrimination blind spots of the global model. After calculating the original update gradient, the agent performs a gradient decoupling operation: it constructs a Riemannian manifold tangent space using local neighborhood samples, projects the original gradient onto the tangent space to obtain private gradient components representing local specific knowledge, and extracts shared gradient components orthogonal to the local data distribution. The agent only sends the shared gradient components to the federated collaboration server, mathematically isolating gradient information containing local privacy features.

[0013] The federated collaboration server, acting as the server node for federated learning, receives shared gradient components from multiple edge intelligent agents via an encrypted communication network. The server maintains a reliability metric for each node, which is dynamically adjusted based on the cosine similarity between the shared gradient components and the global update direction. During aggregation, the server performs a weighted average of the valid shared gradient components according to the reliability metric and updates the global policy parameters.

[0014] Furthermore, the server uses a validation dataset containing both attack and normal samples to calculate and normalize the gradient of the loss function with respect to the input, generating a global gradient residual vector. Simultaneously, it parses the updated global model weights, calculates the global importance score for each feature dimension, converts it into a probability distribution, and generates global attention distribution parameters. These control signals are fed back to the edge to guide the next round of training.

[0015] Furthermore, the aforementioned edge intelligent agent also includes blockchain-based fingerprint verification logic. The agent calls the blockchain notarization interface to obtain the baseline fingerprint corresponding to the laboratory data, calculates the difference metric (such as normalized Hamming distance) between the current feature fingerprint vector and the baseline fingerprint, and constructs a joint state vector together with the fingerprint's topological embedding. The reinforcement learning policy network outputs access control actions based on this state vector, including direct access, sandbox cleaning, blocking alarm, or de-identified synthesis.

[0016] The aforementioned encrypted communication network integrates traffic shaping and priority scheduling functions. The network unit parses the service type identifier in the encrypted frame header, marks control signaling containing global gradient residual vectors and global attention distribution parameters as high-priority queues for priority forwarding, and marks payloads containing model weight data as ordinary-priority queues, thereby ensuring the real-time performance of the control plane in federated collaborative training.

[0017] This invention provides a laboratory data security protection system that integrates federated learning and reinforcement learning. It has the following beneficial effects: 1. This invention performs gradient orthogonal decomposition at the edge intelligent agent, utilizing manifold learning theory to separate the original update gradient of the reinforcement learning policy network into private gradient components and shared gradient components. Only the shared gradient components, orthogonal to the local data distribution, are sent to participate in federated aggregation, while the private gradient components, containing local specific information, are retained locally. This approach mathematically cuts off the path for attackers to infer the characteristics of the original experimental data in the laboratory by analyzing the uploaded gradients, helping to solve the privacy leakage problem in the federated learning collaborative training process and ensuring the security of sensitive laboratory data.

[0018] 2. This invention utilizes a manifold topological resonance mechanism to dynamically adjust the physical structure of the reinforcement learning policy network based on the local topological dimension of the laboratory data. By generating a mask matrix to freeze the forward and backward propagation of redundant neurons, the model capacity can adaptively match the geometric complexity of the data being processed. This mechanism helps avoid overfitting or underfitting of fixed-structure models when processing heterogeneous laboratory data. It also helps reduce the computational resource consumption and memory usage of edge gateway nodes, thereby improving the deployment efficiency of the security protection system on resource-constrained laboratory instruments.

[0019] 3. This invention combines adversarial course enhancement with a reliability-based federated aggregation mechanism. It uses the global gradient residual vector calculated by the server to constrain the generation direction of local virtual samples, enabling edge nodes to generate high-value adversarial samples targeting the weaknesses of the global model. Combined with a weighted aggregation strategy based on node reliability indicators, the system can automatically reduce the weight of abnormal or malicious nodes, so that the laboratory data security protection model can still maintain convergence stability when facing sample scarcity or malicious poisoning attacks, thus improving the system's robustness and generalization ability in identifying new network attack behaviors. Attached Figure Description

[0020] Figure 1 This is a system architecture diagram in an embodiment of the present invention; Figure 2 This is a block diagram of the internal functional modules of the edge intelligent agent in an embodiment of the present invention; Figure 3 This is a block diagram of the internal functional modules of the federated collaboration server in an embodiment of the present invention.

[0021] Among them, 100 is the edge intelligent agent; 101 is the adaptive fingerprint extraction module; 102 is the manifold topology resonance module; 103 is the adversarial curriculum enhancement module; 104 is the dynamic policy decision module; 105 is the gradient orthogonal decomposition module; 200 is the federated collaborative server; 201 is the reputation weighted aggregation module; 202 is the global residual calculation module; 203 is the attention distribution analysis module; and 300 is the encrypted communication network. Detailed Implementation

[0022] The technical solutions in 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.

[0023] See attached document Figure 1-3 This invention provides a laboratory data security protection system that integrates federated learning and reinforcement learning, including an edge intelligent agent 100, a federated collaborative server 200, and an encrypted communication network 300.

[0024] Edge intelligent agent 100 is deployed at the gateway nodes of laboratory instruments, and federated collaborative server 200 is deployed in the central data center. Edge intelligent agent 100 and federated collaborative server 200 establish a connection through encrypted communication network 300.

[0025] The edge intelligent agent 100 is used to perform feature perception, dynamic access control decision-making, and preliminary processing of model parameters for scientific research data on a local laboratory end. The edge intelligent agent 100 includes an adaptive fingerprint extraction module 101, a manifold topological resonance module 102, an adversarial curriculum enhancement module 103, a dynamic policy decision-making module 104, and a gradient orthogonal decomposition module 105.

[0026] The adaptive fingerprint extraction module 101 is connected to the instrument's network communication interface. The adaptive fingerprint extraction module 101 is configured to acquire raw communication data generated by the instrument and determine the weighted region for feature extraction based on the received global attention distribution parameters. The adaptive fingerprint extraction module 101 extracts features from the weighted region to generate a semantic fingerprint vector, simultaneously combining environmental metadata to generate an environmental feature vector, and calculates the Hamming distance between the semantic fingerprint vector and a preset reference fingerprint.

[0027] The manifold topology resonance module 102 is connected to the adaptive fingerprint extraction module 101. The manifold topology resonance module 102 maintains the Riemannian manifold space of local historical data and maps the semantic fingerprint vector at the current moment into this Riemannian manifold space. The manifold topology resonance module 102 calculates the local topological dimension of the location of the semantic fingerprint vector and generates network structure adjustment instructions for adjusting the physical structure of the neural network based on this local topological dimension.

[0028] The adversarial curriculum enhancement module 103 receives the global gradient residual vector from the federated collaboration server 200. Using the currently acquired semantic fingerprint vector as seed data, the adversarial curriculum enhancement module 103 generates counterfactual fingerprint samples using a generative adversarial network. During generation, the adversarial curriculum enhancement module 103 uses the global gradient residual vector as a directional constraint, ensuring that the feature variation direction of the counterfactual fingerprint samples aligns with the discriminative weakness direction of the global model. The generated counterfactual fingerprint samples are then injected into the experience replay pool.

[0029] The dynamic policy decision module 104 is connected to the adaptive fingerprint extraction module 101, the manifold topology resonance module 102, and the adversarial curriculum enhancement module 103. Responding to network structure adjustment instructions, the dynamic policy decision module 104 dynamically adjusts the layer depth and width of the reinforcement learning policy network by activating or quiescent neurons. The dynamic policy decision module 104 constructs a state vector containing Hamming distance and fingerprint topology embedding, inputs it to the adjusted policy network, and outputs access control actions. The dynamic policy decision module 104 also calculates the original gradient of the policy network based on environmental feedback and samples in the experience replay pool.

[0030] The gradient orthogonal decomposition module 105 is connected to the dynamic policy decision module 104. The gradient orthogonal decomposition module 105 is configured to construct the tangent space basis of the Riemannian manifold space. The gradient orthogonal decomposition module 105 decomposes the original gradient projection into parallel components parallel to the tangent space and perpendicular components perpendicular to the tangent space. The gradient orthogonal decomposition module 105 filters out the parallel components and extracts the perpendicular components as data to be uploaded to the federated collaboration server 200.

[0031] The federated coordination server 200 is used to coordinate the aggregation of model parameters and the synchronization of global defense strategies among multiple edge intelligent agents 100. The federated coordination server 200 includes a reputation-weighted aggregation module 201, a global residual calculation module 202, and an attention distribution analysis module 203.

[0032] The reputation-weighted aggregation module 201 receives vertical components from each edge intelligent agent 100 through the encrypted communication network 300. Based on preset node reputation weights, the reputation-weighted aggregation module 201 performs weighted aggregation calculations on all received vertical components and updates the global policy parameters using the calculation results.

[0033] The global residual calculation module 202 is connected to the reputation weighted aggregation module 201. The global residual calculation module 202 evaluates the updated global policy parameters using a pre-set validation dataset, calculates the gradient of the loss function with respect to the input, and generates a global gradient residual vector.

[0034] The attention distribution analysis module 203 is used to analyze the input weight distribution of the global policy parameters. The attention distribution analysis module 203 extracts the input feature regions that contribute the most to the model's decision and generates a global attention heatmap. The federated collaboration server 200 distributes the updated global policy parameters, global gradient residual vector, and global attention heatmap to each edge intelligent agent 100.

[0035] The relevant modules of the system of the present invention and their specific implementation principles will be disclosed in detail below.

[0036] As a local computing unit deployed on the laboratory instrument side, the Edge Intelligent Agent 100 is configured with an embedded gateway, an edge server, or a board integrated into the instrument control computer.

[0037] The Edge Intelligent Agent 100 achieves data interaction between its internal modules through an internal high-speed bus or memory sharing mechanism. The Edge Intelligent Agent 100 integrates high-performance computing units, including but not limited to GPUs (Graphics Processing Units) or NPUs (Neural Processing Units), to accelerate the inference and training processes of deep learning models.

[0038] The adaptive fingerprint extraction module 101 is primarily responsible for multimodal feature perception and authenticity verification of scientific research data. The adaptive fingerprint extraction module 101 intercepts the raw data stream output by the instrument in real time through the data acquisition interface. For different types of scientific instruments, the raw data stream includes image data from an electron microscope, waveform data from a mass spectrometer, or sequence data from a sequencer. The adaptive fingerprint extraction module 101 has a pre-built feature extractor based on a deep neural network, which may include a multilayer convolutional neural network (CNN) or a long short-term memory network (LSTM).

[0039] During feature extraction, the adaptive fingerprint extraction module 101 receives global attention distribution parameters from the federated collaboration server 200. Global attention distribution parameters It is a weight vector that matches the input dimension of the feature extractor, representing regions or frequency bands in the data sample that have high discriminative value under the current global threat situation (e.g., specific byte segments in the file header or high-frequency components of waveforms).

[0040] The adaptive fingerprint extraction module 101 divides the raw data into For each feature region, the feature responses of each region are weighted element-wise using global attention distribution parameters to calculate the weighted semantic feature vector. The calculation formula is as follows: ; In the formula, The generated semantic feature vector; It is a non-linear activation function; This represents the total number of feature regions. For the first The global attention weights corresponding to each feature region are determined by... The result was obtained through analysis; This is the weight matrix of the projection layer, used to map high-dimensional features to a low-dimensional semantic space; For the first The original feature vectors of each feature region; This is the bias term for the projection layer.

[0041] Simultaneously, the adaptive fingerprint extraction module 101 collects the instrument's current operating environment data, including timestamps, sensor readings (such as temperature and voltage), user IDs, and system logs. It then performs one-hot encoding or normalization on this discrete data to generate an environmental feature vector. .

[0042] In addition, the adaptive fingerprint extraction module 101 is configured to call the blockchain notarization interface to obtain the base fingerprint corresponding to the data through the hash index of the data. The baseline fingerprint is the original feature value stored on the blockchain at the moment of data generation, serving as a benchmark for verifying data integrity. The adaptive fingerprint extraction module 101 calculates the semantic feature vector. Compared with the benchmark fingerprint Normalized Hamming distance between Hamming distance is used to quantify the degree of difference between the current data content and the original evidence content. The calculation formula is as follows: In the formula, The normalized Hamming distance takes values ​​in the range [0,1]. The bit length of the fingerprint vector; The index of the vector bit; The first binary segment of the semantic feature vector Bit value; For the benchmark fingerprint The value of the digit; This represents the XOR operation.

[0043] The adaptive fingerprint extraction module 101 uses the Hamming distance The output is sent to the dynamic policy decision module 104 as a state input, and the larger the value... Values ​​typically indicate the risk of data tampering or contamination.

[0044] The manifold topology resonance module 102 is primarily responsible for quantifying the geometric complexity of the data and dynamically adjusting the network structure accordingly. The manifold topology resonance module 102 maintains a fixed-length historical fingerprint buffer (e.g., the most recent 1000 records) to construct a Riemannian manifold space capable of characterizing the local data distribution. The manifold topological resonance module 102 utilizes historical fingerprint data in the buffer pool, employs the k-nearest neighbor algorithm to construct a local adjacency graph, and estimates the current semantic feature vector based on the maximum likelihood estimation method. Local topological dimension in manifold space Local topological dimension This reflects the inherent structural complexity of current scientific research data. The higher the dimension, the more complex the data pattern, requiring a deeper network structure to fit it.

[0045] Manifold topological resonance module 102 is based on local topological dimension Generate network structure adjustment instructions. These instructions are used to define the reinforcement learning policy network. The physical scale at the current moment, including network layer depth. and layer width The specific mapping relationship is as follows: ; ; In the formula, The layer depth of the policy network; The layer width of the policy network; For local topological dimensions; This is the layer depth scaling factor; This is the layer depth offset constant; This is the scaling factor for the layer width; This is the layer width offset constant.

[0046] The manifold topology resonance module 102 sends a network structure adjustment command to the dynamic policy decision module 104. The dynamic policy decision module 104 generates a mask matrix corresponding to the topology dimension to adjust the network structure beyond the specified dimensions. and The connection weights of neurons within a certain range are reset to zero (i.e., their forward and backward propagation is frozen), thereby achieving dynamic reconstruction of the network structure and adaptation of computing power.

[0047] The adversarial curriculum enhancement module 103 is primarily responsible for addressing the cold start and sample scarcity issues in reinforcement learning. Internally, the adversarial curriculum enhancement module 103 deploys a generative adversarial network (GAN), which includes a generator. and discriminator The adversarial course enhancement module 103 receives the global gradient residual vector issued by the federated collaboration server 200. The global gradient residual vector is the gradient vector obtained by evaluating the global model on the validation set, indicating the feature direction where the global model has the largest current discrimination error. The adversarial course enhancement module 103 uses the current true semantic feature vector... As a conditional seed, combined with a random noise vector Using generators Generate counterfactual fingerprint samples To enable the generated samples to specifically attack the weaknesses of the global model, the adversarial course enhancement module 103 in the generator... The loss function introduces a direction constraint term, and the total loss function The definition is as follows: ; In the formula, This represents the total loss value of the generator; To generate the standard adversarial loss term for adversarial networks; The weight hyperparameters for the direction constraint term; For the generated counterfactual fingerprint sample; The true semantic feature vector; This is the global gradient residual vector.

[0048] The direction constraint term forces the generated counterfactual samples to maintain the same direction of mutation relative to the true samples as the global gradient residual vector. The adversarial curriculum enhancement module 103 will then use the generated counterfactual fingerprint samples... As a virtual state, and based on preset expert rules (such as blocking if the fingerprint distance is too large), corresponding virtual actions and virtual rewards are generated, encapsulated as experience tuples, and injected into the experience replay pool of the dynamic strategy decision module 104.

[0049] The dynamic policy decision module 104 is configured as the core computing unit for executing access control and runs a decision algorithm based on deep reinforcement learning. The dynamic policy decision module 104 constructs the current joint state vector in real time. This joint state vector is composed of user behavior features, device environment features, normalized Hamming distance input from the adaptive fingerprint extraction module 101, and topological embedding vector calculated by the manifold topological resonance module 102.

[0050] The dynamic policy decision module 104 inputs the joint state vector into the reinforcement learning policy network after structural adjustment using a mask matrix. This reinforcement learning policy network employs a multilayer perceptron structure, and the effective connection weights between its neurons are controlled by a mask matrix generated by the manifold topology resonance module 102 to adapt to the current data complexity. The reinforcement learning policy network outputs a probability distribution for a preset action space. The dynamic policy decision module 104 selects and executes the target access control action based on this probability distribution. The preset action space specifically includes: Direct access means allowing data to pass through the gateway; Sandbox cleaning involves redirecting data to an isolated environment for deep testing; alarm blocking involves cutting off the connection and sending alarm information to the administrator; and data masking and synthesis involves differential privacy processing of sensitive fields in the data.

[0051] The dynamic policy decision module 104 collects environmental feedback rewards after actions are executed. The reward function is designed to comprehensively consider security and efficiency: a large negative reward is given if a data breach occurs or the system is flagged by an intrusion detection system (IDS); a positive reward is given if data is transmitted normally and no false alarms are triggered; and a small negative reward (efficiency penalty) is given if a time-consuming sandbox cleaning action is performed but no threat is found. The original update gradient .

[0052] The gradient orthogonal decomposition module 105 is primarily responsible for decoupling local specific knowledge from global general knowledge. The gradient orthogonal decomposition module 105 acquires the fingerprint manifold data maintained by the manifold topological resonance module 102 and extracts the current semantic feature vector. of The nearest neighbor samples form a local neighborhood matrix.

[0053] The gradient orthogonal decomposition module 105 uses the singular value decomposition (SVD) method to decompose the local neighborhood matrix, selecting the first... The right singular vectors corresponding to the maximal singular values ​​form the basis matrix VV of the tangent space. The column vectors span the tangent space The gradient orthogonal decomposition module 105 updates the original gradient. Projective decomposition, calculating parallel components parallel to the tangent space. and the perpendicular component to the tangent space The calculation formula is as follows: ; ; In the formula, These are the parallel components of the gradient; Let be the basis matrix of the tangent space; This is the transpose of the basis matrix; This is the original update gradient; This represents the vertical component of the gradient; It is an identity matrix.

[0054] parallel component Characterizes specific knowledge (such as instrument-specific noise patterns) that is only valid in the tangential direction of the local data manifold and is used only to update the local model. The gradient orthogonal decomposition module 105 extracts the vertical component. It characterizes a general logical rule orthogonal to the local data distribution for the vertical components. Homomorphic encryption and compression are performed, and the data is sent to the federated collaboration server 200 via the encrypted communication network 300.

[0055] In this embodiment, the adaptive fingerprint extraction module 101, in conjunction with the global attention mechanism, enables the perception layer to dynamically adapt to the threat situation, ensuring the effectiveness of feature extraction. The manifold topological resonance module 102 allows the neural network structure to adaptively scale with the geometric complexity of the data, reducing energy consumption when processing simple data and releasing computing power when processing complex data, thus optimizing resource allocation. The targeted generation of the adversarial curriculum enhancement module 103 solves the training problem of reinforcement learning when abnormal samples are scarce, improving the robustness of the model. The mathematical projection of the gradient orthogonal decomposition module 105 rigorously separates local specific data from global general logic at the geometric level, achieving federated co-evolution while providing strict privacy protection for sensitive laboratory data.

[0056] The Federated Collaboration Server 200 is deployed in the central data center. As a data processing device based on federated learning and reinforcement learning architecture, the hardware carrier of the Federated Collaboration Server 200 is configured as a high-performance rack server or server cluster. It has network interface cards that support data transmission rates of 10 Gigabit or higher and large-capacity memory resources to support the parallel computing needs of large-scale federated learning tasks.

[0057] The federated collaboration server 200 establishes a network connection with the edge intelligent agents 100 distributed in various laboratories through the encrypted communication network 300. The federated collaboration server 200 is configured to issue global defense strategy parameters and receive local model gradients uploaded by the edge intelligent agents 100, and realize the iteration of the security protection model based on the parameter aggregation update mechanism. The federated collaboration server 200 specifically includes a reputation weighted aggregation module 201, a global residual calculation module 202, and an attention distribution analysis module 203 in its logical functional architecture.

[0058] The reputation-weighted aggregation module 201 is primarily responsible for implementing weighted aggregation of multi-source gradients within the federated learning architecture to suppress the impact of gradient data uploaded by abnormal nodes on global model parameters. The reputation-weighted aggregation module 201 is configured to receive vertical gradient components uploaded from multiple edge intelligent agents 100.

[0059] To address the varying data quality across nodes in a laboratory network environment, the reputation-weighted aggregation module 201 maintains a dynamically updated list of node reputations. Upon receiving the first... Round communication Gradient vertical components uploaded by 100 edge intelligent agents Then, the reputation-weighted aggregation module 201 first uses a preset private key to analyze the gradient vertical components. Decryption is performed, and the digital signature is verified to confirm the integrity of the data source. Subsequently, the reputation-weighted aggregation module 201 calculates the cosine similarity between the vertical component of the gradient and the direction of the previous global update, serving as the basis for evaluating the quality of this gradient. The reputation-weighted aggregation module 201 then calculates the gradient quality based on the node's current reputation value. The global policy parameters are updated by performing a weighted average of all valid vertical gradient components. The updated formula is as follows: ; In the formula, The updated global policy parameters; These are the global policy parameters at the current moment; The learning rate for global federated learning; The total number of edge intelligent agents participating in this round of aggregation is 100; For the first The node at the th The reputation value of the wheel; For the first The vertical components of the gradient uploaded by each node.

[0060] Meanwhile, the reputation-weighted aggregation module 201 dynamically updates the reputation value of the nodes based on the quality assessment results of this gradient. The update of the reputation value follows the exponential moving average rule, and the calculation formula is as follows: ; ; In the formula, The updated node reputation value; For the sake of reputation, history is a factor of forgetting; To correct the linear unit function; The cosine similarity function; This is the arithmetic mean vector of all uploaded gradients in this round.

[0061] Through this mechanism, the reputation-weighted aggregation module 201 reduces the weight of nodes with negative cosine similarity in subsequent federated aggregations.

[0062] The global residual calculation module 202 is primarily responsible for calculating the loss gradient information of the global model and transforming this information into constraints that guide the edge reinforcement learning agent in adversarial training. Internally, the global residual calculation module 202 stores a pre-built standard validation dataset, which includes labeled attack sample data and normal research sample data.

[0063] In global strategy parameters After the update is complete, the global residual calculation module 202 inputs the validation dataset into the updated global model and calculates the loss value between the predicted output and the true label. The global residual calculation module 202 is configured to calculate the gradient of the loss function with respect to the input data, which reflects the direction in which the input features should change to increase the loss function value. The global residual calculation module 202 averages and normalizes the input gradients of all validation samples to generate a global gradient residual vector. The calculation formula is as follows: ; In the formula, This is the global gradient residual vector; To verify the total number of samples in the dataset; For the first Input features of each validation sample; For the first The true label of the verification sample; The loss function; Indicates input data Find the partial derivative.

[0064] Global gradient residual vector It will be distributed to the edge intelligent agent 100 and used as a directional constraint for generating counterfactual samples during the training process of reinforcement learning.

[0065] The attention distribution analysis module 203 is primarily responsible for extracting the feature weight distribution of the global model, enabling the edge intelligent agent 100 to simultaneously focus on high-weight feature regions. The attention distribution analysis module 203 analyzes the updated global policy parameters... This involves obtaining the weight magnitude of each feature dimension in the first-layer weight matrix of the model. A larger weight magnitude indicates that the feature dimension has a higher contribution to the decision logic of the global model.

[0066] The attention distribution analysis module 203 calculates the global importance score for each feature dimension and uses the Softmax function to transform it into a probability distribution form of global attention distribution parameters. The calculation formula is as follows: ; In the formula, For the global attention distribution parameters corresponding to the first... The components of each feature dimension; This represents the number of neurons in the first hidden layer of the global model. Connecting the first layer weight matrix of the global model to the first layer weight matrix The input feature and the first The weight values ​​of each hidden neuron; This represents the total number of dimensions of the input features.

[0067] The attention distribution analysis module 203 will generate global attention distribution parameters. Distributed to each edge intelligent agent 100 for weighted feature extraction by the adaptive fingerprint extraction module 101.

[0068] In this embodiment, the federated collaborative server 200 establishes a gradient aggregation mechanism based on node reputation through the reputation-weighted aggregation module 201, which reduces the interference of abnormal gradient data on the convergence of the global model; it extracts the input gradient features of the global model through the global residual calculation module 202, providing adversarial constraints based on a global perspective for the reinforcement learning agent at the edge; and it realizes dynamic synchronization of the feature attention region through the weight parsing of the attention distribution analysis module 203.

[0069] The encrypted communication network 300 is configured to establish a data transmission link between the edge intelligent agent 100 and the federated collaboration server 200. The encrypted communication network 300 is built on existing internet infrastructure or a dedicated local area network for laboratories. The physical layer hardware of the encrypted communication network 300 includes high-performance routers, switches, firewall devices, and fiber optic or 5G wireless communication modules. The encrypted communication network 300 employs encrypted tunneling technology based on TLS 1.3 (Transport Layer Security) and is configured to establish a fully forward-secure encrypted session between the edge intelligent agent 100 and the federated collaboration server 200.

[0070] The encrypted communication network 300 integrates a key negotiation and authentication unit. This unit is configured to generate a temporary shared key unique to the current session during the communication connection establishment phase, using an elliptic curve-based Diffie-Hellman key exchange algorithm.

[0071] The encrypted communication network 300 is pre-installed with digital certificates issued by a trusted certificate authority. The key negotiation and authentication unit verifies the legitimacy of the communication peer by verifying the signature chain of the digital certificate. The key negotiation and authentication unit calculates the session master key based on the temporary public keys exchanged between the two parties. The calculation formula is as follows: ; In the formula, The generated session master key; This is a key derivation function based on HMAC; is a generator of the elliptic curve; Temporary private random numbers generated for Edge Intelligent Agent 100; Temporary private random numbers generated for the Federation Collaboration Server 200; It is a large prime number; For pre-master key; Generate a random number for the client; This is a random number generated by the server.

[0072] because and The session master key is regenerated in each session and is not persistently stored. It has forward security.

[0073] The encrypted communication network 300 further includes a secure payload encapsulation unit. This unit is configured to encapsulate the gradient vertical components uploaded by the edge intelligent agent 100 and the global policy parameters issued by the federated collaboration server 200 into frames. The secure payload encapsulation unit uses the AES-GCM (Advanced Encryption Standard – Galois Counter Mode) algorithm to perform authentication and encryption processing on the data payload. AES-GCM uses a counter mode for encryption and generates a Message Authentication Code (MAC) using Galois field multiplication. The secure payload encapsulation unit constructs the encrypted frame structure for the transport layer. The construction logic is as follows: ; In the formula, For the final transport layer encrypted frame to be sent; The authentication data header includes the frame sequence number, protocol version number, and service type identifier; Initialize the vector; For encryption and authentication functions; The data payload to be transmitted; This is the session encryption key derived from the session master key.

[0074] The function's output contains ciphertext. With certification label Verify the authentication label. The receiving end can detect whether the data has been tampered with during transmission.

[0075] In addition, the encrypted communication network 300 is equipped with a traffic shaping and priority scheduling unit. The traffic shaping and priority scheduling unit parses the encrypted frame header. The traffic shaping and priority scheduling unit uses the service type identifier to distinguish between reinforcement learning control signaling and federated learning weight data. The traffic shaping and priority scheduling unit will include the global gradient residual vector. With global attention distribution parameters Data packets are mapped to high-priority transmission queues to ensure real-time feedback for adversarial course training; the traffic shaping and priority scheduling unit maps data packets containing model weight update data to ordinary priority transmission queues and uses a sliding window mechanism for congestion control.

[0076] In this embodiment, the encrypted communication network 300 establishes a trusted communication link between the edge and the cloud through a key negotiation mechanism based on ECDHE, ensuring that historical session data will not be decrypted even if the private key is leaked over a long period of time; the confidentiality and integrity of data are protected during transmission through the encapsulation of the AES-GCM algorithm; and the efficiency of mixed traffic transmission, which combines large-volume transmission of federated learning data with high-frequency control signaling of reinforcement learning, is optimized through hierarchical scheduling of traffic shaping and priority scheduling units.

Claims

1. A laboratory data security protection system integrating federated learning and reinforcement learning, characterized in that, include: An encrypted communication network is configured to establish a persistent, secure session connection between the edge intelligent agent and the federated collaborative server for the protection of laboratory data. An edge intelligent agent, configured as a computing node deployed in laboratory data processing, executes local reinforcement learning training tasks. It receives global attention distribution parameters and global gradient residual vectors through the encrypted communication network, performs weighted feature extraction on laboratory data to generate feature fingerprint vectors, dynamically adjusts the physical structure of the reinforcement learning policy network according to the data distribution characteristics, generates virtual samples to assist the training of the reinforcement learning policy network, and decouples the calculated original update gradients, separating private gradient components and shared gradient components, and outputs only the shared gradient components for participation in federated learning collaborative updates. The federated collaboration server, configured as a server node for federated learning, receives the shared gradient components from multiple edge intelligent agents through the encrypted communication network, maintains node reliability indices corresponding to each edge intelligent agent, performs federated weighted aggregation of the shared gradient components based on the node reliability indices to update global policy parameters, and uses the validation dataset to calculate the global gradient residual vector and parses the updated global policy parameters to generate the global attention distribution parameters.

2. The laboratory data security protection system integrating federated learning and reinforcement learning according to claim 1, characterized in that, The edge intelligent agent includes an adaptive fingerprint extraction module, which is configured as follows: The weighted region for feature extraction is determined using the global attention distribution parameters, and the laboratory data is subjected to element-wise weighting to generate a semantic fingerprint vector as the feature fingerprint vector. The benchmark fingerprint corresponding to the laboratory data is obtained by calling the blockchain notarization interface. The normalized Hamming distance between the semantic fingerprint vector and the benchmark fingerprint is calculated, and the normalized Hamming distance is used as a state input to the reinforcement learning policy network to represent the risk of data tampering.

3. The laboratory data security protection system integrating federated learning and reinforcement learning according to claim 1, characterized in that, The edge intelligent agent includes a manifold topology resonance module, which is configured as follows: Maintain the Riemannian manifold space composed of local historical data, map the current feature fingerprint vector to the Riemannian manifold space, and use the maximum likelihood estimation method to calculate the local topological dimension of the location of the feature fingerprint vector as the data distribution feature. Based on the local topology dimension, a network structure adjustment instruction is generated. The network structure adjustment instruction is used to define the target layer depth and target layer width of the reinforcement learning policy network at the current time. Generate a mask matrix corresponding to the local topological dimension, and set the neuron connection weights in the reinforcement learning policy network that exceed the target layer depth and target layer width to zero.

4. The laboratory data security protection system integrating federated learning and reinforcement learning according to claim 1, characterized in that, The edge intelligent agent includes an adversarial curriculum enhancement module, which is configured as follows: An internally deployed generative adversarial network is used to generate counterfactual fingerprint samples as virtual samples by combining the current feature fingerprint vector as seed data and combining it with random noise vectors. A directional constraint term is introduced into the loss function of the generator. The directional constraint term is configured to force the mutation direction of the counterfactual fingerprint sample relative to the real sample to be consistent with the direction of the global gradient residual vector, thereby generating adversarial samples that target the weaknesses of the global model. The counterfactual fingerprint sample is encapsulated as a virtual experience tuple and injected into the experience replay pool.

5. The laboratory data security protection system integrating federated learning and reinforcement learning according to claim 1, characterized in that, The edge intelligent agent includes a dynamic policy decision module, which is configured as follows: Call the blockchain evidence storage interface to obtain the baseline fingerprint corresponding to the laboratory data, calculate the normalized Hamming distance between the feature fingerprint vector and the baseline fingerprint, and construct a joint state vector containing the normalized Hamming distance and fingerprint topological embedding. The joint state vector is input into the structurally adjusted reinforcement learning policy network, which outputs access control actions for the laboratory data. The access control actions are selected from a preset action space, which includes direct access, redirection to an isolated environment for sandbox cleaning, disconnection alarm, and desensitization synthesis for differential privacy processing of sensitive fields; The original update gradient of the reinforcement learning policy network is calculated based on the environmental feedback rewards and samples in the experience replay pool.

6. The laboratory data security protection system integrating federated learning and reinforcement learning according to claim 3, characterized in that, The edge intelligent agent includes a gradient orthogonal decomposition module, which is configured as follows: The nearest neighbor samples of the feature fingerprint vector are extracted to form a local neighborhood matrix, and the tangent space basis of the Riemannian manifold space is constructed using the singular value decomposition method. The original update gradient is projected onto the tangent space spanned by the tangent space basis to obtain the parallel component as the private gradient component. The parallel component represents specific knowledge that is only valid in the tangent direction of the local data manifold. The vertical component is obtained by subtracting the parallel component from the original update gradient as the shared gradient component. The vertical component represents a general logical rule orthogonal to the local data distribution and is sent to the encrypted communication network.

7. The laboratory data security protection system integrating federated learning and reinforcement learning according to claim 1, characterized in that, The federated collaboration server includes a reputation-weighted aggregation module, which is configured as follows: The received shared gradient component is decrypted using a preset private key and the digital signature is verified. The cosine similarity between the shared gradient component and the previous round of global update direction is calculated. Based on the node reputation value as the node reliability index, a weighted average of all valid shared gradient components is used to update the global policy parameters. The node reputation value is dynamically updated according to the exponential moving average rule. If the cosine similarity is negative, the node reputation value is reduced to suppress the impact of abnormal gradients on the federated learning global model.

8. The laboratory data security protection system integrating federated learning and reinforcement learning according to claim 1, characterized in that, The federated collaboration server includes a global residual calculation module, which is configured as follows: The standard validation dataset containing both attack and normal samples is input into the global model corresponding to the updated global policy parameters, and the gradient of the loss function with respect to the input data is calculated. The input gradients of all validation samples are averaged and normalized to generate the global gradient residual vector, which indicates the direction in which the input features should change in order to increase the value of the loss function.

9. The laboratory data security protection system integrating federated learning and reinforcement learning according to claim 1, characterized in that, The federated collaboration server includes an attention distribution analysis module, which is configured as follows: Analyze the updated global policy parameters to obtain the weight magnitude of each feature dimension in the first-layer weight matrix of the model; Calculate the global importance score for each feature dimension, and use the Softmax function to convert the global importance score into a probability distribution form to generate the global attention distribution parameters, which are used to instruct the edge intelligent agent to simultaneously pay attention to feature regions with higher weight magnitudes.

10. The laboratory data security protection system integrating federated learning and reinforcement learning according to claim 1, characterized in that, The encrypted communication network is configured as follows: The key negotiation and authentication unit is configured to generate a temporary session shared key with forward security using an elliptic curve-based key exchange algorithm during the session initialization phase. The security payload encapsulation unit is configured to encapsulate the shared gradient components and the global policy parameters in a frame-by-frame manner using the Galois counter mode, thereby generating an encrypted frame containing an authentication tag. The traffic shaping and priority scheduling unit is configured to parse the service type identifier in the encrypted frame header, mark the control signaling containing the global gradient residual vector and the global attention distribution parameters as a high-priority queue for priority forwarding, and mark the payload containing model weight data as a normal-priority queue for congestion control.

Images