Liping Yi
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Liping Yi has a total of 65 co-authors. Here are the top co-authors by number of shared publications:
- Gang Wang
Nankai University
19 shared publications - Xiaoguang Liu
Nankai University
17 shared publications - Han Yu
Nanyang technological university, Singapore, Singapore
14 shared publications - Xiaoxiao Li
The University of British Columbia
7 shared publications Gang Wang
Nankai University
4 shared publicationsXiaoguang Liu
Nankai University
4 shared publications
Latest Publications
- Open Access
- 28 May 2026
- Published by Association for Computing Machinery (ACM)
- p. 541-542
FL@FM-TheWebConf'26: The 3rd International Workshop on Federated Foundation Models for the Web
- 28 May 2026
- Published by Association for Computing Machinery (ACM)
- p. 541-542
Open AccessFL@FM-TheWebConf'26: The 3rd International Workshop on Federated Foundation Models for the Web
- 30 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 75  (8) , 2811-2825
An Efficient Data Management Based on Adaptive Data Model for High-Cardinality Time-Series Database
With recent advancements in IoT and performance monitoring, time-series management systems have garnered significant attention. Time-series data consists of two components: a series key (identifier) with a set of corresponding data points. Prior efforts only focus on optimizing the storage of data points. However, as the scale of data grows rapidly, storing series keys and their indexes with 60-minute granularity consumes more than 67.9% of the cost incurred by storing the whole dataset, also leading to unacceptably high query latency. Furthermore, the growing cardinality of series keys also slows down the construction of real-time indexes. Previous methods often rely on a data model or user-selected hot data mechanism for indexing. However, they either assume that users are aware of query patterns or suffer from high storage overhead, time-consuming index construction and high query latency. To pursue low storage costs, rapid real-time index construction and efficient query processing simultaneously, we propose a novel Adaptive Data Model (AdpDM) for large-scale high-cardinality time-series data management. It involves four core modules: 1) a grouping algorithm based on the similarity of series keys is designed to group time-series data for real-time group-level index construction while a grouping strategy reuse mechanism is developed to improve insertion throughput. And dynamic adjustment scheme for merging tagv could provide a stable real-time query latency. 2) a group-level inverted index instead of the traditional series-level inverted index is constructed to reduce the costs of storing indexes and query latency. 3) a dictionary compression algorithm compresses series keys and corresponding indexes to reduce storage caused by growing data cardinality. 4) a query merging schema and dynamic cache mechanism to reduce memory usage and unnecessary I/O. We conduct extensive experiments to evaluate AdpDM and advanced time-series database, the results present that AdpDM reduces at least 32% storage cost of series keys, 89% storage cost of indexes, and 24.8-94.1% query latency while achieving about 2.8× higher writing throughput.
- 30 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 75  (8) , 2811-2825
An Efficient Data Management Based on Adaptive Data Model for High-Cardinality Time-Series Database
With recent advancements in IoT and performance monitoring, time-series management systems have garnered significant attention. Time-series data consists of two components: a series key (identifier) with a set of corresponding data points. Prior efforts only focus on optimizing the storage of data points. However, as the scale of data grows rapidly, storing series keys and their indexes with 60-minute granularity consumes more than 67.9% of the cost incurred by storing the whole dataset, also leading to unacceptably high query latency. Furthermore, the growing cardinality of series keys also slows down the construction of real-time indexes. Previous methods often rely on a data model or user-selected hot data mechanism for indexing. However, they either assume that users are aware of query patterns or suffer from high storage overhead, time-consuming index construction and high query latency. To pursue low storage costs, rapid real-time index construction and efficient query processing simultaneously, we propose a novel Adaptive Data Model (AdpDM) for large-scale high-cardinality time-series data management. It involves four core modules: 1) a grouping algorithm based on the similarity of series keys is designed to group time-series data for real-time group-level index construction while a grouping strategy reuse mechanism is developed to improve insertion throughput. And dynamic adjustment scheme for merging tagv could provide a stable real-time query latency. 2) a group-level inverted index instead of the traditional series-level inverted index is constructed to reduce the costs of storing indexes and query latency. 3) a dictionary compression algorithm compresses series keys and corresponding indexes to reduce storage caused by growing data cardinality. 4) a query merging schema and dynamic cache mechanism to reduce memory usage and unnecessary I/O. We conduct extensive experiments to evaluate AdpDM and advanced time-series database, the results present that AdpDM reduces at least 32% storage cost of series keys, 89% storage cost of indexes, and 24.8-94.1% query latency while achieving about 2.8× higher writing throughput.
- 17 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. PP  (15361233) , 1-14
pFedLoRA: Model-Heterogeneous Personalized Federated Learning with Homogeneous Low-Rank Adapter Sharing on Mobile Edge Devices
Federated learning (FL) is an emerging machine learning paradigm in which a central server coordinates multiple participants (FL clients) collaboratively to train on decentralized data. In practice, FL often faces data, system, and model heterogeneity, which inspires the field of Model-Heterogeneous Personalized Federated Learning (MHPFL). However, existing MHPFL methods rely on extra public data or ignore the relationship between local private heterogeneous models and shared homogeneous models across clients. This leads to unsatisfactory model performance, computational overheads, and communication costs. To bridge this gap, we propose a novel and efficient model-heterogeneous personalized Federated learning framework (pFedLoRA) based on sharing homogeneous Low-Rank Adapter (LoRA) which is popular for fine-tuning pre-trained models. Specifically, we devise a lightweight homogeneous adapter, rather than apply the typical LoRA, to facilitate each client's heterogeneous local model training with our proposed iterative training for global-local bidirectional knowledge exchange. The homogeneous small local adapters are aggregated on the FL server to generate a global adapter. We theoretically prove its $\mathcal {O}(1/T)$ non-convex convergence rate. Experiments on 5 datasets demonstrate pFedLoRA outperforms 9 state-of-the-art baselines in model accuracy with $11.81 \times$ computation and $7.41\times$ communication cost saving.
- 17 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. PP  (15361233) , 1-14
pFedLoRA: Model-Heterogeneous Personalized Federated Learning with Homogeneous Low-Rank Adapter Sharing on Mobile Edge Devices
Federated learning (FL) is an emerging machine learning paradigm in which a central server coordinates multiple participants (FL clients) collaboratively to train on decentralized data. In practice, FL often faces data, system, and model heterogeneity, which inspires the field of Model-Heterogeneous Personalized Federated Learning (MHPFL). However, existing MHPFL methods rely on extra public data or ignore the relationship between local private heterogeneous models and shared homogeneous models across clients. This leads to unsatisfactory model performance, computational overheads, and communication costs. To bridge this gap, we propose a novel and efficient model-heterogeneous personalized Federated learning framework (pFedLoRA) based on sharing homogeneous Low-Rank Adapter (LoRA) which is popular for fine-tuning pre-trained models. Specifically, we devise a lightweight homogeneous adapter, rather than apply the typical LoRA, to facilitate each client's heterogeneous local model training with our proposed iterative training for global-local bidirectional knowledge exchange. The homogeneous small local adapters are aggregated on the FL server to generate a global adapter. We theoretically prove its $\mathcal {O}(1/T)$ non-convex convergence rate. Experiments on 5 datasets demonstrate pFedLoRA outperforms 9 state-of-the-art baselines in model accuracy with $11.81 \times$ computation and $7.41\times$ communication cost saving.
- 6 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. 36  (6) , 9010-9022
FedRSL: Representation Subspace Learning in Model-Heterogeneous Federated Learning
Model-heterogeneous federated learning (MHFL), supporting FL collaboration across clients with heterogeneous models, has become a more practical FL paradigm. Existing MHFL methods enable knowledge fusion over heterogeneous client models by sharing partial homogeneous parameters or extracted label-wise average representations, suffering from model performance bottlenecks and privacy leakage risks. To bridge this gap, we propose a novel model-heterogeneous Federated learning method with homogeneous Representation Subspace Learning (FedRSL) instead of sharing model parameters or representations. In FedRSL, each client’s local heterogeneous model comprises a feature extractor and a prediction header. (1) We construct a homogeneous representation subspace for each client to learn local representation knowledge, and the server aggregates them to generate the global representation subspace for representation knowledge fusion. (2) To facilitate representation learning capability while maintaining efficient communication and computation, we design a lightweight linear model as the homogeneous low-rank linear representation subspace. For each local data sample, its personalized representation extracted by the feature extractor is processed by the global representation subspace to produce the corresponding generalized representation. (3) To effectively bi-transfer global generalized and local personalized knowledge, we reduce the distance between the local personalized representation and the corresponding generalized representation. Experiments on 3 computer vision and 1 natural language processing benchmark datasets over 6 baselines demonstrate that FedRSL obtains state-of-the-art model accuracy (up to 5.51% accuracy improvement) while consuming low communication and computation overheads.
- 6 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. 36  (6) , 9010-9022
FedRSL: Representation Subspace Learning in Model-Heterogeneous Federated Learning
Model-heterogeneous federated learning (MHFL), supporting FL collaboration across clients with heterogeneous models, has become a more practical FL paradigm. Existing MHFL methods enable knowledge fusion over heterogeneous client models by sharing partial homogeneous parameters or extracted label-wise average representations, suffering from model performance bottlenecks and privacy leakage risks. To bridge this gap, we propose a novel model-heterogeneous Federated learning method with homogeneous Representation Subspace Learning (FedRSL) instead of sharing model parameters or representations. In FedRSL, each client’s local heterogeneous model comprises a feature extractor and a prediction header. (1) We construct a homogeneous representation subspace for each client to learn local representation knowledge, and the server aggregates them to generate the global representation subspace for representation knowledge fusion. (2) To facilitate representation learning capability while maintaining efficient communication and computation, we design a lightweight linear model as the homogeneous low-rank linear representation subspace. For each local data sample, its personalized representation extracted by the feature extractor is processed by the global representation subspace to produce the corresponding generalized representation. (3) To effectively bi-transfer global generalized and local personalized knowledge, we reduce the distance between the local personalized representation and the corresponding generalized representation. Experiments on 3 computer vision and 1 natural language processing benchmark datasets over 6 baselines demonstrate that FedRSL obtains state-of-the-art model accuracy (up to 5.51% accuracy improvement) while consuming low communication and computation overheads.
- Top Cited Papers
- 20 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Knowledge and Data Engineering
- Vol. 38  (3) , 1905-1918
pFedMoE: Data-Level Personalization With Mixture of Experts in Model-Heterogeneous Personalized Federated Learning
With growing client diversity, model-heterogeneous personalized federated learning (MHPFL) supports collaboration over structure-heterogeneous client models. However, existing MHPFL methods only achieve client-level personalization but ignore inherent discrepancies within each client’s different data samples, leading to limited model performance. To this end, we propose a novel model-heterogeneous personalized Federated learning with Mixture of Experts (pFedMoE) to achieve a fine-grained data-level personalization. As the first work that incorporates MoE in MHPFL, it introduces three innovations: (1) Different clients hold heterogeneous local models, we add a small proxy global homogeneous feature extractor shared by clients for knowledge exchange. (2) To achieve a fine-grained data-level personalization, we construct a personalized local MoE for each client: a local expert (local heterogeneous client model’s feature extractor), a global expert (global proxy homogeneous feature extractor), and a local personalized gating network, which dynamically balances the generalization and personalization of the local model at the data sample level. (3) We customize a lightweight linear gating network to capture the generalized and personalized data characteristics of each local data sample. We theoretically prove its$\mathcal {O}(1/T)$ convergence rate. Experiments on 3 benchmark image datasets, 1 real-world image dataset and 1 real-world text dataset against 9 baselines demonstrate its state-of-the-art model accuracy with up to 2.79% accuracy improvement while saving up to 43.12% computational overheads and keeping satisfactory communication costs.- 20 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Knowledge and Data Engineering
- Vol. 38  (3) , 1905-1918
Top Cited PaperspFedMoE: Data-Level Personalization With Mixture of Experts in Model-Heterogeneous Personalized Federated Learning
With growing client diversity, model-heterogeneous personalized federated learning (MHPFL) supports collaboration over structure-heterogeneous client models. However, existing MHPFL methods only achieve client-level personalization but ignore inherent discrepancies within each client’s different data samples, leading to limited model performance. To this end, we propose a novel model-heterogeneous personalized Federated learning with Mixture of Experts (pFedMoE) to achieve a fine-grained data-level personalization. As the first work that incorporates MoE in MHPFL, it introduces three innovations: (1) Different clients hold heterogeneous local models, we add a small proxy global homogeneous feature extractor shared by clients for knowledge exchange. (2) To achieve a fine-grained data-level personalization, we construct a personalized local MoE for each client: a local expert (local heterogeneous client model’s feature extractor), a global expert (global proxy homogeneous feature extractor), and a local personalized gating network, which dynamically balances the generalization and personalization of the local model at the data sample level. (3) We customize a lightweight linear gating network to capture the generalized and personalized data characteristics of each local data sample. We theoretically prove its$\mathcal {O}(1/T)$ convergence rate. Experiments on 3 benchmark image datasets, 1 real-world image dataset and 1 real-world text dataset against 9 baselines demonstrate its state-of-the-art model accuracy with up to 2.79% accuracy improvement while saving up to 43.12% computational overheads and keeping satisfactory communication costs. - 3 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13  (3) , 5018-5031
CSQoS: Continual Sparse QoS Measurement for Edge Clouds With GNN-Based Variational Bayesian
Edge computing, an emerging paradigm, utilizes decentralized edge nodes to offer low-latency, high-quality network services. Quality of service (QoS) is a crucial metric to measure the network service quality, and network resource scheduling relies on QoS measurement results. However, current QoS measurement methods often measure all QoS data among edge nodes; these dense measurement approaches introduce significant costs. In addition, edge nodes may adopt varied network access manners; these factors cause fluctuations in QoS between edge nodes. However, existing QoS measurement works often focus on measuring exact QoS values while merely considering QoS fluctuations, resulting in unreliable measured QoS data. In addition, existing QoS measurement methods often cannot support online QoS measurement, leading to stale offline QoS data affecting network resource scheduling. To tackle the two issues, we propose a novel continual sparse QoS (CSQoS) range measurement method with four innovative designs: First, to reduce measurement costs, we propose to measure QoS by only sampling partial QoS data and using them to impute unmeasured QoS data. To achieve sparse QoS imputation, we propose a novel variational Bayesian model (BayGNN) with an edge-enhanced graph neural network (GNN) as the encoder for feature extraction and a multilayer perceptron (MLP) as the decoder to predict unmeasured QoS data. Second, to assess QoS data ranges, we design the proposed BayGNN model to produce uncertainty simultaneously. Third, to fulfill reliable online QoS predictions, we incorporate continual learning and residual connections in BayGNN. Experimental results on two real-world datasets demonstrate that CSQoS has minimal QoS imputation error with the lowest measurement costs, reducing root mean square error (RMSE) of 17.6% and sampling costs of 20%.
- 3 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13  (3) , 5018-5031
CSQoS: Continual Sparse QoS Measurement for Edge Clouds With GNN-Based Variational Bayesian
Edge computing, an emerging paradigm, utilizes decentralized edge nodes to offer low-latency, high-quality network services. Quality of service (QoS) is a crucial metric to measure the network service quality, and network resource scheduling relies on QoS measurement results. However, current QoS measurement methods often measure all QoS data among edge nodes; these dense measurement approaches introduce significant costs. In addition, edge nodes may adopt varied network access manners; these factors cause fluctuations in QoS between edge nodes. However, existing QoS measurement works often focus on measuring exact QoS values while merely considering QoS fluctuations, resulting in unreliable measured QoS data. In addition, existing QoS measurement methods often cannot support online QoS measurement, leading to stale offline QoS data affecting network resource scheduling. To tackle the two issues, we propose a novel continual sparse QoS (CSQoS) range measurement method with four innovative designs: First, to reduce measurement costs, we propose to measure QoS by only sampling partial QoS data and using them to impute unmeasured QoS data. To achieve sparse QoS imputation, we propose a novel variational Bayesian model (BayGNN) with an edge-enhanced graph neural network (GNN) as the encoder for feature extraction and a multilayer perceptron (MLP) as the decoder to predict unmeasured QoS data. Second, to assess QoS data ranges, we design the proposed BayGNN model to produce uncertainty simultaneously. Third, to fulfill reliable online QoS predictions, we incorporate continual learning and residual connections in BayGNN. Experimental results on two real-world datasets demonstrate that CSQoS has minimal QoS imputation error with the lowest measurement costs, reducing root mean square error (RMSE) of 17.6% and sampling costs of 20%.
- Open Access
- 3 August 2025
- Published by Association for Computing Machinery (ACM)
- p. 2725-2734
PMKLC: Parallel Multi-Knowledge Learning-based Lossless Compression for Large-Scale Genomics Database
- 3 August 2025
- Published by Association for Computing Machinery (ACM)
- p. 2725-2734
Open AccessPMKLC: Parallel Multi-Knowledge Learning-based Lossless Compression for Large-Scale Genomics Database
- Open Access
- 23 July 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
SampDedup: Sampling Prediction for Efficient Inline Data Deduplication on Non-volatile Memory
Data deduplication is an effective technique for reducing redundant data storage space in various storage systems. Generally, deduplication consists of four steps: chunking, fingerprinting, fingerprint lookup, and data management. Recently, Non-volatile Memory (NVM) as an emerging storage device has received widespread attention. Directly applying the deduplication technique on NVM for storage cost savings faces many challenges: a) deduplication on NVM devices suffers from computation bottleneck instead of the I/O bottleneck faced by deduplication on traditional storage devices (such as HDD and SSD); b) new fingerprint indexes and metadata are required to be re-designed to adapt to NVM characteristics; c) inline deduplication on NVM is more sensitive to the latency. To solve these challenges, we propose a novel Samp ling prediction-based inline data Dedup lication method ( SampDedup ) on NVM devices. It aims to ensure high deduplication ratios while reducing computation costs and latency by optimizing data chunking , fingerprinting , and fingerprint lookup . a) For data chunking , a sampling prediction-based chunking method ( SampChunk ) is proposed to leverage chunk similarity to distinguish duplicate chunks and skip them for chunking. This method can be easily integrated into most sliding-window based and non-window based CDC chunking algorithms. b) For fingerprinting , the commonly used SHA-1 algorithm is further optimized to reduce the extra computational overhead introduced by SampChunk, and an asynchronous fingerprinting method is proposed to reduce the fingerprinting latency of unique chunks. c) For fingerprint lookup , we design a header fingerprint index and metadata table for each data chunk constructed by SampChunk on NVM, and we use a fast-read buffer to replace the traditional slow LRU cache to improve search efficiency. Experiments on 4 real-world datasets demonstrate that SampDedup consistently presents high inline data deduplication ratios on NVM with different workloads and data partitioning algorithms while saving more than \(90\% \) chunking time compared with state-of-the-art deduplication baselines.- 23 July 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Architecture and Code Optimization
Open AccessSampDedup: Sampling Prediction for Efficient Inline Data Deduplication on Non-volatile Memory
Data deduplication is an effective technique for reducing redundant data storage space in various storage systems. Generally, deduplication consists of four steps: chunking, fingerprinting, fingerprint lookup, and data management. Recently, Non-volatile Memory (NVM) as an emerging storage device has received widespread attention. Directly applying the deduplication technique on NVM for storage cost savings faces many challenges: a) deduplication on NVM devices suffers from computation bottleneck instead of the I/O bottleneck faced by deduplication on traditional storage devices (such as HDD and SSD); b) new fingerprint indexes and metadata are required to be re-designed to adapt to NVM characteristics; c) inline deduplication on NVM is more sensitive to the latency. To solve these challenges, we propose a novel Samp ling prediction-based inline data Dedup lication method ( SampDedup ) on NVM devices. It aims to ensure high deduplication ratios while reducing computation costs and latency by optimizing data chunking , fingerprinting , and fingerprint lookup . a) For data chunking , a sampling prediction-based chunking method ( SampChunk ) is proposed to leverage chunk similarity to distinguish duplicate chunks and skip them for chunking. This method can be easily integrated into most sliding-window based and non-window based CDC chunking algorithms. b) For fingerprinting , the commonly used SHA-1 algorithm is further optimized to reduce the extra computational overhead introduced by SampChunk, and an asynchronous fingerprinting method is proposed to reduce the fingerprinting latency of unique chunks. c) For fingerprint lookup , we design a header fingerprint index and metadata table for each data chunk constructed by SampChunk on NVM, and we use a fast-read buffer to replace the traditional slow LRU cache to improve search efficiency. Experiments on 4 real-world datasets demonstrate that SampDedup consistently presents high inline data deduplication ratios on NVM with different workloads and data partitioning algorithms while saving more than \(90\% \) chunking time compared with state-of-the-art deduplication baselines. - Open Access
- 22 April 2025
- Published by Association for Computing Machinery (ACM)
- p. 3543-3551
MSDZip: Universal Lossless Compression for Multi-source Data via Stepwise-parallel and Learning-based Prediction
- 22 April 2025
- Published by Association for Computing Machinery (ACM)
- p. 3543-3551
Open AccessMSDZip: Universal Lossless Compression for Multi-source Data via Stepwise-parallel and Learning-based Prediction
- Open Access
- 11 April 2025
- Published by Association for the Advancement of Artificial Intelligence (AAAI) in Proceedings of the AAAI Conference on Artificial Intelligence
- Vol. 39  (21) , 22146-22154
pFedES: Generalized Proxy Feature Extractor Sharing for Model Heterogeneous Personalized Federated Learning
Federated learning (FL), as a privacy-preserving collaborative machine learning paradigm, has attracted significant interest from industry and academia. To allow each data owner (FL client) to train a heterogeneous and personalized local model based on its local data distribution, system resources and requirements on model structure, the field of model-heterogeneous personalized federated learning (MHPFL) has emerged. Existing MHPFL approaches either rely on the availability of a public dataset with special characteristics to facilitate knowledge transfer, incur high computational and communication costs, or face potential model leakage risks. To address these limitations, we propose a model-heterogeneous personalized Federated learning approach based on generalized proxy feature Extractor Sharing (pFedES) for supervised image classification tasks. (1) We devise a shared small proxy homogeneous feature extractor before each client's heterogeneous local model. (2) Clients train them via the proposed iterative learning to enable the exchange of global generalized knowledge and local personalized knowledge. (3) The small proxy local homogeneous extractors produced after local training are uploaded to the server for aggregation to facilitate knowledge fusion across clients. We theoretically prove pFedES converges with a non-convex convergence rate O(1/T). Experiments on 3 benchmark datasets against 9 baselines demonstrate that pFedES performs state-of-the-art model accuracy while maintaining efficient communication and computation.
- 11 April 2025
- Published by Association for the Advancement of Artificial Intelligence (AAAI) in Proceedings of the AAAI Conference on Artificial Intelligence
- Vol. 39  (21) , 22146-22154
Open AccesspFedES: Generalized Proxy Feature Extractor Sharing for Model Heterogeneous Personalized Federated Learning
Federated learning (FL), as a privacy-preserving collaborative machine learning paradigm, has attracted significant interest from industry and academia. To allow each data owner (FL client) to train a heterogeneous and personalized local model based on its local data distribution, system resources and requirements on model structure, the field of model-heterogeneous personalized federated learning (MHPFL) has emerged. Existing MHPFL approaches either rely on the availability of a public dataset with special characteristics to facilitate knowledge transfer, incur high computational and communication costs, or face potential model leakage risks. To address these limitations, we propose a model-heterogeneous personalized Federated learning approach based on generalized proxy feature Extractor Sharing (pFedES) for supervised image classification tasks. (1) We devise a shared small proxy homogeneous feature extractor before each client's heterogeneous local model. (2) Clients train them via the proposed iterative learning to enable the exchange of global generalized knowledge and local personalized knowledge. (3) The small proxy local homogeneous extractors produced after local training are uploaded to the server for aggregation to facilitate knowledge fusion across clients. We theoretically prove pFedES converges with a non-convex convergence rate O(1/T). Experiments on 3 benchmark datasets against 9 baselines demonstrate that pFedES performs state-of-the-art model accuracy while maintaining efficient communication and computation.