Unmanned aerial vehicle federated learning selective homomorphic encryption method based on resource constraint
By introducing a globally mask-optimized selective homomorphic encryption method into UAV federated learning, the problems of high cost of full encryption and unbalanced privacy protection are solved, achieving a balance between resource conservation and global privacy protection, and adapting to low-altitude economic applications where UAV computing and communication are limited.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-14
- Publication Date
- 2026-04-07
AI Technical Summary
In the context of drone federated learning, full homomorphic encryption is too costly and existing solutions cannot take into account global privacy protection, resulting in an imbalance in privacy protection capabilities.
Under the constraints of energy consumption and bandwidth, a global mask optimization selective homomorphic encryption method is adopted, a greedy selection strategy is used to optimize parameter encryption, and a global unified mask is generated by combining privacy gain density to ensure a balance between global privacy protection and resource consumption.
It effectively reduces communication and energy consumption costs, improves the overall level of privacy protection, ensures system security and fairness, and adapts to resource-constrained environments for drones.
Smart Images
Figure CN121814296A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence and privacy protection, specifically relating to a resource-constrained drone federated learning selective homomorphic encryption method. Background Technology
[0002] With the widespread application of drones in low-altitude economic scenarios such as environmental monitoring, disaster relief, logistics delivery, and urban governance, a large amount of distributed data is being continuously collected and processed. To achieve collaborative modeling without centralized transmission of raw data, federated learning (FL) has become an important technical approach. However, while traditional federated learning avoids directly uploading raw data, it still carries the risk of privacy breaches. Attackers could indirectly recover the features of the original data by performing gradient inversion, membership inference, and other attacks through uploaded parameter updates.
[0003] To enhance security, homomorphic encryption has become a mainstream approach. However, full encryption can lead to enormous computational costs in scenarios like drones where computing and communication resources are limited. Existing improved homomorphic encryption schemes for federated learning, such as MaskCrypt, attempt to rank the importance of all user training parameters and encrypt parameters in a fixed proportion to reduce computational costs. However, this approach lacks protection for global privacy, focusing instead on protecting the highly important parameters of a few users, resulting in suboptimal privacy protection. Therefore, it is necessary to introduce a more effective global masking optimization algorithm based on the MaskCrypt scheme, and to incorporate energy consumption and bandwidth constraints to improve the balance between privacy and performance for drones in homomorphic encryption-based federated learning scenarios. Summary of the Invention
[0004] This invention aims to address the problem of excessively high cost of full homomorphic encryption and the inability of mask selection schemes to ensure global privacy in the context of drone federated learning (FL) scenarios. To adapt to the drone environment, a homomorphic encryption method with global mask optimization under energy consumption and bandwidth constraints is proposed: while ensuring aggregation correctness and global privacy protection capabilities, communication and energy consumption overhead is reduced, and the privacy imbalance caused by protecting only the highly important parameters of a few users is avoided. This achieves an overall approximate minimization of global model privacy risks and controlled resource consumption.
[0005] The present invention achieves the above-mentioned technical objectives through the following technical means.
[0006] A resource-constrained UAV federated learning selective homomorphic encryption method includes the following process:
[0007] Step 1: System Initialization and Parameter Sensitivity Calculation
[0008] Initialize the local model parameters, remaining energy, and communication bandwidth of the UAV, and pre-deploy homomorphic encryption public key pairs. Based on this initialization, each UAV performs local model training, obtaining the absolute value of the difference between local parameter updates and gradient information. Calculate the sensitivity index (the product of the gradient and the absolute value of the difference between parameter updates) for each parameter based on the absolute value of the difference between local parameter updates and gradient information, and perform normalization processing to obtain the sensitivity set for each parameter of the UAV. Each UAV uploads the sensitivity set, remaining energy, and communication bandwidth to the server.
[0009] Step 2: Global Mask Consensus Generation
[0010] The server collects the sensitivity set, remaining energy, and communication bandwidth set of each drone terminal, introduces communication bandwidth and energy consumption constraints, and models the problem as a greedy choice problem of "privacy benefit density (parameter sensitivity / resource cost)", obtains a globally unified mask, and distributes it to the drone terminals.
[0011] Step 3: Parameter Encryption and Upload
[0012] The drone receives a globally unified mask and divides its local parameters into an "encrypted transmission part" and a "plaintext transmission part" based on the mask. The "encrypted transmission part" is encrypted using a homomorphic encryption algorithm to obtain the "encrypted transmission part." Each drone then uploads both the "encrypted transmission part" and the "plaintext transmission part" to the server.
[0013] Step 4: Server aggregation and client decryption update
[0014] The server performs direct aggregation on the "plaintext transmission portion" and homomorphic addition aggregation on the "encrypted transmission portion," obtaining aggregated "plaintext transmission portions" and "encrypted transmission portions," which are then sent to each drone. The drone receives the aggregated "plaintext transmission portion" and "encrypted transmission portion," decrypts the aggregated "encrypted transmission portion" using a pre-deployed homomorphic encryption private key, and combines it with the aggregated "plaintext transmission portion" to obtain the complete parameter update content. Each drone uses this complete parameter update content to complete its local model synchronization for this round.
[0015] Furthermore, the specific process of step 1 is as follows:
[0016] Step 1.1: System Initialization
[0017] The system includes a collection of drone clients. In addition, a centralized server is required. If this is the initial round, the specific parameter settings are as follows:
[0018] 1) Initialize model parameters: Pre-deploy the global model parameter vector w for all drones. 0 .
[0019] 2) Resource status settings: Initialize the remaining energy of each drone k. and available bandwidth
[0020] 3) Key Deployment: The system pre-generates a homomorphic encryption public key pk and a private key sk. The server and the drone share the pk, while the drone holds the sk required for decryption. Throughout the federated learning process, the server only performs homomorphic aggregation operations based on the public key and does not have decryption capabilities.
[0021] This step obtains the initial global model parameters, the drone's remaining energy / bandwidth status, and the public / private key pair, preparing for subsequent operations.
[0022] Step 1.2: Local Model Training and Parameter Update Calculation for UAV
[0023] In each iteration t, drone k is based on its local dataset. The global model w obtained by the drone in the previous round t Perform local training. Let the loss function be... The gradient is
[0024]
[0025] in, It is the local average gradient vector of drone k in the t-th round of training; This represents the size of the local dataset for drone k, i.e., the number of samples; the summation symbol... Represents the local dataset for drone k. The summation of all samples in the dataset, where (x, y) represents the dataset. One piece of data in the table. l(w t ;x,y) represents the loss value calculated for the data point (x,y), while Then it means l(w) t (x, y) represents the global model w obtained by the UAV in the previous round. t The gradient of a vector.
[0026] The updated local parameter vector after completing t rounds of local model training is denoted as: The absolute value of the parameter difference is
[0027]
[0028] The output of this step is the gradient. Difference between parameter update
[0029] Step 1.3: Parameter sensitivity calculation and uploading
[0030] Based on the sensitivity definition in MaskCrypt, the sensitivity index for each parameter dimension i is... The calculation method is as follows
[0031]
[0032] in, Let be the absolute value of the update difference for the i-th parameter; This is the gradient corresponding to this parameter.
[0033] Then, the sensitivity vector Normalization is performed, and the calculation method is as follows:
[0034]
[0035] Where d represents the number of parameters on the drone side.
[0036] Obtain the sensitivity set of UAV end parameters Finally, drone k will collect the sensitivity data. Remaining energy Available bandwidth Uploaded to the server for subsequent global mask consensus.
[0037] Furthermore, the specific process of step 2 is as follows:
[0038] Step 2.1: Introduce communication bandwidth and energy consumption constraints
[0039] In step 1, each drone client Its parameter sensitivity set has been uploaded. Remaining energy Available bandwidth The server integrates it as follows:
[0040]
[0041] This information set serves as input for optimization modeling, used to generate a globally uniform mask x. t Define a globally uniform mask x t ∈{0,1} d ,in This indicates that the i-th parameter is selected as the encrypted parameter in round t; otherwise, it is transmitted in plaintext. Next, we introduce communication bandwidth and energy consumption constraints:
[0042] 1) Communication bandwidth constraints
[0043] The amount of communication between the drone and the server in each round is limited by the available bandwidth. Since the size of the ciphertext parameter is much larger than that of the plaintext parameter, they need to be calculated separately, and the total amount of uploaded data must not exceed the bandwidth limit. The plaintext size of the i-th parameter is denoted as... The size after encryption is denoted as The upload data volume constraint is:
[0044]
[0045] Where, x t，i Represents the globally uniform mask x t The i-th value; N is the total number of parameters in the local model on the UAV; D(x t ) represents the global uniform mask x t The total communication bandwidth required to upload and download parameters.
[0046] 2) Energy consumption constraints
[0047] Remaining energy of the drone Energy consumption consists of two parts: computational energy consumption (energy consumed during local training, parameter encryption, and decryption of the ciphertext portion of the global model); and communication energy consumption (energy consumed during wireless transmission of plaintext and ciphertext parameters). Let the energy consumed by transmitting the i-th parameter in plaintext be denoted as . Its encryption, decryption, and ciphertext transmission consume energy of The energy consumption constraint is then...
[0048]
[0049] Where, x t，i Represents the globally uniform mask x t The i-th value; N is the total number of parameters in the local model on the drone; E(x) t ) represents the global uniform mask x t The total energy consumption required for this round.
[0050] Step 2.2: Greedy Selection Algorithm Steps
[0051] The algorithm needs to balance the importance of parameters and the cost of resources consumed. Therefore, this algorithm introduces privacy benefit density to prioritize the encryption of parameters with high "cost-effectiveness", which means setting the corresponding mask to 1.
[0052] Privacy benefit density is defined as
[0053]
[0054] in, This represents the sensitivity value of the i-th parameter in the t-th round; This represents the resource cost required to encrypt the i-th parameter in the t-th round. The resource cost... Energy consumption cost and communication cost It is composed of elements that can be pre-normalized or weighted according to system configuration to ensure that different resource dimensions are compared under the same scale. For example, in one implementation, it can be defined as: Where α∈[0,1] is the preset weight parameter.
[0055] In UAV federated learning scenarios, homomorphic encryption operations introduce both computational energy consumption and communication overhead, and the global model parameters have a high dimensionality. Using a globally optimal search method would lead to excessively high server-side computational complexity, which is detrimental to system real-time performance. Therefore, this invention addresses the specific problem of "parameter-level homomorphic encryption selection" in UAV federated learning. Under the premise of satisfying energy consumption and bandwidth constraints, it adopts a greedy selection strategy based on privacy gain density to reduce the overall computational complexity of the system while ensuring global privacy protection.
[0056] Next, we begin the greedy selection algorithm steps, based on the data received from the server. Calculate the privacy benefit density of all parameters Sort the results in descending order to obtain a priority list.
[0057]
[0058] in, This represents the j-th largest privacy benefit density value after sorting.
[0059] Starting from the beginning position (1) of the sorted sequence, first set the global mask x t Set all parameters to 0, and sequentially determine whether to perform homomorphic encryption transmission on the parameter, i.e., whether to set the mask position corresponding to the parameter to the encrypted state. Assume we are currently determining whether the (j)th parameter should be encrypted.
[0060]
[0061] Continue selecting sequentially until the entire sequence is traversed, ultimately obtaining a globally uniform mask x. t , where x t，i =1 indicates that the i-th parameter is selected for encryption, x t，i =0 indicates that the i-th parameter is transmitted in plaintext.
[0062] The globally unified mask generated through the above greedy selection process completes the decision on the server side and remains consistent across all drones participating in this round of federated learning. This ensures that all drones reach a unified consensus on the parameter-level homomorphic encryption strategy, guaranteeing that the mask has homomorphic encryption executability under the resource constraints of each drone.
[0063] Furthermore, the specific process of step 3 is as follows:
[0064] Step 3.1: Divide, encrypt, and upload parameters.
[0065] The drone receives the globally unified mask x sent by the server. t The local parameter vector is determined based on this mask. It is divided into two parts, namely the part that needs to be encrypted for transmission (mask value x). t，i =1) and plaintext transmission part (mask value x) t，i =0), that is
[0066]
[0067] in, For encrypted transmission, This refers to the part transmitted in plaintext.
[0068] For the parts that require encrypted transmission The drone uses the homomorphic encryption public key pk distributed during system initialization to perform homomorphic encryption operations and obtain the ciphertext.
[0069]
[0070] Among them, Enc pk () indicates that the homomorphic encryption public key pk is used to encrypt the function.
[0071] The drone will encrypt the ciphertext portion. Part of plaintext transmission Uploaded together to the server, the uploaded content is
[0072]
[0073] Furthermore, the specific process of step 4 is as follows:
[0074] Step 4.1: Server Aggregation and Distribution
[0075] The server receives the upload from step 3. The plaintext and encrypted transmission portions are aggregated separately. For the plaintext portion, aggregation is performed directly.
[0076]
[0077] For the ciphertext portion, the homomorphic addition property is used for encrypted aggregation.
[0078]
[0079] We obtain the aggregated plaintext and ciphertext portions. The data was distributed to each drone terminal.
[0080] Step 4.2: Unmanned Aerial Vehicle (UAV) Decryption and Model Synchronization
[0081] The drone uses the private key sk to receive the encrypted portion. Decrypt
[0082]
[0083] in, This is the aggregated result of the decrypted encrypted portion; Dec sk () indicates that the function is decrypted using the homomorphic encryption private key sk.
[0084] Then the parameters obtained from decryption With plain text Combined, restore the complete global model parameter update.
[0085]
[0086] The drone updates its local model accordingly.
[0087]
[0088] This round of federal learning has concluded.
[0089] The present invention also provides a system for implementing the above method, characterized in that it includes the following functional modules:
[0090] System initialization module: Used to initialize the local model parameters w on the UAV. 0 Register the remaining energy of the drone With bandwidth It also completes the pre-deployment of the homomorphic encryption public key pk and private key sk, ensuring that each client can directly call them in subsequent training and encrypted communication.
[0091] Local training and sensitivity calculation module: used to enable the drone to perform training based on its local dataset in each round of training. For global model parameters w t Training is performed, and the parameter update difference is calculated. and gradient Based on this, a set of local parameter sensitivities is generated. At the same time Energy and bandwidth information is uploaded to the server.
[0092] Global Mask Generation Module: The server collects sensitivity sets and resource status information from all drones. Mask selection is modeled as a "privacy gain density" optimization problem with energy consumption and bandwidth constraints. A greedy algorithm is used to jointly sort and filter parameter sensitivity and resource constraints, ultimately generating a globally unified mask vector M. tThe data was then distributed to each drone terminal.
[0093] Encryption and Upload Module: Used for receiving globally unified mask M on the drone end. t The parameters are then divided according to the mask: homomorphic encryption is performed on the "protected part" to obtain a subset of encrypted parameters; the "non-sensitive part" remains in plaintext. The drone then uploads both the encrypted and plaintext parameters to the server.
[0094] The server-side aggregation and decryption update module: The server directly averages the received "plaintext parameter portion" and performs homomorphic addition aggregation on the "encrypted parameter portion" to generate the aggregated update result; this result is then sent to each drone. The drone uses its local private key sk to decrypt the aggregated encrypted portion and merges it with the plaintext portion to obtain the complete parameter update content w. t+1 This completes the model synchronization for this round.
[0095] The beneficial effects of this invention are:
[0096] (1) Effectively reduce communication and energy consumption costs and enhance practicality in resource-constrained environments.
[0097] To address the issue mentioned in the background section regarding the significant computational and communication costs associated with full homomorphic encryption in UAV scenarios, this invention introduces energy consumption and bandwidth constraints into the system design and employs global mask optimization to precisely select the parameters requiring encryption, thereby avoiding indiscriminate full encryption. Compared to traditional solutions, this invention minimizes redundant encryption operations while ensuring global privacy protection, effectively reducing communication bandwidth consumption and energy expenditure. This makes the method better suited to the low-altitude, economical application environment of UAVs, where computational and communication resources are limited, thus improving the deployability and sustainability of federated learning.
[0098] (2) Improve the overall level of privacy protection and ensure the overall security and fairness of the system.
[0099] Addressing the issue mentioned in the background technology that "existing methods such as MaskCrypt only protect highly important parameters of some users, easily leading to privacy imbalances," this invention employs a global mask consensus mechanism. While considering parameter sensitivity, it combines a greedy optimization method based on privacy benefit density, enabling global parameter selection to simultaneously balance privacy protection and resource consumption. This mechanism avoids the drawback of protecting only the important node parameters of a single drone while neglecting global privacy, ensuring that all drone clients enjoy balanced security guarantees in the overall collaboration. Ultimately, this invention not only improves the overall privacy protection level of the system but also enhances the controllability and robustness of the method, thereby guaranteeing the fairness and credibility of federated learning in multi-user drone scenarios. Attached Figure Description
[0100] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0101] Figure 1 This is a schematic diagram of the process of the UAV federated learning global mask selection homomorphic encryption method under the constraints of energy consumption and bandwidth. Detailed Implementation
[0102] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments:
[0103] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.
[0104] This invention provides a resource-constrained UAV federated learning selective homomorphic encryption method, the overall process of which is illustrated below. Figure 1 As shown. The method includes the following main processes:
[0105] like Figure 1 As shown, the first step includes system initialization, local model training for the drone, and sensitivity calculation and uploading:
[0106] System Initialization: In one specific embodiment, the system comprises several UAV terminals and a centralized server, connected via a wireless communication network. Each UAV includes at least: an onboard computing unit for performing local model training, parameter encryption and decryption, a wireless communication module for uploading parameters and distributing models to the server, a power management module for detecting remaining energy, and a storage module for storing local datasets, model parameters, and key information. The centralized server includes at least: a high-performance computing module for performing global mask generation and parameter aggregation, a large-capacity storage module for storing sensitivity sets and encrypted parameters uploaded by each UAV, and a wireless network interface module for concurrent communication with multiple UAVs. During the initialization phase, the server predetermines the global model structure and generates public-private key pairs supporting homomorphic encryption, unifying local training parameters: local training epochs (e.g., 1-5 epochs), learning rate range (e.g., 0.001-0.01), and public-private key pairs, which are then distributed to the UAVs via the wireless network. During the initialization phase, each UAV reads its remaining power and available bandwidth through its power management module and communication module.
[0107] UAV local model training: The UAV trains the received global model based on locally collected data to obtain local model parameters. If a gradient anomaly occurs, the current round of training is terminated, the anomaly information is uploaded, and the UAV does not participate in the current round of global mask generation.
[0108] Sensitivity calculation and upload: Compare the model parameters before and after training, calculate the parameter update difference and its corresponding gradient, and calculate the sensitivity of each parameter dimension according to the preset formula and upload it to the server. If the remaining energy and bandwidth cannot meet the minimum communication requirements, only the resource status information will be uploaded, and the drone will not participate in this round of global mask generation.
[0109] like Figure 1 As shown, the second step involves a greedy selection algorithm to generate the global mask:
[0110] Global Mask Generation: The server calculates the privacy benefit density based on parameter sensitivity and resource cost. Energy consumption and bandwidth are both weighted at 0.5 (this can be modified according to specific implementation preferences), and all parameters are sorted in descending order of benefit density. Starting with the highest priority parameter, the server checks each parameter one by one. If the current selection will not cause any drone's communication bandwidth or energy consumption to exceed limits, the parameter is marked as "encrypted." Otherwise, the parameter is skipped, and the server continues to the next parameter. Finally, a unified global mask is generated and distributed to all drones.
[0111] like Figure 1 As shown, the third step includes parameter partitioning, encryption, and uploading:
[0112] Parameter partitioning, encryption, and uploading: The drone divides its local parameters into encrypted and plaintext parts based on a globally unified mask. Homomorphic encryption is performed on the encrypted part, and both the plaintext and ciphertext parameters are uploaded to the server. If the encryption process fails or times out, an error message is uploaded, and the drone will not participate in this round of aggregation.
[0113] like Figure 1 As shown, the fourth step includes server aggregation and distribution, and drone-side decryption and model synchronization:
[0114] Server aggregation and distribution: The server performs aggregation directly on plaintext parameters and homomorphic addition aggregation on ciphertext parameters. After obtaining the aggregation result, it distributes it to the drone.
[0115] Drone-side decryption and model synchronization: The drone uses its private key to decrypt the aggregated ciphertext portion and merges it with the plaintext result to form a complete model update. If decryption fails, only the plaintext aggregated portion is used to update the model, and the drone re-participates in the complete process in the next round.
[0116] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.
Claims
1. A resource-constrained UAV federated learning selective homomorphic encryption method, characterized in that, The steps include: Step 1): Perform system initialization on the drone, including presetting local model parameters, registering remaining energy and available bandwidth, and pre-deploying homomorphic encryption public and private keys; During local training, the parameter update difference and gradient information are calculated, and a sensitivity set is generated and uploaded to the server accordingly. Step 2): The server collects the sensitivity set, remaining energy and bandwidth status of all drones, introduces communication and energy consumption constraints, models the mask selection as a privacy gain density optimization problem, and uses a greedy algorithm based on parameter-level homomorphic encryption decision to generate a globally unified mask and send it to the drones. Step 3): The UAV receives the globally unified mask, divides the parameters according to the mask, performs homomorphic encryption on the parts that need to be protected, keeps the non-sensitive parts in plaintext form, and uploads the encrypted part and the plaintext part together to the server. Step 4): The server performs direct aggregation on the plaintext part and aggregates the encrypted part using the homomorphic addition property to generate a complete aggregation result and send it to the drone. The drone uses its private key to decrypt the encrypted part and combines it with the plaintext part to form a complete parameter update, thereby achieving local model synchronization.
2. The method according to claim 1, characterized in that, The parameter sensitivity is calculated by multiplying the absolute value of the parameter update difference with the corresponding gradient, and the sensitivity vector is normalized to generate a sensitivity set.
3. The method according to claim 1, characterized in that, The process of generating the globally unified mask includes: calculating the privacy gain density based on parameter sensitivity and resource cost; sorting the privacy gain density values of all parameters in descending order; and selecting encryption parameters one by one according to priority under the condition of satisfying communication bandwidth and energy consumption constraints, and finally generating a globally unified mask applicable to all UAVs.
4. The method according to claim 1, characterized in that, The communication constraint is that the sum of the uploaded size of encrypted parameters and the uploaded size of plaintext parameters must not exceed the available bandwidth of the UAV.
5. The method according to claim 1, characterized in that, The energy consumption constraint is that the total energy consumed by encryption calculation, decryption calculation, and uploading encryption parameters must not exceed the remaining energy of the drone.
6. The method according to claim 1, characterized in that, The homomorphic encryption algorithm is a public-key encryption algorithm that supports homomorphic addition. When the server does not have decryption capabilities, it only performs homomorphic addition on the encrypted part and can complete parameter aggregation without decryption.
7. A system for implementing the method as described in any one of claims 1 to 6, characterized in that, include: The system initialization module is used to initialize the local model parameters, energy consumption status and communication bandwidth of the UAV, and to complete the deployment of homomorphic encryption public and private keys; The local training and sensitivity calculation module is used to perform local training and calculate parameter update differences, gradients, and sensitivity sets; the global mask generation module is used by the server to generate a globally unified mask based on privacy gain density optimization under energy consumption and bandwidth constraints. The encryption and upload module is used by the drone to perform homomorphic encryption on the parts that need to be protected based on a global mask and then upload them. The server aggregation and decryption update module is used to perform plaintext direct aggregation and encrypted homomorphic aggregation, and the model is updated after being decrypted by the drone.