A privacy protection method for crowd intelligence perception based on edge computing and federated learning

Abstract

Mobile crowd sensing, as a new type of perception data collection paradigm, has been applied to various social scenarios, but it also faces serious privacy information leakage problems. Existing methods mainly protect the original data under the traditional centralized architecture, and do not fully consider the problems of internal theft and data interception by external attackers. Based on this, the present invention proposes a crowd sensing privacy protection method based on edge computing and federated learning. First, the participants perform model training locally, perturb the model parameter values ​​using localized differential privacy, and upload them to the edge computing node after process encryption using homomorphic encryption; then, the edge computing node performs edge aggregation on the local model parameters, and uploads the edge model parameters updated in collaboration with the participants to the perception platform; finally, the perception platform receives the encrypted edge model parameters, performs global aggregation and processing operations in ciphertext form, thereby effectively preventing the leakage of participants' privacy information.

Description

Technical Field

[0001] The present invention belongs to the field of mobile crowd intelligence perception, and specifically relates to a privacy protection method based on edge computing and federated learning. Background Art

[0002] Mobile Crowd Sensing (MCS) is a new form of data collection in the field of Internet of Things (IoT). It combines mobile sensing and crowdsourcing ideas, and can quickly and easily collect a large amount of sensing data through various sensors built into user mobile devices, and study and process the data through the sensing platform, so as to better provide users with personalized services. MCS can be applied to various fields, such as intelligent transportation, situational education, health care, environmental monitoring, etc., promoting the comprehensive construction of smart city services.

[0003] In the process of performing perception tasks, participants will inevitably share perception data with the platform, and these perception data often contain various important private information, such as the identity information, location information, and behavior information of the participants. However, the perception platform, as a data processing and control center, is often untrustworthy. When it aggregates and processes the data of participants, it may leak the sensitive information of participants. At the same time, when collecting and submitting data, participants may also be threatened by internal attackers stealing data and external attackers intercepting data. These will cause serious privacy issues and hinder the development of MCS.

[0004] Research on MCS privacy protection mainly focuses on the traditional centralized MCS architecture. The perception data of participants will directly interact with the perception platform, which will lead to the centralized leakage of participants' privacy. At the same time, the perception platform will bring high transmission delay and cost when processing massive data, which is not suitable for large-scale perception tasks. Based on the marginality and efficiency of edge computing, the present invention uses edge computing nodes to separate the direct interaction between participants and the perception platform, builds a new distributed group intelligence perception platform, and further considers the unreliability of edge computing nodes.

[0005] The current privacy protection method mainly protects the original data collected by the participants. If attacked, the sensitive data of the participants will be directly leaked, which will have serious consequences. Therefore, federated learning can be used to enable participants to train model parameters locally, and upload the model parameters to the edge computing nodes and perception platform for updating to avoid direct data leakage. However, model information may also lead to indirect leakage of private data. The existing methods do not consider the usability and efficiency issues when using privacy protection technology. Therefore, the present invention integrates different privacy protection technologies, hoping to achieve a balance between privacy and efficiency, thereby providing participants with stronger security protection. Summary of the invention

[0006] In view of the above-mentioned existing research methods and existing problems, the present invention simultaneously introduces edge computing (EC) and federated learning (FL), and combines local differential privacy (LDP) with homomorphic encryption (HE), and proposes an efficient mobile crowd intelligence perception privacy protection method. EC, as a new computing model that performs calculations at the edge of the network, can transform the traditional centralized MCS architecture into a distributed MCS architecture, thereby preventing the perception platform from leaking the privacy data of participants to a certain extent and improving transmission efficiency. In the FL mode, participants keep their private data locally without uploading it to the edge computing node and the perception platform, thereby preventing the participant data from being directly leaked due to attacks. At the same time, LDP is used to perturb the model parameter values ​​locally, and HE is used to perform process encryption when uploading to the edge computing node and the perception platform, thereby efficiently protecting the model information of the participants. The system model structure proposed in the present invention is as follows: Figure 1 As shown, it mainly includes three parts: local training of participants, aggregation and update of edge computing nodes, and aggregation and processing of perception platforms.

[0007] Participants perform local training. Participants collect data through their personal smart terminal devices according to the perception tasks issued by the requester, use the perception data to train the model locally, and apply LDP and HE privacy protection technologies to achieve noise perturbation and process encryption. Then, the local model parameters are uploaded to the edge computing node for edge aggregation and iterative updates. The global model is built while preventing the leakage of participants' privacy information, and providing a reliable and effective source of information for the requester.

[0008] Edge computing node aggregation and update. Edge computing nodes have more storage and computing resources than mobile devices. Deploying them at the edge of the network can not only request content and services from the perception platform, but also collaborate with participants to process information. It mainly collects local perturbations and encrypted model parameters sent by participants for edge aggregation, and sends the updated edge model parameters to the perception platform. Since edge computing nodes are untrusted, aggregation and update operations are performed in encrypted form with the support of HE technology.

[0009] Aggregation and processing of the perception platform. The perception platform is the data processing and control center in the MCS, with powerful storage and computing capabilities. After receiving the encrypted edge model parameters sent by the edge computing node, the perception platform uses HE technology to aggregate and update the global model in ciphertext form, and sends the final processing results to the requester in encrypted form to complete the perception task.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) By introducing edge computing nodes in MCS, the centralized MCS architecture is transformed into a distributed MCS architecture, and a new edge computing-supported MCS platform is built, which is conducive to avoiding the centralized leakage of participant data, while reducing the burden on the perception platform, and is more suitable for large-scale perception tasks. (2) By implementing the perception task under the privacy-preserving machine learning mode of FL, the sensitive data of the participants can be retained locally, avoiding the leakage of their original data information by the edge computing nodes and the perception platform. (3) By combining the two privacy protection technologies of LDP and HE, information obfuscation and process encryption are achieved. In this way, sensitive information will not be leaked between participants, nor will it be leaked from edge computing nodes and perception platforms. It can also resist attacks from external data interception, effectively ensuring the privacy of participants. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is the overall framework of the present invention.

[0012] Figure 2 It is the overall flow chart of the present invention. DETAILED DESCRIPTION

[0013] The overall implementation process of the present invention is as follows Figure 2 In the specific implementation steps, there are three main stages: local training of participants, aggregation and update of edge computing nodes, and aggregation and processing of perception platforms.

[0014] Participant local training: Participants receive a public / private key pair (pk, sk) generated by the key generation center, where the private key sk is used to decrypt the model parameters encrypted by the public key pk. The private key sk cannot be accessed except by the participant's own edge computing node and perception platform. Assume that t1 is the number of local updates of an edge aggregation, t2 is the number of edge aggregations of a global aggregation, and T is the total number of local updates, which is an integer multiple of t1t2. In round t, participant x i Download the latest encryption model parameters from the corresponding edge computing node Then, decrypt using the private key sk get And the existing local model parameters Replace with To update the local model, prevent overfitting due to training only on local data. Then, participant x i Based on the perception data D collected by the mobile device i And the decrypted model parameters Jointly train local models Where η represents the fixed learning rate of the participant during local training, Indicates the participants' current model parameters The average gradient of the local data is

[0015] Get a new round of model parameters by local training every t1 rounds Afterwards, participant x i The Laplace mechanism of ε-LDP technology is used to add random perturbation noise to the updated local model parameters to achieve information fuzzification: in, represents Laplace noise, Δf represents local sensitivity, and ε represents the privacy budget set for each participant.

[0016] Participant x i Use the public key pk to perturb the local model parameters Encrypted Then the perturbed and encrypted local model parameters And the related local loss function It is sent to the edge computing node for a new round of edge aggregation and update. In the process of executing HE, the Paillier algorithm only supports the calculation of integer fields, that is, the model parameter value that needs to be encrypted should be Z nAn element in the range. However, model parameters are usually in the form of floating-point numbers. To implement encryption of model parameters, an additional rounding factor L needs to be introduced, which is greater than the maximum number of decimal places of the model parameters. Therefore, for floating-point parameters, by multiplying by 10 L Encoding it as an integer parameter can be used for encryption, and the opposite operation is performed when decoding.

[0017] Aggregation and update of edge computing nodes: In round t, the edge computing nodes download the latest encrypted global model parameters from the perception platform. And the existing edge model parameters Replace with Then, each edge computing node randomly selects k participants in the ratio of ρ to participate in this round of training, that is, k←max(Kρ,1). After each t1 round of local training of the participants, the edge computing node will receive the new round of perturbations and encrypted local model parameters of the corresponding participants. And the local loss function Then the local model parameters of the participants are aggregated at the edge, that is, a weighted average is used for the received model parameters: Where N represents the number of training data for each edge computing node, N i represents the dataset size of the ith participant, S t It represents the set of k participants randomly selected by the edge computing node in round t. Due to the symmetry of the Laplace mechanism, the noise is almost eliminated during the edge aggregation operation, so it will not have a significant impact on the accuracy of the model.

[0018] After edge aggregation, the edge computing node sends the aggregated encrypted edge model parameters to all corresponding participants for a new round of iterative updates. After each t2 rounds of edge model aggregation, the edge computing node will send a signal to each participant to stop model training and send the final encrypted edge model parameters to all participants. And the marginal loss function Sent to the perception platform for global aggregation and update, where This means that every t1t2 rounds of local updates, participants communicate with the perception platform.

[0019] Aggregation and processing of the perception platform: In the global cycle t = 0, the perception platform first initializes the model parameter ω0, and sends the original model parameter ω0 to all edge computing nodes, which are then broadcast by the edge computing nodes to the corresponding participants to initialize the local model parameters and perform preliminary training. When t>0, the perception platform sends the latest encrypted global model parameters to the edge computing nodes for collaborative training with the participants. After each t2 rounds of edge aggregation, the perception platform receives the encrypted edge aggregation results of the edge computing nodes and local joint training Then perform the aggregate averaging of global model parameters: Where J represents the number of edge computing nodes participating in the training, M represents the total amount of training data for all edge computing nodes, and M j represents the number of training data for the jth edge computing node,

[0020] The perception platform also needs to receive the edge loss function from the edge computing node And aggregate to get the global loss function, that is The perception platform then sends the latest encrypted global model parameters And the global loss function It is sent to all edge computing nodes for the next iteration of training. When the global model update round reaches the maximum value T, the perception platform sends a signal to each edge computing node to stop model training and sends the final encrypted global model parameters to the requester, thus completing the perception task.

[0021] The above-mentioned implementation method is a preferred implementation method of the present invention, but the implementation method of the present invention is not limited to the above-mentioned method. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A crowd-sensing privacy protection method based on edge computing and federated learning, characterized in that: The method is applied to a new distributed crowd-intelligence perception architecture including a participant local training module, an edge computing node aggregation and update module, and a perception platform aggregation and processing module, and includes: In the participant local training module, participants collect original perception data through smart terminal devices, use the federated learning paradigm to perform model parameter training locally, and use local differential privacy and homomorphic encryption technology to perturb and encrypt model parameter values. Then, the perturbed and encrypted local model parameters and related local loss functions are sent to the edge computing node, thereby preventing the participants' privacy information from being maliciously stolen and intercepted. The specific implementation methods of local training model parameters and perturbation and encryption include: In round t, participant x i The calculation formula for training local model parameters using the federated learning paradigm is: in, represents the local model parameters obtained by the participant using the private key sk to decrypt the edge model parameters of the previous round, η represents the fixed learning rate of the participant during the local training process, Indicates the participants’ current model parameters The average gradient of local data; the mechanism of using localized differential privacy to perturb the model parameter value is a symmetric Laplace mechanism, and the calculation formula is in, represents Laplace noise, Δf represents local sensitivity, and ε represents the privacy budget set for each participant; The homomorphic encryption technology used is the Paillier algorithm, which uses the public key pk to perturb the local model parameters Encrypted In the edge computing node aggregation and update module, after receiving the perturbations and encrypted local model parameters and local loss functions sent by the participants, the edge computing nodes perform edge aggregation by weighted average, and iteratively update with the participants. Then, the final encrypted edge model parameters and edge loss functions are sent to the perception platform to avoid serious privacy leakage caused by the perception platform's centralized collection of all participant information. In the perception platform aggregation and processing module, after receiving the encrypted edge model parameters and edge loss function jointly trained by the edge computing node and the local, the perception platform aggregates and averages the global model parameters in ciphertext form, and aggregates to obtain the global loss function, and then iteratively trains with the edge computing node. When the global model update round reaches the maximum value, the final encrypted global model parameters are sent to the requester to complete the perception task. Since the perception platform cannot obtain the private key and the local model parameters of the participants are unknown, it effectively prevents the privacy information of the participants from being leaked to untrusted perception platforms.

2. The method for protecting privacy based on edge computing and federated learning according to claim 1, characterized in that: In the edge computing node aggregation and update module, the edge computing node receives the perturbations and encrypted local model parameters sent by the participants. Then, through the formula Perform weighted averaging to obtain edge model parameters Where N represents the number of training data for each edge computing node, N i represents the dataset size of the ith participant, k represents the number of participants randomly selected by the edge computing node in proportion to ρ, S t Represents the set of k participants randomly selected by the edge computing node in round t.

3. The method for protecting privacy based on edge computing and federated learning according to claim 2 is characterized in that: After the edge computing node aggregation and update module is executed, in the perception platform aggregation and processing module, the perception platform receives the encrypted edge model parameters of the edge computing node and the local joint training. The global model parameters are obtained by aggregation and averaging in ciphertext form. The calculation formula is Where J represents the number of edge computing nodes participating in the training, M represents the total amount of training data for all edge computing nodes, and M j represents the number of training data for the jth edge computing node,

Images