Geological disaster multi-institution privacy data collaborative investigation method based on federal learning
By using LSTM-GCN and Transformer-XGBoost models under the federated learning framework, combined with differential privacy and homomorphic encryption techniques, the problems of information leakage and model accuracy in collaborative training of geological disaster data are solved, and efficient and secure multi-agency collaborative geological disaster prediction is achieved.
Patent Information
- Application Number
- CN202511096512.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-06
- Publication Date
- 2025-11-21
AI Technical Summary
Existing technologies pose a risk of data-sensitive information leakage in geological disaster prediction. Furthermore, traditional federated learning struggles to capture the spatiotemporal correlation of geological data, leading to decreased model prediction accuracy. Additionally, the lack of verification mechanisms against malicious entities impacts model robustness.
A collaborative investigation method for privacy-preserving geological disaster data from multiple institutions is adopted based on federated learning. An initial model is built through a central server, while institutions collect and preprocess data locally. The model is trained using a hybrid LSTM-GCN network and a Transformer-XGBoost ensemble model. Gradients are processed using differential privacy and homomorphic encryption. The central server performs weighted averaging and credibility verification to ensure that the data does not leave the institution.
It enables collaborative training of privacy-preserving data among multiple institutions, protects sensitive information, improves the model's prediction accuracy and robustness, provides early warning of geological disasters, and ensures data security and model reliability.
Smart Images
Figure CN120996741A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of geological disasters, and particularly relates to a geological disaster multi-agency private data collaborative investigation method based on federated learning. BACKGROUND
[0002] Current geological disaster prediction mainly relies on data collaborative analysis of multiple agencies (geological, meteorological, surveying and mapping departments, etc.). Traditional centralized training requires the aggregation of raw data to a central server, which has the risk of sensitive geological information leakage (such as terrain data, disaster point coordinates). Although the existing federated learning scheme realizes local training of data, when dealing with spatio-temporal correlation data of geological disasters, the general federated framework is difficult to capture the spatio-temporal correlation of geological data (such as rainfall trend, crust displacement continuity), resulting in a decrease in model prediction accuracy; no verification mechanism is designed for malicious agencies to submit invalid / forged gradients (such as sensor fault data), affecting the robustness of the model. SUMMARY
[0003] Therefore, the present application provides a geological disaster multi-agency private data collaborative investigation method based on federated learning to solve the problems in the prior art.
[0004] In order to achieve the above-mentioned purpose, the present application provides the following technical scheme:
[0005] The geological disaster multi-agency private data collaborative investigation method based on federated learning comprises the following steps:
[0006] Step one, the central server constructs an initial geological disaster prediction model and generates global model parameters; each agency registers identity information with the central server and obtains a secure communication key;
[0007] Step two, each agency collects geological disaster-related raw data locally, pre-processes the raw data and extracts features;
[0008] Step three, federated cycle training is repeated until convergence;
[0009] Step four, the central server weights and averages the encrypted gradients according to the data volume of each agency or a preset trust weight, generates a global aggregated gradient, updates the global model parameters using the aggregated gradient to form a new version of the global model, verifies the trustworthiness, triggers an exception handling protocol when the verification fails, and enters the next round of model distribution when the verification is successful;
[0010] Step five, the central server distributes the updated global model to all participating agencies; return to step three to start a new round of federated training until the model converges or reaches a predetermined training round.
[0011] Further, all the raw data is only stored locally and does not leave the internal network of the agency.
[0012] Further: the process of each training cycle in step three is as follows:
[0013] (1) Download the global model
[0014] At the beginning of each round of training, each institution downloads the current latest global model parameters from the central server to the local;
[0015] (2) Local model training
[0016] Based on the downloaded global model, the institution uses local private data to perform local model training and calculate model gradients;
[0017] (3) Gradient obfuscation processing
[0018] The local gradient is processed for privacy protection, including differential privacy processing and homomorphic encryption processing;
[0019] Differential privacy processing is to add noise to the gradient to prevent the gradient from leaking original data information; homomorphic encryption processing is to encrypt the gradient to ensure that the uploaded content cannot be reversed to the original data;
[0020] (4) Upload encrypted gradient
[0021] The processed encrypted gradient is uploaded to the central server for global model update.
[0022] Further: the global model uses an LSTM-GCN hybrid network model;
[0023] Wherein, the calculation formula of the objective function is:
[0024]
[0025] Wherein, W is the global model parameter matrix (LSTM-GCN weight); K is the total number of institutions participating in training, is the local loss function of client k; λ is the L2 regularization coefficient;
[0026] The parameter update formula is as follows:
[0027]
[0028] Agg is the abbreviation of Aggregation, which refers to the operation of the central server to calculate the weighted average of encrypted gradients from multiple clients; η g is the global model learning rate;
[0029] The gradient aggregation formula is:
[0030]
[0031] where, is the encrypted gradient with added noise; D k is the local dataset of client k; t is the training round index.
[0032] Further: the local model adopts a Transformer-XGBoost integrated model;
[0033] Then the gradient calculation formula of the model is:
[0034]
[0035] The parameter update formula is:
[0036]
[0037] where, V k is the local model parameter (Transformer weight) of client k; f g is the global model function (LSTM-GCN); f l is the local model function (Transformer feature extractor); x i , y i are geological disaster samples and labels, respectively; is the loss function; η l is the local model learning rate; s is the number of local training iteration steps.
[0038] Further: the encrypted gradient with added noise is:
[0039]
[0040] The spatiotemporal correlation noise is:
[0041]
[0042] where, N(μ,Σ) is a Gaussian noise matrix (spatiotemporal correlation); σ 2 is the noise intensity coefficient; t i , s i are the timestamp and spatial coordinates of sample i, respectively; τ t , τ s are the time and space decay coefficients, respectively; ||·|| is the spatiotemporal distance metric.
[0043] The guarantee of differential privacy is:
[0044]
[0045] where, Δ2 is the gradient sensitivity, and δ is the failure probability.
[0046] Further: the local model extracts the spatio-temporal feature zi, wherein the calculation formula of zi is zi=Transformer-XGBoost(V k ,x i ); the global model processes the cross-region feature During gradient transmission, only the global model gradient is uploaded The local parameter V k Never leave the device.
[0047] Further: the credibility verification is that the server detects gradient anomaly, judges whether the mechanism exists malicious submission or data failure; the gradient anomaly includes that the gradient value is too large or too small, and the statistical distribution deviation.
[0048] The present application has the following advantages: the present application realizes the collaborative optimization of privacy and efficiency by fusing the spatio-temporal feature protection mechanism in the federated learning framework. The cross-region geological features are aggregated by the LSTM-GCN global model, and the high-dimensional spatio-temporal features are extracted by the Transformer-XGBoost local model; the spatio-temporal feature collaborative modeling provides a technical paradigm for the multi-institutional privacy and security collaboration.
[0049] Other features and advantages of the present application will be described in the following description, and some will become apparent from the description, or will be learned from the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS
[0050] In order to more intuitively illustrate the prior art and the present application, the following exemplary drawings are given. It should be understood that the specific shape, structure shown in the drawings should not be regarded as the limiting conditions for realizing the present application; for example, based on the technical concept disclosed in the present application and the exemplary drawings, those skilled in the art can easily make routine adjustments or further optimization to some units (components) such as increase / decrease / attribute division, specific shape, positional relationship, connection mode, size ratio relationship, etc.
[0051] Figure 1 The implementation flowchart of the multi-institutional privacy data collaborative investigation method based on federated learning provided by an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION
[0052] The following embodiments of the present application are illustrated by way of specific examples, and other advantages and effects of the present application will be readily appreciated by those skilled in the art upon reading the forward disclosure. Obviously, the described embodiments are only a part of embodiments of the present application, rather than all the embodiments. It should be understood that these embodiments are only used to further explain the present application, and cannot be understood as limiting the protection scope of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.
[0053] Please refer to Figure 1 , the geological disaster multi-agency private data collaborative investigation method based on federated learning includes the following steps:
[0054] Step one, system initialization
[0055] The central server constructs an initial geological disaster prediction model, generates global model parameters, and serves as the starting point for federated learning.
[0056] Each agency (such as the geological bureau, the meteorological bureau, and the surveying and mapping institute) registers identity information with the central server, obtains a secure communication key, and ensures the encryption and authentication of subsequent data transmission.
[0057] The initial geological disaster prediction model uses the LSTM-GCN model; the LSTM network captures the time dependence of geological disasters, such as rainfall trends and displacement changes. The graph convolution network (GCN) aggregates spatial correlation features across regions, such as geological structures and the influence of adjacent disaster points.
[0058] The central server updates the global parameters by weighted averaging of the encrypted gradients of each agency.
[0059] Differential privacy and homomorphic encryption are used for privacy protection; differential privacy adds spatiotemporal related noise (noise intensity decays with spatiotemporal distance) to prevent gradient reverse engineering of original data. Homomorphic encryption ensures that the uploaded gradient cannot be decrypted.
[0060] Local data and model parameters never leave the agency.
[0061] In addition, the central server detects gradient anomalies (such as value range anomalies and distribution deviations), and intercepts malicious / inactive data submission
[0062] Step two, local data processing
[0063] Each agency collects geological disaster related data (such as terrain, rainfall, and soil sensor data) locally, performs data cleaning (removes outliers), normalization (unifies dimensions), and feature extraction (such as terrain slope and rainfall intensity features).
[0064] In addition, all original data is only stored locally and does not leave the internal network of the institution.
[0065] Step three, federal cycle training, repeat until convergence (iterative execution); the process of each training cycle is as follows:
[0066] (1) Download the global model
[0067] At the beginning of each round of training, each institution downloads the current latest global model parameters from the central server to the local.
[0068] The global model uses an LSTM-GCN hybrid network model, where LSTM is a long short-term memory network and GCN is a graph convolution.
[0069] The calculation formula of the objective function is as follows:
[0070]
[0071] Where W is the global model parameter matrix (LSTM-GCN weight); K is the total number of participating training institutions, is the local loss function of client k; λ is the L2 regularization coefficient.
[0072] The parameter update formula is as follows:
[0073]
[0074] Agg is the abbreviation of Aggregation, which refers to the operation of the central server to calculate the weighted average of encrypted gradients from multiple clients; η g is the global model learning rate.
[0075] The gradient aggregation formula is:
[0076]
[0077] Where, is the encrypted gradient after adding noise; D k is the local data set of client k; t is the training round index.
[0078] (2) Local model training
[0079] Based on the downloaded global model, the institution uses local private data to train the local model, calculates the model gradient, for example, using the stochastic gradient descent (SGD) optimization algorithm.
[0080] In this embodiment, the local model uses a Transformer-XGBoost integrated model.
[0081] The gradient calculation formula of the model is:
[0082]
[0083] The parameter update formula is:
[0084]
[0085] where V k is the local model parameter (Transformer weight) of the client k; f g is the global model function (LSTM-GCN); f l is the local model function (Transformer feature extractor); x i , y i are geological disaster samples and labels, respectively; is the loss function; η l is the local model learning rate; s is the number of local training iterations.
[0086] (3) Gradient confusion processing
[0087] The local gradient is processed for privacy protection, which includes differential privacy processing and homomorphic encryption processing.
[0088] Differential privacy processing is to add noise to the gradient to prevent the gradient from leaking original data information; homomorphic encryption processing is to encrypt the gradient to ensure that the uploaded content cannot be reversed to the original data.
[0089] where the encrypted gradient with noise added is:
[0090]
[0091] The spatiotemporal correlation noise is:
[0092]
[0093] where N(μ,Σ) is a Gaussian noise matrix (spatiotemporal correlation); σ 2 is the noise intensity coefficient; t i , s i are the timestamp and spatial coordinates of sample i, respectively; τ t , τ s are the time and space decay coefficients, respectively; ||·|| is the spatiotemporal distance measure.
[0094] The guarantee of differential privacy is:
[0095]
[0096] where Δ2 is the gradient sensitivity, and δ is the failure probability.
[0097] (4) Upload encrypted gradients
[0098] Upload the processed encrypted gradients to the central server for global model updating.
[0099] In this embodiment, the local model extracts spatio-temporal features zi, where the calculation formula for zi is: zi = Transformer-XGBoost(V k ,x i ).
[0100] Global model processes cross-regional features
[0101] During gradient transmission, only upload global model gradients Local parameters V k Never leave the device.
[0102] In this embodiment, the advantage of weighted aggregation is that institutions with large amounts of data (such as provincial geological bureaus) have higher weights; but remote monitoring stations (small samples) can still contribute to region-specific rules.
[0103] For example: XX village belongs to monitoring station C jurisdiction, input the following data into the model, 24h rainfall is 185mm (over alert value), slope displacement is 12.7mm / 24h, soil moisture content is 43%.
[0104] The prediction process is as follows: Transformer extracts rainfall pulse features, XGBoost outputs local risk probability of 78%, LSTM-GCN fuses regional geological data, and outputs final probability of 92%. The result is to issue a red alert 42 hours in advance, successfully avoiding risks.
[0105] Step four, global aggregation and verification
[0106] (1) Aggregation server weighted average; the central server weights the encrypted gradients according to the data volume or pre-set trusted weight of each institution (uses FedAvg algorithm) to generate global aggregated gradients.
[0107] (2) Update the global model; use the aggregated gradients to update the global model parameters to form a new version of the global model.
[0108] (3) Trustworthiness verification; the server detects gradient anomalies (such as excessively large / small gradient values, statistical distribution deviation), and determines whether the institution has malicious submission or data failure.
[0109] When the verification fails, trigger the exception handling protocol (such as excluding abnormal institutions, starting the audit mechanism); when the verification passes, enter the next round of model distribution.
[0110] Step five, model distribution and iteration
[0111] (1) Distribute new global model; the central server distributes the updated global model to all participating agencies.
[0112] (2) Iterative training; return to step three and start a new round of federated training until the model converges (performance meets standards) or reaches the predetermined training round.
[0113] The present application protects the privacy and security of the data of multiple agencies in the process of geological disaster prevention. Specifically, multiple agencies (geology, meteorology, surveying and mapping) train collaboratively under the premise of not sharing original data, avoiding the risk of sensitive information (such as disaster point coordinates, terrain data) leakage.
[0114] The local model captures fine-grained spatio-temporal regularity (such as rainfall pulse, slope displacement); the global model integrates cross-regional geological correlation (such as stratigraphic continuity, disaster chain effect). Solving the problem of insufficient modeling of spatio-temporal characteristics of geological disasters in traditional federated learning. It is beneficial to increase the accuracy of geological disaster prediction.
[0115] Abnormal gradient detection eliminates fault / forged data interference and improves model robustness; the federated framework supports real-time dynamic updating of the model, helping to provide early warning in high-risk areas; for example, XX village successfully avoids risks.
[0116] The present application realizes the cooperative optimization of privacy protection and prediction efficiency, and can also provide a safe and reliable technical paradigm for multi-agency geological disaster joint defense.
[0117] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present application shall be included in the protection scope of the present application.
Claims
1. A collaborative investigation method for multi-agency privacy data on geological disasters based on federated learning, characterized in that: Includes the following steps: Step 1: The central server constructs an initial geological disaster prediction model and generates global model parameters; each institution registers its identity information with the central server and obtains a secure communication key. Step 2: Each institution collects relevant raw data on geological disasters locally, preprocesses the raw data, and extracts features; Step 3: Federated loop training, repeated until convergence; Step 4: The central server performs a weighted average of the encryption gradients based on the data volume or preset trust weights of each institution to generate a global aggregate gradient; the global model parameters are updated using the aggregate gradient to form a new version of the global model; and the trustworthiness is verified; when the verification fails, an exception handling protocol is triggered; when the verification passes, the next round of model distribution begins. Step 5: The central server distributes the updated global model to all participating institutions; return to Step 3 to start a new round of federated training until the model converges or reaches the predetermined number of training rounds.
2. The collaborative investigation method for multi-agency privacy data of geological disasters based on federated learning according to claim 1, characterized in that, All the raw data described herein will be stored locally and will not leave the organization's internal network.
3. The collaborative investigation method for multi-agency privacy data of geological disasters based on federated learning according to claim 1, characterized in that, The process for each training cycle in step three of the federated loop training is as follows: (1) Download the global model At the start of each training round, each institution downloads the latest global model parameters from the central server to its local machine; (2) Local model training The organization trains the model locally using local privacy data based on the downloaded global model and calculates the model gradient. (3) Gradient obfuscation processing Privacy protection measures are applied to local gradients, including differential privacy processing and homomorphic encryption. Differential privacy involves adding noise to the gradient to prevent it from leaking the original data information; homomorphic encryption encrypts the gradient to ensure that the uploaded content cannot be reversed to reconstruct the original data. (4) Upload the encryption gradient The processed encrypted gradients are uploaded to the central server for global model updates.
4. The collaborative investigation method for multi-agency privacy data of geological disasters based on federated learning according to claim 1, characterized in that, The global model adopts a hybrid LSTM-GCN network model; The objective function is calculated using the following formula: Where W is the global model parameter matrix (LSTM-GCN weights); K is the total number of institutions participating in the training. Let be the local loss function for client k; λ be the L2 regularization coefficient. The parameter update formula is as follows: Where t is the training round index. This refers to the collection of cryptographic gradients uploaded by all participating institutions; Agg is short for Aggregation, specifically referring to the operation of a central server calculating a weighted average of cryptographic gradients from multiple clients; η g The global model learning rate; The gradient aggregation formula is: in, The encryption gradient after adding noise; D k t represents the local dataset of client k; t is the training round index.
5. The collaborative investigation method for multi-agency privacy data of geological disasters based on federated learning according to claim 1, characterized in that, The local model uses the Transformer-XGBoost ensemble model; The gradient calculation formula for this model is: The parameter update formula is: Among them, V k f represents the local model parameters (Transformer weights) for client k; g For the global model function (LSTM-GCN); f l For the local model function (Transformer feature extractor); x i y i These are geological disaster samples and labels; Let η be the loss function; l is the local model learning rate; s is the number of local training iterations.
6. The collaborative investigation method for multi-agency privacy data of geological disasters based on federated learning according to claim 1, characterized in that, The encryption gradient after adding noise is: Spatiotemporal correlated noise is: Where N(μ,Σ) is the Gaussian noise matrix (spatiotemporal correlation); σ 2 t is the noise intensity factor; i s i These are the timestamp and spatial coordinates of sample i, respectively; τ t ,τ s These are the attenuation coefficients in time and space, respectively; ||·|| is the measure of spatiotemporal distance. The guarantees of differential privacy are: Where Δ2 is the gradient sensitivity and δ is the failure probability.
7. The collaborative investigation method for multi-agency privacy data of geological disasters based on federated learning according to claim 1, characterized in that, The local model extracts spatiotemporal features zi, where zi is calculated using the formula: zi = Transformer - XGBoost(V k ,x i Global models handle cross-regional features. During gradient propagation, only the global model gradient is uploaded. Local parameter V k Never leave the device.
8. The collaborative investigation method for multi-agency privacy data of geological disasters based on federated learning according to claim 1, characterized in that, Trustworthiness verification involves the server detecting gradient anomalies to determine whether the organization has submitted maliciously or the data is invalid. Gradient anomalies include gradient values that are too large or too small, and statistical distribution deviations.