학술

Hiding in plain sight: Adversarial identity concealment for facial gender recognition in edge computing

PLOS ONE
CC BY
이 매체는 공공·자유 라이선스로 본문을 직접 표시합니다.

Figures
Abstract
With the development of the edge computing, privacy protection problems face new challenges due to the end-to-edge distributed computation on different devices of different owners. In order to protect privacy with high efficiency, we propose an idea of transmitting only the information related to main tasks to remote offloading devices, and keeping the private information on the user device itself or nearby safe devices. This paper proposes a distributed layout privacy protection strategy to ensure the bidirectional tasks of gender recognition and privacy protection. The key is to design a favorable feature extraction module with privacy concealment, which runs on user device or nearby safe device through adversarial fine-tuning. And other modules like attribute feature analysis and inference module and attribute classification module, run on other servers. Two effective privacy protected gender identification methods of the local and global fine-tuning method are proposed. The local fine-tuning method freezes either of the feature extraction module or the inference modules, while global fine-tuning method adjusts the weights of all the modules simultaneously. Experimental results show although the local fine-tuning method is more accurate and effective, the convergence speed of the global fine-tuning method is faster and the training process is more concise.
Citation: Yao Y, Zhao S, Ni Z (2026) Hiding in plain sight: Adversarial identity concealment for facial gender recognition in edge computing. PLoS One 21(11): e0362254. https://doi.org/10.1371/journal.pone.0362254
Editor: Feng Ding, Nanchang University, CHINA
Received: November 3, 2025; Accepted: October 22, 2026; Published: November 20, 2026
Copyright: © 2026 Yao et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: No data was generated by this study. The following existing data source was used: the Labeled Faces in the Wild (LFW) dataset from the University of Massachusetts Amherst, available at http://vis-www.cs.umass.edu/lfw/ via web address. The LFW dataset is publicly available for non-commercial research purposes and requires no special access permissions.
Funding: The work of Zhengwei Ni was supported in part by the Tongxiang General Artificial Intelligence Research Institute Project TAGI2-B-2024-0017. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
1 Introduction
With the popularization of intelligent Internet devices such as smart phones and wearable devices, and the real-time performance requirements of various intelligent applications, the limited network bandwidth in the cloud computing architecture will become a bottleneck. Mobile Edge Computing [1] is to offload the computing tasks of users’ intelligent applications to the adjacent edge servers, which can reduce the overhead of local resources, alleviate the communication delay and communication congestion caused by cloud computing, improve the task computing efficiency, and meet the requirements of user application service quality. However, since computing of the tasks are to be offloaded and distributed to edge servers or other service nodes, and the owners of the edge nodes or servers may be in various forms, an emerging privacy protection problem for edge computing arise. For instance, the intermediate processing results of the server probably contain private data, and the attacker may violate the privacy by attacking the task data offloaded to the edge servers [2] or a remote server. Privacy may also be leaked during transmission. The privacy attribute protection provided by the existing edge computing is not perfect.
So far, many privacy protection methods have been proposed. Most of them consider one-time transformation, scrambling, and encryption [3–5], which will introduce a lot of computation load not ignorable for the system.Moreover, the terminal in edge computation environment is very sensitive to computation load increase.
At present, there are many edge intelligent applications including various face attribute features extraction. While identifying certain attributes, there may be a risk of extracting the characteristics of identity attributes at the same time, since there are many correlations between identity features and these task related attributes. It is a great challenge to protect user privacy from being stolen by edge servers, while dispersing tasks on each node without adding extra computation.
Therefore, in this paper, we propose a face image privacy protection method suitable for edge intelligence. We focus on face gender recognition as the main task taking the identity information as the privacy information to discuss the privacy protection strategy. A novel network model is proposed with two feature inference networks for the attributes of gender and identity respectively, and a privacy-concealment feature extraction network with functions of identity feature concealment and gender feature extraction. Through two-way adversarial adjustment through the gradient reversal layer (GRL) [6–8], the information output by the privacy-concealment feature extraction network only includes gender feature, without identity attribute features.
By using the gradient reversal layer (GRL), the identity-sensitive information is blanked at the near-user terminal, so that there is no sensitive information transmitted to the remote server. At the same time, the information related to the main task to transfer to remote devices.
Since this method adopts the adversarial training method [9,10]and has many modules, it is hard to use the global training method to achieve the purpose of gender recognition and identity privacy concealment at one step as most existing networks do. The parameters can easily get trapped in local optima. To solve it, we propose multi-step training methods based on sub-module training which employ specific loss functions to iteratively update specific module parameters.
The contributions of this paper can be summarized as follows:
- This paper proposes a privacy protection strategy for edge intelligence, which is characterized by end-to-edge distributed computation on different devices of different owners. We propose to transmit only the information related to main tasks to remote offloading devices, and keep the private information on the user device itself or nearby safe devices.
- This paper adopts the shared privacy-concealment feature extraction network to realize gender feature extraction and identity privacy feature concealment. It is deployed on the safe device close to the user end to protect privacy information. Gradient reversal layer (GRL) is added to the network to facilitate the task of privacy identity concealment.
- The two-fold functions are hard to realize at one time. This paper designs multi-step privacy-concealment methods of local and global fine-tuning based on adversarial training. The difference is that local fine-tuning method freezes either of the feature extraction module or the inference modules, while the global fine-tuning adjusts the modules simultaneously.
- As the network architecture has multiple modules, the fine-tuning scheme can adjust specific module through independent iterative training and corresponding specific loss functions. Compared with the traditional global training method, it is not easy to fall into the local optimum, and the convergence speed will be faster.
- Experiments are performed on the LFW dataset, showing that the proposed strategy, whether local fine-tuning or global fine-tuning, achieves a high gender recognition rate and the identity privacy concealment at last. Though the local fine-tuning achieves better privacy protection performance than global fine-tuning, global fine-tuning is simpler and converges faster.
The rest of this paper is organized as follows. In Section II, we discuss the related work. In Section III, we present our ideas and the proposed approach. Our method is experimentally evaluated in Section IV. Section V summarizes the full paper.
2 Related works
In edge intelligence, uploading data to edge servers introduces challenges to privacy protection. Privacy protection for machine learning applications may cause significant computing costs [11–13], making privacy protection in edge computing more challenging. Meanwhile, most of present works of image privacy protection are for face recognition applications [14,15], but not for recognition of other facial attributes.
For example, Wang et al. [16] propose a privacy preserving facial recognition technology based on convolutional neural networks, which is used for facial feature extraction. In order to overcome the problem of privacy leakage, a nearest neighbor algorithm for calculating cosine similarity on encrypted feature vectors is proposed. But the encryption process introduces enormous computational complexity. Cheng et al. [17] propose a privacy protection scheme based on secure nearest neighbor algorithm, which can calculate and compare the cosine similarity of feature vectors under real ciphertext. Then the encrypted data is distributed to edge computing devices through cloud server. Similar to the work in [16] the encryption process introduces enormous computational complexity. The encryption enormous computational cost is not ignorable to the terminal in edge computation environment, which is very sensitive to computation load increase.
With the proposal of federated learning [18], some researchers introduce federated learning into edge computing. In federated learning, multiple participating devices have their own local data. These participants do not need to upload data to a central server, but instead train a unified model through shared model updates [19]. For example, Liu et al. [20] propose a privacy protection framework named P2FEC combining federated learning and edge computing, which can build a unified deep learning model across multiple users or devices without uploading their data to a centralized server. Lu et al. [21] propose a privacy protection data sharing mechanism for distributed multi-party Internet of Things(IoT) applications, which incorporates federated learning into the licensed blockchain. However, federated learning requires a large amount of communication between the client and servers, which lead to low communication efficiency. At the same time, due to the scattered data on different clients, federated learning cannot use all the data for training like traditional machine learning methods. Different clients may have different amounts and qualities of data, which lead to data imbalance and a decrease in model accuracy.
Current research on edge intelligence generally focuses on optimizing the performance and efficiency of deploying deep learning technology to edge intelligence environments. They solve problems such as Quality of Experience(QoE) [22], task deadline time [23], energy consumption, transmission cost and communication delay [24]. The enormous computational burden brought in by privacy protection remains a problem on edge intelligence.
As edge computing is a relatively new field, there are relatively few related research articles on privacy protection. Thus, we survey other works on multi-server privacy protection such as the Secure Multi-party Computing method [25]. These multi-party security computing methods typically use multiple cloud servers [26,27]. The Secure Multi-party Computing method effectively solves the problem of data privacy and security by computing together among multiple participants without sharing the original data [28]. Ma et al. [29] propose a secure face authentication scheme, that adopts a convolutional neural network to extract the feature vector of the face, use two non-collusion servers to respectively store the user’s encrypted face features and complete face verification. All data is transmitted in ciphertext state, and no other device except the face authentication server can decrypt the data. Cai et al. [30] propose a secure face recognition scheme in a joint outsourcing cloud environment, which can effectively protect the privacy of user data for a semi-trusted model. Haghighat et al. [31] use secure circuits and the secure multi-party computing technology of homomorphic encryption to realize face recognition in the encrypted domain. Secure Multi-party Computing typically involves a large amount of computation and communication between multiple parties, leading to higher computation and communication costs. At the same time, when processing large-scale data, the encryption process significantly slows down computation speed.
The machine learning algorithm operating on the encrypted features, not only serves the purpose that the original data cannot be obtained, i.e., the purpose of privacy protection, but also has the advantages of accurate reasoning and recognition in further steps. However, its encryption technology introduces long calculation delay and high energy consumption to the system [16,17], which requires us to further improve the encryption algorithm.
In terms of privacy protection, early works mainly focus on facial privacy protection in the field of face recognition and location privacy protection in the field of intelligent transportation [32]. Generally, they depend on one time transformation, scrambling, and encryption [33,34]. But it often incurs high computation load not ignorable for the system. The terminal in edge computation environment is very sensitive to computation load increase [13,35]. Therefore, to use machine learning to protect privacy in edge computing, the first problem to be solved is the high computational complexity.
Given that many applications focus on facial attributes rather than facial recognition, we are committed to making contributions in this field. We use adversarial methods in machine learning to directly extract features that do not contain privacy information, which cannot be recovered in subsequent processing. Our method obtains privacy-free features related to face attributes through adversarial training, which only introduces extra computation costs in the training phase. It does not introduce any additional computational burden in real-time computation phase, thus successfully solves the problem of high computational complexity in present privacy works.
3 Methodology
In this paper, we take the gender recognition based on face as the main task, the identity as the privacy information to discuss privacy protection strategies in edge computing scenarios. We propose adversarial training methods to ensure the bidirectional tasks of face attribute feature extraction and privacy feature concealment. Especially, we design a good feature extractor to separate the gender information and the sensitive identity information intertwined in the original data domain. As shown in Fig.1, we divide the network into three modules: the privacy concealment feature extraction module with gender feature extraction, the attribute feature analysis and inference module and the attribute classification module. Different modules run on different types of entities as shown in Fig. 1.
The privacy concealment feature extraction module runs on the safe terminals that are relatively close to the user, which can ensure that no privacy features are transmitted to distant edge servers [36,37]. The attribute feature analysis and inference module and the attribute classification module can run on the devices far away from the user, because there is no privacy feature in the modules anymore.
We expect that the output of the privacy concealment feature extraction module can be used in the attribute feature analysis and inference module to achieve the two-fold functions of gender recognition and identity privacy concealment. This training process requires adversarial fine-tuning of the attribute feature analysis and inference modules and the privacy concealment feature extraction module.
3.1 Blocks in the architecture
3.1.1 Privacy concealment feature extraction module.
This module is composed of a convolutional neural network (CNN). The input to the module is the image dataset with gender classification labels and identity classification labels. The image input is represented as , where X represents the image input space, and the image classification label is , where Y = {1,2,3...k} represents the image classification label space. When classifying gender, k = 2; when classifying identity , the value of d depends on the specific number of identities. The recognition of gender and identity uses the same dataset. In the training stage, the face image formed by the image input x after preprocessing is transformed into an D-dimensional feature vector after the mapping transformation of , and the preliminary feature extraction is completed.
The privacy-concealment feature extraction module is to extract the low-level features of the attributes. The higher the network level, the more the network will focus on the main task and will also try to eliminate sensitive information. We employ the GRL to perform privacy concealment. What the GRL does is to multiply the error passed to this layer by a negative number (), which will make the networks before and after the GRL have opposite training goals to achieve a confrontational effect [9].
3.1.2 Attribute feature analysis and inference module.
The feature maps extracted by the privacy-concealment feature extraction module will be input to the attribute feature analysis and inference module. The output of the privacy concealment feature extraction module will be divided into two branches, and input into the gender inference and classification module and the identity inference and classification module . represents the weight parameter for gender feature analysis, and represents the weight parameter for the identity feature analysis.
This gender inference module consists of a multi-layer CNN and fully connected layers. We expect that the gender can be recognized by the gender inference module. If it is not the case, fine-tuning on the gender related modules is required to ensure that gender information can be accurately extracted.
The feature maps output by the privacy-concealment feature extraction module is sent to the module for identity inference. When the privacy concealment is not achieved, it will infer the identity information successfully. If this is the case, privacy concealment-feature extraction module will be fine-tuned using the GRL to continue the privacy concealment process.
3.2 Attribute classification performance evaluation and optimization objective
3.3 Local fine-tuning training of the model
The purpose of training is to achieve high accuracy of gender recognition and ensure that the accuracy of identity recognition is ultimately lower than 10% for identity privacy protection. If a certain purpose is not achieved, freeze the network of privacy concealment feature extraction or the network of the attribute feature analysis and inference module to iteratively update the network parameters to perform adversarial fine-tuning. The network flow chart is shown in Fig. 2.
3.4 Global fine-tuning training of the model
The global fine-tuning method is different from the local fine-tuning method. If the purpose of training is not achieved, adversarial global fine-tuning of the relevant modules of identity hiding and gender recognition is required, and the network parameters are updated iteratively. No fixed modules are fixed while fine-tuning the privacy feature concealment module and corresponding feature inference module towards the certain function for gender identification or identity protection. The network flow chart is shown in Fig. 3.
4 Experiments
4.1 Dataset
Experiments are performed on the LFW face recognition dataset, which is an unconstrained natural scene face recognition dataset. It consists of more than 13,000 face pictures of famous people with different orientations, expressions and lighting environments. There are more than 5,000 people, of which 1,680 have face pictures more than one. In this experiment, 24 individuals were selected from the LFW database for data training with identity tags 023. Among them, 80% of the pictures are randomly selected for training, and the remaining 20% for testing.
Furthermore, we provide a supplementary clarification regarding the rationale for dataset selection, explicitly articulating the alignment between the unconstrained, in-the-wild characteristics of the LFW dataset and the privacy-preservation objectives of this study. Although contemporary works on facial attribute recognition commonly employ datasets such as CelebA or VGGFace2, which offer more consistent attribute annotations or larger scales, we opt for LFW due to its data distribution encompassing highly diverse variations in pose, expression, and illumination under natural conditions—i.e., the classic “in-the-wild” benchmark. Facial images captured by edge intelligent devices in real-world scenarios are often non-ideal and unconstrained, a condition that LFW emulates effectively. This facilitates a rigorous evaluation, under challenging unconstrained settings, of the proposed method’s effectiveness in suppressing the separability of identity-related privacy features.
4.2 Experimental setup
Fig 4 shows the parameters of the network model. A uses the GRL to adjust the network to finally extract the gender features without identity features. B and C can realize their respective inference functions through fine-tuning. The network parameters are shown in Table 1.
4.3 Results and analysis for local fine-tuning method
In this subsection, we analyze the experimental results of the local fine-tuning method and verify the effectiveness and necessity of the method.
According to Fig 2, several training samples are randomly selected and randomly scrambled in ①. Then perform process ② (gender recognition).
After ②, the gender loss function finally approaches 4, and the gender accuracy can reach 90%. Then perform identity recognition in ④. Then it is found that the identity accuracy cannot reach below 10% but approaches 60%, indicating features extracted by network A still contains a large number of identity privacy features. Therefore, it is necessary to apply the GRL to adjust network A to eliminate the identity features; at the same time, it is also necessary to ensure that the features extracted by network A should also contain the gender features.
Therefore, ⑤ is still needed. After network A is adjusted through the GRL, the identity accuracy is maintained below 4%.
Next, verify whether the main task of gender recognition can be achieved in ②. If it can be achieved, the process ends. If the characteristics of gender are also weakened by the GRL, then we need to carry out ⑥ to further extracting gender features. After several rounds of training, the final gender recognition graphs are shown in Fig 5.
Then verify the accuracy for identity recognition, as shown in Fig 6.
It can be found that the final identity accuracy curve approaches about 3.15%. It means that the identity privacy features are protected, and the whole training process is over.
As last we check the final test results. The the confusion matrix of identity recognition is shown in Fig 7 with the test accuracy for identity 3.19%, indicating that identity privacy protection has achieved good results.
4.4 Results and analysis for global fine-tuning method
In this subsection, we analyze the experimental results of the global fine-tuning method and verify the effectiveness and necessity of the method.
According to process ① in Fig 3, a number of samples are randomly selected from the LFW data set. The process ② gender identification is carried out.
Similar to the results of local fine-tuning, the gender accuracy can reach 90%.
Then execute process ④ for identity identification.
It is found the identity accuracy cannot be lower than 10%. Therefore, it is necessary to perform process ⑤. It is found the identity accuracy is maintained at about 6%. Next, it is necessary to verify whether the gender recognition task can be achieved. Go to process ⑥. After two rounds of training, the final gender recognition accuracy is shown in Fig 8.
Then verify the accuracy for identity in Fig 9. It can be found that the final identity recognition accuracy tends to be around 7.5%. The whole training process is over.
As last we performed tests. The confusion matrix of identity recognition is shown in Fig 10.
It is found the final accuracy of the global fine-tuning identity test is 5.10%. Though it is a little higher than the identity accuracy of 3.19% achieved by local fine-tuning, the task of identity privacy protection is ensured.
4.5 Comparison of local fine-tuning method and global fine-tuning method
Based on the experimental analysis, whether by local fine-tuning or global fine-tuning, the final accuracy for gender recognition can reach a high rate and the identity privacy protection task can also be achieved. In terms of privacy protection, local fine-tuning is more effective than global fine-tuning. In terms of the training process and the number of training iterations, the global fine-tuning method is more concise and converges faster than local fine-tuning method.
5 Method performance analysis
5.1 Computational efficiency analysis
To further substantiate the lightweight advantage of the proposed method in edge computing scenarios, we conduct a theoretical analysis of the computational complexity of each deployment module based on the model architecture. Specifically, we quantify the number of multiply–accumulate operations (MACs), floating-point operations (FLOPs), and model parameters. The MACs for convolutional layers and fully connected layers are defined as follows:
(7)(8)The approximate conversion formula for the number of floating-point operations (FLOPs) is given as:
(9)Within the proposed “end-to-edge” distributed framework, the computational workload of online inference is effectively partitioned. Node A, deployed on the user device (i.e., the privacy-preserving feature extraction module, CNN Unit 1), incurs an extremely low computational cost, amounting to only 21.78 M MACs (approximately 43.56 MFLOPs) with merely 456 parameters. In contrast, Node B, deployed on the edge or cloud server (i.e., the gender inference module, CNN Unit 2), requires about 32.37 M MACs (approximately 64.74 MFLOPs) and contains approximately 5.41 M parameters.
Accordingly, the total computational cost of online inference after deployment is approximately 108.30 MFLOPs. Notably, the user-side device is only required to execute Node A, which imposes virtually no additional computational or storage overhead.Although the above estimates are derived from the specific network scale considered in this work, theoretical analysis indicates that conventional cryptography-based privacy-preserving methods typically require ciphertext transformations on the client side, thereby introducing substantial encryption/decryption or secure computation overhead. In contrast, the proposed method leverages adversarial training to suppress private information directly at the feature extraction stage. As a result, no additional cryptographic burden is incurred during real-time inference in practical deployment.From a system-level deployment perspective, the proposed approach achieves an orders-of-magnitude reduction in computational cost, theoretically demonstrating its strong suitability for resource-constrained edge intelligence applications.
5.2 Comparison of privacy protection paradigms
Naive approaches attempt to preserve privacy by degrading visual information (e.g., reducing resolution or introducing perturbations). Although such methods are simple to deploy, studies on robustness using the LFW dataset indicate that low-quality perturbations—such as blur and noise—lead to a performance drop in primary tasks, including gender recognition. Fundamentally, these approaches reduce identity cues by indiscriminately corrupting visual information, making it difficult to achieve a fine-grained privacy–utility trade-off.
Standard encryption-based schemes [16] [17] provide privacy protection by computing encrypted feature vectors. These approaches primarily ensure data confidentiality during transmission and processing, but are typically accompanied by substantial computational complexity for encryption and decryption, as well as increased communication overhead.
In contrast, the proposed method achieves feature-level decoupling. Compared with standard image compression such as down-sampling (a naive privacy-preserving approach) and conventional lightweight encryption schemes, we systematically delineate the fundamental differences among these paradigms in terms of gender task retention, identity feature suppression, inference computational cost, communication overhead, and applicable scenarios. A detailed comparison is summarized in Table 2.
6 Conclusion
Since the previous deep learning-based methods are not applicable in privacy protection applications where private information and main task information overlap, we propose two privacy protection methods for face edge intelligence applications. They extract the main task features while eliminating the privacy features as much as possible through sharing network layers and two adversarial tuning methods, i.e., local fine-tuning and global fine-tuning of the network. The proposed privacy protection strategy is convenient for distributed layout, such as edge intelligence environment, performing privacy blanking processing on user terminals and safe edge servers relatively close to users, avoiding transmitting private information to remote servers. In the future, we plan to pursue microarchitecture optimization on edge computing hardware and empirical end-side performance measurements (e.g., inference latency and FPS) as the next direction of our work.
References
- 1.
Li H, Shou G, Hu Y, Guo Z. Mobile Edge Computing: Progress and Challenges. In: 2016 4th IEEE MobileCloud Conf. 2016. p. 83–4.
- 2.
Ma L, Pei Q, Xiao H, Li H, Li Z, Fan K. Edge Computing Enhanced Privacy Preserving for Location Based Services. In: IEEE INFOCOM 2019 WKSHPS. 2019. p. 1–6.
- 3.
Cheung S-cS, Rafique MU, Tan W-t. Privacy-Preserving Distributed Deep Learning with Privacy Transformations. In: 2018 IEEE WIFS, Hong Kong, China. 2018. p. 1–7.
- 4.
Chen Z, Li L, Peng H, Liu Y, Zhu H, Yang Y. An evaluation method of image scrambling degree based on pixel distribution. In: 2018 SocialSec, Santa Clara, CA, USA. 2018. 206–11.
- 5. Zhang W, Wang H, Hou D, Yu N. Reversible Data Hiding in Encrypted Images by Reversible Image Transformation. IEEE Trans Multimedia. 2016;18(8):1469–79.
- 6.
Ganin Y, Lempitsky V. Unsupervised domain adaptation by backpropagation. In: ICML 2015 Proceedings Vol. 37. 2015. p. 1180–9.
- 7. Cui L, Chen Z, Wang A, Hu J, Park BB. Development of a Robust Cooperative Adaptive Cruise Control With Dynamic Topology. IEEE Trans Intell Transport Syst. 2022;23(5):4279–90.
- 8. Chen Z, Park BB. Cooperative Adaptive Cruise Control With Unconnected Vehicle in the Loop. IEEE Trans Intell Transport Syst. 2022;23(5):4176–86.
- 9.
Chen X, Zhang N. Layer-wise adversarial training approach to improve adversarial robustness. In: 2020 IJCNN, Glasgow, UK. 2020. p. 1–8.
- 10.
Javanmardi M, Tasdizen T. Domain adaptation for biomedical image segmentation using adversarial training. In: ISBI 2018, Washington (DC). 2018. p. 554–8.
- 11. Wang Y, Poor HV. Decentralized Stochastic Optimization With Inherent Privacy Protection. IEEE Trans Automat Control. 2023;68(4):2293–308.
- 12.
Takano S, Nakao A, Yamagchi S, Oguchi M. Secure Feedback to Edge Servers in Distributed Machine Learning Using Rich Clients. In: COMPSAC 2023, Torino, Italy. 2023. p. 934–9.
- 13. Lv Z, Cheng C, Lv H. Blockchain-Based Decentralized Learning for Security in Digital Twins. IEEE IoT J. 2023;10(24):21479–88.
- 14. Wang T, Zhang Y, Zhao R, Wen W, Lan R. Identifiable Face Privacy Protection via Virtual Identity Transformation. IEEE Signal Process Lett. 2023;30:773–7.
- 15. Zhong Y, Deng W. OPOM: Customized Invisible Cloak Towards Face Privacy Protection. IEEE Trans Pattern Anal Mach Intell. 2023;45(3):3590–603. pmid:35588415
- 16. Wang X, Xue H, Liu X, Pei Q. A Privacy-Preserving Edge Computation-Based Face Verification System for User Authentication. IEEE Access. 2019;7:14186–97.
- 17.
Cheng Y, Meng H, Lei Y, Tan X. Research on Privacy Protection Technology in Face Identity Authentication System Based on Edge Computing. In: AIID 2021, Guangzhou, China. 2021. p. 438–49.
- 18.
McMahan HB, et al. Communication-efficient learning of deep networks from decentralized data. In: AISTATS. 2016.
- 19. Yang W, Xiang W, Yang Y, Cheng P. Optimizing Federated Learning With Deep Reinforcement Learning for Digital Twin Empowered Industrial IoT. IEEE Trans Ind Inf. 2023;19(2):1884–93.
- 20. Liu G, Wang C, Ma X, Yang Y. Keep your data locally: Federated-learning-based data privacy preservation in edge computing. IEEE Network. 2021;35(2):60–6.
- 21. Lu Y, Huang X, Dai Y, Maharjan S, Zhang Y. Blockchain and Federated Learning for Privacy-Preserved Data Sharing in Industrial IoT. IEEE Trans Ind Inf. 2020;16(6):4177–86.
- 22. Li S, Huang J, Hu J, Cheng B. QoE-DEER: A QoE-Aware Decentralized Resource Allocation Scheme for Edge Computing. IEEE Trans Cogn Commun Netw. 2022;8(2):1059–73.
- 23.
Ma X, Zhou A, Zhang S, Wang S. Cooperative Service Caching and Workload Scheduling in Mobile Edge Computing. In: INFOCOM 2020, Toronto, ON, Canada. 2020. p. 2076–85.
- 24. Galanopoulos A, Salonidis T, Iosifidis G. Cooperative Edge Computing of Data Analytics for the Internet of Things. IEEE Trans Cogn Commun Netw. 2020;6(4):1166–79.
- 25. Akbari-Nodehi H, Maddah-Ali MA. Secure Coded Multi-Party Computation for Massive Matrix Operations. IEEE Trans Inform Theory. 2021;67(4):2379–98.
- 26. Iwamura K, Kamal AAAM. Communication-efficient secure computation of encrypted inputs using (k, n) threshold secret sharing. IEEE Access. 2023;11:51166–84.
- 27. Kamal AAAM, Iwamura K. (Server-Aided) Two-Party Multiplication of Encrypted Shares Using (k, n) Threshold Secret Sharing With N ≥ k Servers. IEEE Access. 2021;9:113117–29.
- 28. Ah-Fat P, Huth M. Optimal Accuracy-Privacy Trade-Off for Secure Computations. IEEE Trans Inform Theory. 2019;65(5):3165–82.
- 29. Ma Y, Wu L, Gu X, He J, Yang Z. A Secure Face-Verification Scheme Based on Homomorphic Encryption and Deep Neural Networks. IEEE Access. 2017;5:16532–8.
- 30.
Cai Y, Tang C. Securely outsourced face recognition under federated cloud environment. In: ISPDC 2016, Fuzhou, China. 2016. p. 269–76.
- 31. Haghighat M o h a m m a d, Zonouz S a m a n, Abdel-Mottaleb M o h a m e d. CloudID: Trustworthy cloud-based and cross-enterprise biometric identification. Expert Syst Appl. 2015;42:7905–16.
- 32. Li C, Zhang Y, Luo Y. A federated learning-based edge caching approach for mobile edge computing-enabled intelligent connected vehicles. IEEE T-ITS. 2023;24(3):3360–9.
- 33. Huang Z, Zhang M, Zhang Y. Toward Efficient Encrypted Image Retrieval in Cloud Environment. IEEE Access. 2019;7:174541–50.
- 34. Xia Z, Jiang L, Liu D, Lu L, Jeon B. BOEW: A Content-Based Image Retrieval Scheme Using Bag-of-Encrypted-Words in Cloud Computing. IEEE Trans Serv Comput. 2022;15(1):202–14.
- 35.
Malhotra K, Kumar Y. Challenges to implement Machine Learning in Embedded Systems. In: ICACCCN 2020, Greater Noida, India. 2020. p. 477–81.
- 36. Tari Z, Yi X, Premarathne US, Bertok P, Khalil I. Security and Privacy in Cloud Computing: Vision, Trends, and Challenges. IEEE Cloud Comput. 2015;2(2):30–8.
- 37.
Shah M. Paillier Cryptosystem based Reversible Data Hiding with Extended Application in Ciphertext Domain. Phd thesis. University of Science and Technology of China; 2020.

관련 뉴스