Intelligent ship data privacy protection method and system

Abstract

The invention discloses an intelligent ship data privacy protection method and system, and the method comprises the steps: S1, constructing a global model network structure in a central server of a ship system, and constructing a local model network structure at a local end of each ship; s2, acquiring local multi-source data of the intelligent ship; s3, performing initialization setting on the noise model of each ship; the method comprises the following steps: S1, acquiring local multi-source data, S2, acquiring features of the local multi-source data, S4, analyzing privacy sensitivity of each data dimension based on the acquired features of the local multi-source data, S5, modeling data gradient distribution by adopting a multi-modal Gaussian mixture model, S6, counting information based on local and global sensitivity, and S7, realizing structure and scale dual adaptive noise injection. When the method is used, the problem that in the prior art, due to the fact that data distribution of a differential privacy mechanism is not matched in the Non-IID federated learning environment, privacy protection and model performance of intelligent ship multi-source data are difficult to consider at the same time is solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent ship technology, and more specifically, to a method and system for protecting data privacy on intelligent ships. Background Technology

[0002] As artificial intelligence technology is increasingly applied in smart shipping, intelligent ships generate massive amounts of data in areas such as navigation control, energy management, equipment monitoring, and cargo transportation. This data includes sensitive information such as ship location, navigation parameters, equipment status, fuel consumption, and cargo information. This data is not only vast in quantity but also involves critical issues such as ship safety, operational efficiency, and equipment health. While the processing and utilization of this data have played a significant role in improving the efficiency and safety of the shipping industry, they have also brought challenges to data privacy and information security.

[0003] Information leakage, particularly during data transmission and model training, has become a core challenge for collaborative systems between intelligent ships and smart ports. In traditional data processing methods, ship data is often uploaded to a central server for analysis and training, inevitably posing a risk of data leakage. For example, without effective protection, real-time ship location data, navigation routes, and fuel consumption information could be accessed maliciously or misused. Therefore, in collaborative systems between intelligent ships and smart ports, protecting ship data privacy without affecting the effective use of data is a pressing technical challenge that needs to be addressed.

[0004] Federated Learning (FL), a typical privacy-preserving distributed learning framework, offers a new approach to solving the aforementioned problems. This framework allows multiple clients to train models locally, sharing only local gradients or model parameters with a central server for global aggregation. This enables multi-party collaborative modeling without leaking the original data, effectively reducing the risk of centralized data leakage, achieving localized processing of ship data, and protecting the security of ship privacy data.

[0005] While federated learning offers significant advantages in privacy protection, the intermediate parameters, local gradients, and update directions transmitted during training can be exploited by attackers to infer sensitive information about ships. First, federated learning requires frequent uploading of local model updates to a central server for global aggregation. This makes it possible for sensitive ship information to be leaked during communication, especially under conditions of frequent changes in ship data and long training periods, where communication latency and bandwidth limitations can further increase the risk of leakage. Second, due to differences in equipment, operating environment, and navigation status among each ship, the generated data is often non-independent and identically distributed (Non-IID), posing a challenge to the effectiveness of federated learning, particularly in balancing privacy protection and model performance.

[0006] Differential privacy, as a lightweight privacy protection mechanism, can be challenging. While DP-based privacy strategies enhance the security of federated learning systems, noise injection can distort the direction and magnitude of model parameter updates, reduce the effective utilization of information during training, and weaken model performance, especially in non-IID environments. Therefore, how to protect privacy while maintaining model accuracy remains a critical technical challenge.

[0007] Therefore, designing an adaptive noise injection mechanism that can effectively protect privacy while improving model accuracy has become a key issue that needs to be addressed in current intelligent ship data protection technologies. This invention proposes a dual adaptive differential privacy mechanism based on structure and scale, providing a more precise privacy protection scheme for the diversity, heterogeneity, and high sensitivity of intelligent ship data. By dynamically adjusting the structure and scale of the noise, this invention minimizes the impact of noise on the accuracy of the ship data model while ensuring privacy protection, thus effectively solving the privacy protection problem under the high security requirements of intelligent ships. Summary of the Invention

[0008] To address the aforementioned technical problems, the purpose of this invention is to overcome the data distribution mismatch in non-IID federated learning environments where existing differential privacy mechanisms suffer from difficulties in balancing privacy protection and model performance for multi-source data from intelligent ships. This invention provides a data privacy protection method for intelligent ships based on dual adaptive differential privacy of structure and scale. This method addresses the diversity, heterogeneity, and high sensitivity of data generated during intelligent ship operation. Considering the high information security requirements of ship data security, equipment monitoring, and port scheduling, it designs a differential privacy mechanism combining structural and scale adaptation, embedding it into the federated learning training process to achieve efficient and secure multi-party collaborative modeling for shipping scenarios.

[0009] To achieve the above objectives, the present invention provides a method for protecting the data privacy of intelligent ships, the method comprising:

[0010] Step S1: Construct a global model network structure in the central server of the ship system, and construct a local model network structure on the local end of each ship; wherein, the global model network structure and the local model network structure have the same structure.

[0011] Step S2: Acquire local multi-source data of the intelligent ship; wherein, the data includes location data, navigation parameters, equipment operating status, cargo information and business data;

[0012] Step S3: Initialize the noise model settings for each ship;

[0013] Step S4: Based on the characteristics of the acquired local multi-source data, analyze the privacy sensitivity of each data dimension, obtain the sensitivity value of each dimension, and store it in the local sensitivity analysis cache.

[0014] Step S5: A multimodal Gaussian mixture model is used to model the data gradient distribution, and the optimal number of modes and the position of the modal mean are determined using the Bayesian information criterion. The modal mean is set according to a symmetrical distribution during initialization to ensure that the disturbance noise meets the zero-mean unbiasedness requirement.

[0015] Step S6: Based on local and global sensitivity statistics, the client calculates the mixed sensitivity value. The client combines the local sensitivity and global sensitivity to calculate the mixed sensitivity value, and adjusts the noise scale accordingly.

[0016] Step S7: Based on the hybrid sensitivity value, the noise variance of each mode is calculated using a dynamic variance adjustment strategy, and the noise intensity is gradually reduced during training iterations using an exponential decay strategy until the set minimum noise variance threshold is reached, thereby achieving dual adaptive noise injection of structure and scale.

[0017] Preferably, after step S7, the method further includes:

[0018] Step S8: In order to cope with the possibility that changes in the ship's operating environment may cause gradient distribution to drift during long-term training, an update mechanism is adopted to recalculate the number of modes and the mean set in each round of calculation to ensure that the noise model always adapts to the current training data.

[0019] Step S9: Anonymize the model gradients or parameters injected with noise to generate perturbed model update data, and upload it to the federated learning aggregation server through an encrypted communication channel.

[0020] In step S10, the federated learning aggregation server receives disturbance model update data from multiple smart ships, merges the update data using an aggregation algorithm to generate a global model, and distributes the updated global model to each ship for continued local training.

[0021] Preferably, in step S3, the modality weights and variances are set symmetrically during initialization and dynamically adjusted during training.

[0022] Preferably, in step S4, the client calculates the mixing sensitivity value according to the following formula:

[0023]

[0024] in, For local sensitivity, λ represents the global sensitivity, and λ is the fusion weighting factor.

[0025] Preferably, in step S7, the variance adjustment coefficient is calculated based on the change in the current mixing sensitivity and the mixing sensitivity of the previous round, and the noise variance is reduced round by round in combination with the exponential decay strategy until the set minimum value is reached.

[0026]

[0027] Where κ is the scaling factor, which controls the amplification effect of changes in sensitivity on variance.

[0028] The present invention also provides an intelligent ship data privacy protection system, the system comprising:

[0029] The data acquisition module is used to acquire local data from the intelligent ship.

[0030] Sensitivity analysis module, used to perform sensitivity analysis based on the characteristics of the multi-source data;

[0031] The noise structure modeling module is used to model the gradient distribution of data based on a multimodal Gaussian mixture model and determine the optimal modal structure.

[0032] The hybrid sensitivity calculation module is used to calculate the hybrid sensitivity value by fusing local and global gradient statistics.

[0033] The noise injection module is used to dynamically adjust the noise variance based on the mixed sensitivity value and the noise attenuation strategy, and inject the noise into the model gradient corresponding to the multi-source data.

[0034] The parameter upload module is used to upload the perturbed model update parameters to the federated learning aggregation server;

[0035] The global model receiving module is used to receive the updated global model and apply it to local training.

[0036] Preferably, the mixed sensitivity calculation module calculates the mixed sensitivity value according to the following formula:

[0037]

[0038] in, For local sensitivity, λ represents the global sensitivity, and λ is the fusion weighting factor.

[0039] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a method for protecting data privacy of intelligent ships.

[0040] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a method for protecting data privacy of intelligent ships.

[0041] According to the above technical solution, the beneficial effects of the intelligent ship data privacy protection method and system provided by this invention are as follows: Under the same privacy budget conditions, this invention can effectively improve the adaptability between differential privacy noise and intelligent ship data distribution, achieving a balance between privacy protection and model performance in intelligent ship federated learning training. In complex shipping scenarios with highly heterogeneous data and limited privacy budgets, this method can still maintain high model accuracy and system stability, and has significant application value for intelligent ship navigation safety, equipment health management, and port collaborative optimization.

[0042] Other features and advantages of the present invention will be described in detail in the following detailed description section; and all parts not covered in the present invention are the same as or can be implemented using the prior art. Attached Figure Description

[0043] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the following detailed description to explain the invention, but do not constitute a limitation thereof. In the drawings:

[0044] Figure 1 This is a flowchart illustrating the steps of the intelligent ship data privacy protection method in an example of the present invention;

[0045] Figure 2 This is a schematic diagram of dynamic noise generation in an example of the present invention;

[0046] Figure 3 This is a bar chart showing the accuracy of the present invention in a simulation experiment;

[0047] Figure 4This is a structural block diagram of the intelligent ship data privacy protection system in an example of the present invention;

[0048] Figure 5 This is a schematic diagram of the structure of the electronic device in an example of the present invention. Detailed Implementation

[0049] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0050] Example 1:

[0051] like Figure 1-3 The present invention provides a method for protecting the data privacy of intelligent ships, the method comprising:

[0052] Step S1: Construct a global model network structure in the central server of the ship system, and construct a local model network structure on each ship's local terminal. The global model network structure is identical to the local model network structure. The parameters of the global model will be continuously updated during the training process of each client model, and finally merged into a global model through an aggregation algorithm.

[0053] Step S2: Acquire local multi-source data from the intelligent vessel. This data includes location data, navigation parameters, equipment operating status, cargo information, and other business-related data, which are highly privacy-sensitive. Therefore, before model training, sensitivity analysis and feature extraction are performed on this data to assess the impact of each data dimension on privacy protection, obtain the sensitivity value for each dimension, and store it in a local sensitivity cache. This process helps the system identify sensitive data, ensuring that stricter privacy protection measures are taken during model training.

[0054] Step S3: Initialize the noise model settings for each ship. First, preset the initial variance of each mode to σ. 2 The means form a symmetric set μ = {…, -b, -a, 0, +a, +b, ...}, where {a, b, ...} > 0 to ensure that the expected value of the noise perturbation is zero, satisfying the unbiasedness requirement in the differential privacy mechanism. The number of modes in the noise model is set, and by dynamically adjusting the weights and variances of each component, the mutual conversion between multimodal and single-modal modes is achieved during training, thereby enhancing the modeling capability for complex gradient structures.

[0055] Step S4: Based on the characteristics of the data, analyze the privacy sensitivity of each data dimension, obtain the sensitivity value for each dimension, and store it in the local sensitivity analysis cache. To improve the robustness of sensitivity estimation and effectively mitigate extreme value interference, the selection of the quantile threshold X directly affects the gradient clipping effect in local sensitivity estimation. This embodiment introduces a dynamic threshold adjustment strategy: in the early stages of training, a lower quantile can be used to strictly limit the gradient magnitude and prevent privacy leakage; as the model gradually converges, the quantile is gradually increased to retain more effective gradient information, thereby balancing the relationship between privacy and accuracy.

[0056]

[0057] in, Let be the gradient of the parameters of the i-th client at layer l.

[0058] Step 5: For the noisy structure, a multimodal Gaussian mixture model is used to model the data gradient distribution, and the optimal number of modes and the position of the modal mean are determined using the Bayesian information criterion. The modal mean is set in a symmetrical distribution during initialization to ensure that the disturbance noise meets the zero-mean unbiasedness requirement.

[0059] Specifically, based on the gradient distribution trained on the local model, each ship uses the Bayesian information criterion to model the local gradient distribution of each layer on the client side, automatically determining the current optimal number of modes N and its set of central means D. t,i,l ={C1,C2,...,C N This allows the noise to flexibly adapt to changes in the current training state. For the mean c of each modality... n Find the mean μ that has the smallest Euclidean distance from it. k It is then mapped to the mean set μ set in the initialization phase, and finally assigned the corresponding weight to the component.

[0060]

[0061] Here, τ controls the width of the Gaussian kernel, determining the sensitivity of the weights to distance changes. To avoid some modal weights being too low, we set a lower limit ω for the weights. min .

[0062]

[0063] We require all symmetric means to have the same weight to avoid noise introducing systematic bias in the optimization direction.

[0064]

[0065] To enhance the expression of modal differences, we use a nonlinear amplification factor λ to stretch and enhance the weights, thereby strengthening the expression of differences between modes.

[0066]

[0067] Normalization yields the modal weights. The client assigns sampling probabilities and variances for each mode based on these weights, so that the perturbation distribution can focus on the main mode region of the current gradient.

[0068]

[0069] Step 6: For the noise scale, the client calculates the mixed sensitivity value based on local and global sensitivity statistics. The client combines local and global sensitivity to calculate the mixed sensitivity value, and adjusts the noise scale accordingly. Local sensitivity reflects the data characteristics of a single ship, while global sensitivity reflects the overall characteristics across ships.

[0070] Specifically, considering local sensitivity In the early stages of model training, there may be significant fluctuations, leading to instability due to noise perturbations. To address this, we introduce a global sensitivity. To improve the stability of sensitivity estimation, specifically, after each round of communication, the client uploads its local sensitivity, and the server aggregates the global sensitivity and sends it back to the client.

[0071]

[0072] The client then merges global and local sensitivities to calculate the mixed sensitivities of each layer of the model. Achieve hierarchical adaptive adjustment of noise level.

[0073]

[0074] Step 7: Based on the hybrid sensitivity value, the noise variance of each mode is calculated using a dynamic variance adjustment strategy, and the noise intensity is gradually reduced during training iterations using an exponential decay strategy until the set minimum noise variance threshold is reached, thereby achieving dual adaptive noise injection of structure and scale.

[0075] The client's historical mixing sensitivity is used during dynamic adjustment of noise variance. For reference, dynamically adjust the noise variance of the current round.

[0076]

[0077] Where κ is the scaling factor, which controls the amplification effect of changes in sensitivity on variance.

[0078] Finally, an exponential decay strategy is used to reduce the noise variance. The value decreases smoothly with each training round until it reaches the set minimum value.

[0079]

[0080] This mechanism can provide strong privacy protection in the early stages of training while reducing noise in the later stages to accelerate model convergence.

[0081] Finally, each client will be based on normalized weights. For each Gaussian component in each layer The sampling probability and variance share are allocated as the basis for structural adjustment of noise disturbance.

[0082]

[0083] After completing the noise weight and variance allocation, the client assigns sampling probabilities and variance shares to each Gaussian component in each gradient layer based on normalized weights, thus completing the structural perturbation injection.

[0084] Step 8: To address the potential drift in gradient distribution caused by changes in the ship's operating environment during long-term training, this example employs an update mechanism to recalculate the number of modes and the mean set in each round. This strategy can improve system stability while maintaining adaptability, thereby achieving high-precision convergence of the model under privacy budget constraints and adapting to changes in the distribution of ship data at different times and in different environments.

[0085] Step 9: Anonymize the model gradients or parameters injected with noise to generate perturbed model update data, and upload it to the federated learning aggregation server through an encrypted communication channel.

[0086] Step 10: The federated learning aggregation server receives disturbance model update data from multiple smart ships, merges the update data using an aggregation algorithm to generate a global model, and distributes the updated global model to each ship for continued local training.

[0087] The intelligent ship data privacy protection method provided by this invention is applicable to multi-party collaborative modeling scenarios based on federated learning. This method can effectively prevent the leakage of sensitive data such as ship location, operating mode, and business information, while reducing the impact of noise injection on model accuracy and data availability. It has good adaptability and practical value in intelligent ship navigation safety, operation optimization, and privacy protection.

[0088] Furthermore, the effectiveness of the intelligent ship data privacy protection method based on dual adaptive differential privacy of structure and scale in this embodiment is illustrated through simulation experiments. In this case, the number of clients N=10, the number of global communications T=100, and the federated learning model is used to predict the accuracy of image classification on the MNIST dataset.

[0089] Please see Figure 3 , Figure 3 This is a bar chart of global model accuracy provided by an embodiment of the present invention. As can be seen from the figure, the model accuracy can reach a high level under different heterogeneity settings.

[0090] Example 2:

[0091] like Figure 4 As shown, the present invention also provides an intelligent ship data privacy protection system, the system comprising:

[0092] Data acquisition module 1 is used to acquire local data from the intelligent ship;

[0093] Sensitivity analysis module 2 is used to perform sensitivity analysis based on the characteristics of the multi-source data;

[0094] Noise structure modeling module 3 is used to model the data gradient distribution based on a multimodal Gaussian mixture model and determine the optimal modal structure;

[0095] Hybrid sensitivity calculation module 4 is used to calculate the hybrid sensitivity value by fusing local and global gradient statistical information;

[0096] Noise injection module 5 is used to dynamically adjust the noise variance based on the mixed sensitivity value and noise attenuation strategy and inject noise into the model gradient corresponding to the multi-source data;

[0097] Parameter upload module 6 is used to upload the perturbed model update parameters to the federated learning aggregation server;

[0098] Global model receiving module 7 is used to receive the updated global model and apply it to local training.

[0099] In a preferred embodiment of the present invention, the hybrid sensitivity calculation module calculates the hybrid sensitivity value according to the following formula:

[0100]

[0101] in, For local sensitivity, λ represents the global sensitivity, and λ is the fusion weighting factor.

[0102] Example 3:

[0103] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Figure 5 As shown, the electronic device includes: a processor 501, a memory 502, a communication interface 503, and a bus 504; wherein, the processor 501, the memory 502, and the communication interface 503 communicate with each other through the bus 504; the communication interface 503 is used for information transmission between the electronic device and the communication device of the terminal; the processor 501 is used to call program instructions in the memory 502 to execute the above-mentioned intelligent ship data privacy protection method.

[0104] Example 4: The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a method for protecting data privacy of intelligent ships.

[0106] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above embodiments. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention.

[0107] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not describe the various possible combinations separately.

[0108] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0109] Furthermore, various different implementations of the present invention can be combined arbitrarily, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed in the present invention.

Claims

1. A method for protecting the data privacy of intelligent ships, characterized in that, The method includes: Step S1: Construct a global model network structure in the central server of the ship system, and construct a local model network structure on the local end of each ship; wherein, the global model network structure and the local model network structure have the same structure. Step S2: Acquire local multi-source data of the intelligent ship; wherein, the data includes location data, navigation parameters, equipment operating status, cargo information and business data; Step S3: Initialize the noise model settings for each ship; Step S4: Based on the characteristics of the acquired local multi-source data, analyze the privacy sensitivity of each data dimension, obtain the sensitivity value of each dimension, and store it in the local sensitivity analysis cache. Step S5: A multimodal Gaussian mixture model is used to model the data gradient distribution, and the optimal number of modes and the position of the modal mean are determined using the Bayesian information criterion. The modal mean is set according to a symmetrical distribution during initialization to ensure that the disturbance noise meets the zero-mean unbiasedness requirement. Step S6: Based on local and global sensitivity statistics, the client calculates the mixed sensitivity value. The client combines the local sensitivity and global sensitivity to calculate the mixed sensitivity value, and adjusts the noise scale accordingly. Step S7: Based on the hybrid sensitivity value, the noise variance of each mode is calculated using a dynamic variance adjustment strategy, and the noise intensity is gradually reduced during training iterations using an exponential decay strategy until the set minimum noise variance threshold is reached, thereby achieving dual adaptive noise injection of structure and scale.

2. The intelligent ship data privacy protection method according to claim 1, characterized in that, After step S7, the method further includes: Step S8: In order to cope with the possibility that changes in the ship's operating environment may cause gradient distribution to drift during long-term training, an update mechanism is adopted to recalculate the number of modes and the mean set in each round of calculation to ensure that the noise model always adapts to the current training data. Step S9: Anonymize the model gradients or parameters injected with noise to generate perturbed model update data, and upload it to the federated learning aggregation server through an encrypted communication channel. In step S10, the federated learning aggregation server receives disturbance model update data from multiple smart ships, merges the update data using an aggregation algorithm to generate a global model, and distributes the updated global model to each ship for continued local training.

3. The intelligent ship data privacy protection method according to claim 1, characterized in that, In step S3, the modality weights and variances are set symmetrically during initialization and dynamically adjusted during training.

4. The intelligent ship data privacy protection method according to claim 1, characterized in that, In step S4, the client calculates the mixing sensitivity value according to the following formula: in, For local sensitivity, λ represents the global sensitivity, and λ is the fusion weighting factor.

5. The intelligent ship data privacy protection method according to claim 1, characterized in that, In step S7, the variance adjustment coefficient is calculated based on the change in the current mixing sensitivity and the mixing sensitivity of the previous round, and the noise variance is reduced round by round in combination with the exponential decay strategy until the set minimum value is reached. Where κ is the scaling factor, which controls the amplification effect of changes in sensitivity on variance.

6. A smart ship data privacy protection system according to any one of claims 1-5, characterized in that, The system includes: The data acquisition module is used to acquire local data from the intelligent ship. Sensitivity analysis module, used to perform sensitivity analysis based on the characteristics of the multi-source data; The noise structure modeling module is used to model the gradient distribution of data based on a multimodal Gaussian mixture model and determine the optimal modal structure. The hybrid sensitivity calculation module is used to calculate the hybrid sensitivity value by fusing local and global gradient statistics. The noise injection module is used to dynamically adjust the noise variance based on the mixed sensitivity value and the noise attenuation strategy, and inject the noise into the model gradient corresponding to the multi-source data. The parameter upload module is used to upload the perturbed model update parameters to the federated learning aggregation server; The global model receiving module is used to receive the updated global model and apply it to local training.

7. The intelligent ship data privacy protection system according to claim 6, characterized in that, The hybrid sensitivity calculation module calculates the hybrid sensitivity value according to the following formula: in, For local sensitivity, λ represents the global sensitivity, and λ is the fusion weighting factor.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the intelligent ship data privacy protection method as described in any one of claims 1-5.

9. A computer-readable storage medium storing a computer program thereon, characterized in that, When the program is executed by the processor, it implements the intelligent ship data privacy protection method according to any one of claims 1-5.

Images