Underwater edge privacy-preserving task offloading method based on differential federated learning

Abstract

The present invention discloses an underwater edge privacy protection task offloading method based on differential federated learning. First, sensor nodes are deployed in the target waters and a network architecture is constructed. The state space, action space and reward function are defined respectively. The sensor nodes perform local model training after sensing the environmental data. Subsequently, each sensor node uploads the local model parameters to the sea surface buoy for model aggregation, and introduces differential perturbations during the uploading process to protect privacy. After the model aggregation is completed, the sea surface buoy sends the global model to each sensor node, and the node generates a task offloading strategy based on it. Finally, the sensor node encrypts the task data through homomorphic encryption, and securely offloads the encrypted computing task to the corresponding sea surface buoy or low-orbit satellite for execution. This method effectively reduces privacy leakage during underwater task offloading and improves the security of data transmission in the underwater network.

Description

Technical Field

[0001] The present invention belongs to the field of edge computing technology, and specifically relates to an underwater edge privacy protection task offloading method based on differential federated learning. Background Art

[0002] In underwater applications such as marine environmental monitoring, resource exploration, and military reconnaissance, sensor nodes are often deployed underwater, generating a continuous stream of computing tasks that require timely and efficient processing. Traditional methods rely on underwater nodes to process data themselves, but these nodes have limited computing power and power, making it difficult to efficiently complete complex tasks. Furthermore, the unique underwater communication environment makes data transmission susceptible to eavesdropping and tampering, posing a serious risk of privacy breaches.

[0003] In recent years, edge computing-based task offloading technology has been widely used in underwater environments. By offloading computing tasks to surface buoys or low-orbit satellites, it can significantly reduce the energy consumption of underwater nodes and improve task processing efficiency. For example, Nanjing University, in its patent application document "Underwater Computing Task Offloading Method Based on Improved Deep Deterministic Policy Gradient" (filing date: December 4, 2024, application number: 202411768075.4, application publication number: CN 119255300 A, the content of which is still citationable), discloses a method for underwater computing task offloading based on improved deep deterministic policy gradient. This underwater computing task offloading method has the following shortcomings: during the underwater task offloading process, privacy protection issues during task data transmission are ignored, making sensitive information easily stolen or leaked, and security difficult to ensure.

[0004] Some research proposes using federated learning to protect data privacy. For example, Guangzhou University, in its patent application titled "A Secure Communication Method for Industrial IoT Edge Computing Based on Federated Learning" (filing date: May 13, 2025, application number: 202510607873.7, publication number: CN 120151374 A, which is still citationable), discloses a method for offloading tasks to the edge of the Industrial IoT based on federated learning. This task offloading method has the following shortcomings: The parameter exchange during the federated learning process still poses the risk of privacy leakage, lacks additional protection mechanisms such as differential privacy, and has poor privacy protection. Furthermore, the transmission of model parameters lacks effective security protection mechanisms, making it vulnerable to external malicious attacks and data tampering, making it unsuitable for underwater environments.

[0005] In summary, the existing technology has problems such as high risk of privacy leakage and insufficient security during edge task offloading in underwater environments. At the same time, it does not fully consider the privacy protection and security protection measures during the transmission of model parameters, which affects the overall performance and data security of the system. Summary of the Invention

[0006] Purpose of the invention: The purpose of the present invention is to overcome the shortcomings of the existing technology and propose an underwater edge privacy protection task offloading method based on differential federated learning to solve the privacy leakage and data security problems existing in the existing underwater edge computing task offloading process. Specifically, the present invention effectively prevents the leakage of sensitive information by introducing a differential privacy perturbation mechanism in the process of uploading model parameters after local model training of the sensor node; in the task data offloading link, the task data is encrypted by homomorphic encryption technology to ensure the security of the data transmission process. This method aims to significantly reduce the risk of privacy leakage in the task offloading process, improve the overall security and system performance of data transmission, and meet the needs of safe and efficient offloading of edge computing tasks in underwater environments.

[0007] This paper proposes an underwater edge privacy protection task offloading method based on differential federated learning to make up for the shortcomings of the above technologies.

[0008] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions:

[0009] An underwater edge privacy-preserving task offloading method based on differential federated learning includes the following steps:

[0010] S1: Deploy sensor nodes in the target waters and build a network architecture;

[0011] S2: Define the state space, action space and reward function respectively. The sensor node performs local model training after perceiving the environmental data.

[0012] S3: Each sensor node uploads its local model parameters to the surface buoy for model aggregation, and introduces differential perturbations during the upload process to protect privacy. After model aggregation is complete, the surface buoy sends the global model to each sensor node, which then generates a task offloading strategy based on it.

[0013] S4: The sensor node encrypts the task data through homomorphic encryption and securely offloads the encrypted computing task to the corresponding sea surface buoy or low-orbit satellite for execution.

[0014] Furthermore, the step S1 is specifically as follows:

[0015] Deploy sensor nodes in the target waters and form a network with sea surface buoys and low-orbit satellites. Define the low-orbit satellite set as , the sea surface buoy set is , the sensor node set is Each surface buoy is equipped with an edge server to provide computing services. The buoy has built-in acoustic and electromagnetic wave communication modules that can communicate with sensor nodes and low-orbit satellites simultaneously. The low-orbit satellites carry cloud servers that serve as the destination for offloading computing tasks when the surface buoy's computing power is insufficient.

[0016] sensor nodes In each discrete time slot Generate computing tasks , the sensor nodes In the time slot Generate computing tasks The set of , underwater existence An eavesdropping node attempts to monitor or tamper with the computing tasks transmitted by the sensor node. The onshore key management center is responsible for regularly generating and distributing keys; defining the offloading decision variable and ,when =1, indicating a computing task In the time slot Uninstalled to A surface buoy; otherwise =0; when =1, indicating a computing task In the time slot Uninstalled to low-orbit satellites; otherwise =0.

[0017] Furthermore, the step S2 is specifically as follows:

[0018] S2-1: Define the state space, action space and reward function respectively:

[0019] (1) State space: In each time slot , the state space Defined as:

[0020] (1)

[0021] in, is the task data volume; and are the distances from the sensor node to the sea surface buoy and from the sea surface buoy to the low-orbit satellite, respectively; and Sea surface buoys and low-orbit satellites The remaining computing resources; Surface buoy of remaining storage space; is the differential privacy budget, is the key length for homomorphic encryption;

[0022] (2) Action space: Action space is the combination of all values ​​of the uninstall decision variable, defined as: (2)

[0023] (3) Reward function: The goal is to minimize the total processing delay and total energy consumption of the task, while considering the differential privacy budget , the reward function Defined as:

[0024] (3)

[0025] in, is the weight factor, For the task The total processing delay, For the task The total energy consumption, is the privacy penalty coefficient;

[0026] S2-2: The sensor node performs local model training after sensing environmental data:

[0027] In the time slot , the sensor node uses local data to perform local model training, according to the current state and select the current action , get instant rewards and enter a new state , using the experience replay pool for sample storage, the main Q network is in the current state Take the current action The Q value after Defined as:

[0028] (4)

[0029] in, The main Q network is in the previous state Next take the previous action The Q value after For the target Q, the network is in the next state Take the best action Q value, is the learning rate, is the discount factor.

[0030] Furthermore, the step S3 is specifically as follows:

[0031] S3-1: Local model parameters are added with Laplace distributed noise in differential perturbation:

[0032] Before the sensor node uploads the model parameters to the edge server of the sea buoy, all local model parameters are added with Laplace distribution noise in the differential perturbation. The local main Q network model parameters after adding noise are and local target Q network model parameters Defined as:

[0033] (5)

[0034] (6)

[0035] in, are the local main Q network model parameters before noise addition; is the local target Q network model parameter before noise addition, is the Laplace distribution, is the global sensitivity, Budget for differential privacy;

[0036] S3-2: Upload the noisy local model parameters to the sea surface buoy for model aggregation:

[0037] The noisy local model parameters are uploaded to the edge server connected to the sea surface buoy through a secure channel. After each sea surface buoy receives the noisy model parameters from the sensor nodes within its service range, it uses the federated averaging algorithm to aggregate the model, which is defined as:

[0038] (7)

[0039] (8)

[0040] in, is the global main Q network model parameter, is the global target Q network model parameter, Surface buoy The set of sensor nodes under jurisdiction, Surface buoy The number of sensor nodes under jurisdiction;

[0041] S3-3: Global model distribution:

[0042] After the model aggregation is completed, the sea surface buoy sends the global model to each sensor node, and the sensor node generates a task offloading strategy based on it.

[0043] Furthermore, the step S4 is specifically as follows:

[0044] S4-1: The key management center generates and distributes key pairs:

[0045] The onshore key management center generates and distributes key pairs, which are divided into public keys and private keys. The key management center distributes the public key to the sensor nodes and the private key to the sea surface buoys and low-orbit satellites.

[0046] S4-2: Mission data encryption and decryption:

[0047] The sensor node uses the public key to encrypt the task data to be offloaded, and offloads the encrypted data transmission to the corresponding sea surface buoy or low-orbit satellite. After the sea surface buoy or low-orbit satellite receives the encrypted data, it uses the corresponding private key to decrypt it, obtain the original data and complete the task processing.

[0048] The advantages and technical effects of the present invention are as follows:

[0049] The present invention first introduces differential privacy perturbations during the process of uploading local model parameters at sensor nodes, effectively protecting sensitive information in the model parameters and significantly reducing the risk of privacy leakage. Secondly, it uses the federated averaging algorithm for model aggregation to ensure efficient collaborative training and stable convergence of the global model, thereby improving the quality of the task offloading strategy. Finally, by using homomorphic encryption technology for task data, it ensures that task data is always encrypted during transmission, effectively preventing data from being eavesdropped or tampered with, and significantly enhancing the security of the underwater edge computing network. After simulation verification, when the number of underwater eavesdropping node attacks reaches 30 times / minute, the method provided by the present invention reduces the number of privacy leaks by approximately 83.3% compared to the underwater task offloading method based on ordinary federated learning.

[0050] The present invention effectively solves the privacy leakage problem in the process of underwater task offloading and improves the data security of the underwater edge computing system. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is an overall flow chart of an embodiment of the present invention;

[0052] Figure 2 is a network architecture diagram of an embodiment of the present invention;

[0053] Figure 3 Schematic diagram of a key management center distributing public keys and private keys according to an embodiment of the present invention;

[0054] Figure 4 This is a comparison chart of simulation results of the number of privacy leaks using the method provided by the present invention and the underwater task offloading method based on ordinary federated learning in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0056] This embodiment proposes an underwater edge privacy protection task offloading method based on differential federated learning. The overall flow chart is as follows: Figure 1 As shown, the following steps are included:

[0057] S1: Deploy sensor nodes in the target waters and build a network architecture. The specific steps are as follows:

[0058] like Figure 2 As shown in Figure 2, 15 sensor nodes are deployed in the target waters and form a network with 3 sea surface buoys and 2 low-orbit satellites. The low-orbit satellite set is defined as , the sea surface buoy set is , the sensor node set is Each surface buoy is equipped with an edge server to provide computing services. The buoy has built-in acoustic and electromagnetic wave communication modules that can communicate with sensor nodes and low-orbit satellites simultaneously. The low-orbit satellites carry cloud servers that serve as the destination for offloading computing tasks when the surface buoy's computing power is insufficient.

[0059] sensor nodes In each discrete time slot Generate computing tasks , the sensor nodes In the time slot Generate computing tasks The set of , underwater existence An eavesdropping node attempts to monitor or tamper with the computing tasks transmitted by the sensor node. The onshore key management center is responsible for regularly generating and distributing keys; defining the offloading decision variable and ,when =1, indicating a computing task In the time slot Uninstalled to A surface buoy; otherwise =0; when =1, indicating a computing task In the time slot Uninstalled to low-orbit satellites; otherwise =0.

[0060] S2: Define the state space, action space, and reward function respectively. The sensor node performs local model training after sensing the environmental data. The specific steps are as follows:

[0061] S2-1: Define the state space, action space and reward function respectively:

[0062] (1) State space: In each time slot , the state space Defined as:

[0063] (1)

[0064] in, is the task data volume; and are the distances from the sensor node to the sea surface buoy and from the sea surface buoy to the low-orbit satellite, respectively; and Sea surface buoys and low-orbit satellites The remaining computing resources; Surface buoy of remaining storage space; is the differential privacy budget, is the key length for homomorphic encryption;

[0065] (2) Action space: Action space is the combination of all values ​​of the uninstall decision variable, defined as:

[0066] (2)

[0067] (3) Reward function: The goal is to minimize the total processing delay and total energy consumption of the task, while considering the differential privacy budget , the reward function Defined as:

[0068] (3)

[0069] in, is the weight factor, in this embodiment, =0.5, For the task The total processing delay, For the task The total energy consumption, is the privacy penalty coefficient, in this embodiment, =0.4;

[0070] S2-2: The sensor node performs local model training after sensing environmental data:

[0071] In the time slot , the sensor node uses local data to perform local model training, according to the current state and select the current action , get instant rewards and enter a new state , using the experience replay pool for sample storage, the main Q network is in the current state Take the current action The Q value after Defined as:

[0072] (4)

[0073] in, The main Q network is in the previous state Next take the previous action The Q value after For the target Q, the network is in the next state Take the best action Q value, is the learning rate, is the discount factor, in this embodiment, =0.01, =0.5.

[0074] S3: Each sensor node uploads its local model parameters to the sea surface buoy for model aggregation. Differential perturbations are introduced during the upload process to protect privacy. After model aggregation is complete, the sea surface buoy sends the global model to each sensor node, which then generates a task offloading strategy based on the model. The specific steps are as follows:

[0075] S3-1: Local model parameters are added with Laplace distributed noise in differential perturbation:

[0076] Before the sensor node uploads the model parameters to the edge server of the sea buoy, all local model parameters are added with Laplace distribution noise in the differential perturbation. The local main Q network model parameters after adding noise are and local target Q network model parameters Defined as:

[0077] (5)

[0078] (6)

[0079] in, are the local main Q network model parameters before noise addition; is the local target Q network model parameter before noise addition, is the Laplace distribution, is the global sensitivity, Budget for differential privacy;

[0080] S3-2: Upload the noisy local model parameters to the sea surface buoy for model aggregation:

[0081] The noisy local model parameters are uploaded to the edge server connected to the sea surface buoy through a secure channel. After each sea surface buoy receives the noisy model parameters from the sensor nodes within its service range, it uses the federated averaging algorithm to aggregate the model, which is defined as:

[0082] (7)

[0083] (8)

[0084] in, is the global main Q network model parameter, is the global target Q network model parameter, Surface buoy The set of sensor nodes under jurisdiction, Surface buoy The number of sensor nodes under jurisdiction;

[0085] S3-3: Global model distribution:

[0086] After the model aggregation is completed, the sea surface buoy sends the global model to each sensor node, and the sensor node generates a task offloading strategy based on it.

[0087] S4: The sensor node encrypts the task data using homomorphic encryption and securely offloads the encrypted computing task to the corresponding sea surface buoy or low-orbit satellite for execution. The specific steps are as follows:

[0088] S4-1: The key management center generates and distributes key pairs:

[0089] like Figure 3 As shown, the key management center on shore generates and distributes key pairs, which are divided into public keys and private keys. The key management center distributes the public keys to sensor nodes and the private keys to sea surface buoys and low-orbit satellites.

[0090] S4-2: Mission data encryption and decryption:

[0091] The sensor node uses the public key to encrypt the task data to be offloaded, and offloads the encrypted data transmission to the corresponding sea surface buoy or low-orbit satellite. After the sea surface buoy or low-orbit satellite receives the encrypted data, it uses the corresponding private key to decrypt it, obtain the original data and complete the task processing.

[0092] The simulation comparison results of the privacy leakage times using the method provided by the present invention and the underwater task offloading method based on ordinary federated learning are as follows: Figure 4The conventional federated learning-based underwater task offloading method does not utilize differential privacy and homomorphic encryption as used in this example. This example was simulated in Matlab R2023a, using 32GB of computer memory and an x64-based Intel Core i5-12400F CPU. The specific parameters for all simulations in this example are listed in Table 1. All simulations were performed using the same parameter settings to ensure fair comparison.

[0093] Table 1 Simulation parameters

[0094] Simulation parameters Numerical Number of sensor nodes#timg# 15 Number of buoys on the sea surface#timg# 3 Number of low-orbit satellites#timg# 2 Differential Privacy Budget#timg# 0.8 Homomorphic encryption key length#timg# 2048 bits Task data volume#timg# 30 MB Privacy penalty coefficient#timg# 0.4 Learning rate 0.01 Discount factor#timg# 0.5 The number of sensor nodes managed by the sea surface buoy#timg# 5 Number of eavesdropping nodes#timg# 3

[0095] from Figure 4 The simulation results show that as the number of attacks by underwater eavesdropping nodes increases, the number of privacy leaks of the method provided by the present invention remains at a low level. This is because the method provided by the present invention introduces differential privacy perturbations in the process of uploading local model parameters, protecting sensitive information in the model parameters, and using homomorphic encryption technology for task data to ensure that task data is always encrypted during transmission, significantly reducing the risk of privacy leaks. The underwater task offloading method based on ordinary federated learning performs relatively poorly. When the number of attacks by underwater eavesdropping nodes reaches 30 times / minute, the method provided by the present invention reduces the number of privacy leaks by about 83.3% compared to the underwater task offloading method based on ordinary federated learning.

[0096] In summary, the present invention effectively solves the privacy leakage problem in the process of underwater task offloading and improves the data security of the underwater edge computing system.

[0097] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for a person skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to replace some of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions claimed to be protected by the present invention.

Claims

1. A method for underwater edge privacy protection task offloading based on differential federated learning, characterized by: The following steps are involved: S1: Deploy sensor nodes in the target waters and build a network architecture; Deploy sensor nodes in the target waters and form a network with sea surface buoys and low-orbit satellites. Define the low-orbit satellite set as , the sea surface buoy set is , the sensor node set is Each surface buoy is equipped with an edge server to provide computing services. The surface buoy has built-in acoustic and electromagnetic wave communication modules, which can communicate with sensor nodes and low-orbit satellites at the same time. The low-orbit satellite carries a cloud server, which is used as a destination for offloading computing tasks when the computing power of the sea buoy is insufficient; sensor nodes In each discrete time slot Generate computing tasks , the sensor nodes In the time slot Generate computing tasks The set of , underwater existence An eavesdropping node attempts to monitor or tamper with the computing tasks transmitted by the sensor node. The onshore key management center is responsible for regularly generating and distributing keys; defining the offloading decision variable and ,when =1, indicating a computing task In the time slot Uninstalled to A surface buoy; otherwise =0; when =1, indicating a computing task In the time slot Uninstalled to low-orbit satellites; otherwise =0; S2: Define the state space, action space and reward function respectively. The sensor node performs local model training after perceiving the environmental data. S3: Each sensor node uploads its local model parameters to the surface buoy for model aggregation, and introduces differential perturbations during the upload process to protect privacy. After model aggregation is complete, the surface buoy sends the global model to each sensor node, which then generates a task offloading strategy based on it. S3-1: Local model parameters are added with Laplace distributed noise in differential perturbation: Before the sensor node uploads the model parameters to the edge server of the sea buoy, all local model parameters are added with Laplace distribution noise in the differential perturbation. The local main Q network model parameters after adding noise are and local target Q network model parameters Defined as: (1) (2) in, are the local main Q network model parameters before noise addition; is the local target Q network model parameter before noise addition, is the Laplace distribution, is the global sensitivity, Budget for differential privacy; S3-2: Upload the noisy local model parameters to the sea surface buoy for model aggregation: The noisy local model parameters are uploaded to the edge server connected to the sea surface buoy through a secure channel. After each sea surface buoy receives the noisy model parameters from the sensor nodes within its service range, it uses the federated averaging algorithm to aggregate the model, which is defined as: (3) (4) in, is the global main Q network model parameter, is the global target Q network model parameter, Surface buoy The set of sensor nodes under jurisdiction, Surface buoy The number of sensor nodes under jurisdiction; S3-3: Global model distribution: After the model aggregation is completed, the sea surface buoy sends the global model to each sensor node, and the sensor node generates a task offloading strategy based on it; S4: The sensor node encrypts the task data through homomorphic encryption and securely offloads the encrypted computing task to the corresponding sea surface buoy or low-orbit satellite for execution.

2. The underwater edge privacy protection task offloading method based on differential federated learning according to claim 1 is characterized in that: The step S2 is specifically as follows: S2-1: Define the state space, action space and reward function respectively: (1) State space: In each time slot , the state space Defined as: (5) in, is the task data volume; and are the distances from the sensor node to the sea surface buoy and from the sea surface buoy to the low-orbit satellite, respectively; and Sea surface buoys and low-orbit satellites The remaining computing resources; Surface buoy of remaining storage space; is the differential privacy budget, is the key length for homomorphic encryption; (2) Action space: Action space is the combination of all values ​​of the uninstall decision variable, defined as: (6) (3) Reward function: The goal is to minimize the total processing delay and total energy consumption of the task, while considering the differential privacy budget , the reward function Defined as: (7) in, is the weight factor, For the task The total processing delay of For the task The total energy consumption, is the privacy penalty coefficient; S2-2: The sensor node performs local model training after sensing environmental data: In the time slot , the sensor node uses local data to perform local model training, according to the current state and select the current action , get instant rewards and enter a new state , using the experience replay pool for sample storage, the main Q network is in the current state Take the current action The Q value after Defined as: (8) in, The main Q network is in the previous state Next take the previous action The Q value after For the target Q, the network is in the next state Take the best action Q value, is the learning rate, is the discount factor.

3. The underwater edge privacy protection task offloading method based on differential federated learning according to claim 1 is characterized in that: The step S4 is specifically as follows: S4-1: The key management center generates and distributes key pairs: The onshore key management center generates and distributes key pairs, which are divided into public keys and private keys. The key management center distributes the public key to the sensor nodes and the private key to the sea surface buoys and low-orbit satellites. S4-2: Mission data encryption and decryption: The sensor node uses the public key to encrypt the task data to be offloaded, and offloads the encrypted data transmission to the corresponding sea surface buoy or low-orbit satellite. After the sea surface buoy or low-orbit satellite receives the encrypted data, it uses the corresponding private key to decrypt it, obtain the original data and complete the task processing.

Images