Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Zhiyu Chen has a total of 19 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Class-Anchor-Based Federated Continual Learning for Vehicle Part Recognition

    by Fengyu Huang, Jianwei Guo, Gang Liu, Zhiyu Chen

    Federated learning provides a feasible way to support collaborative vehicle part recognition without sharing raw data across distributed vehicle service nodes. However, real vehicle scenarios exhibit spatiotemporal heterogeneity, where client distributions vary and new vehicle models, parts, and viewpoints emerge over time. Such heterogeneity causes catastrophic forgetting and degrades knowledge retention in federated continual learning. To address this problem, this paper proposes FedACA (Federated Continual Learning with Adaptive Class Anchors), a class-anchor-based federated continual learning method for vehicle part image classification. The method uses a frozen Vision Transformer backbone and introduces scene-aware prompt adaptation, adaptive class anchors, non-parametric global prototype selection, and class-aware prompt fusion. These modules stabilize class semantics, reduce feature drift, and improve cross-client knowledge alignment in a data-local federated setting with raw-data non-sharing; however, FedACA does not provide formal privacy guarantees. Experiments on a self-constructed 67-class multi-view vehicle-part dataset and CIFAR-100 show that, under the task-aware class-masking protocol, FedACA achieves final-stage average accuracies of 97.34% ± 0.72% and 95.71% ± 0.63%, respectively, over three independent random seeds, outperforming the evaluated federated continual learning baselines. These results demonstrate that FedACA improves recognition performance and historical knowledge retention under spatiotemporally heterogeneous federated continual learning settings.

  • Open Access

    Graph Attributed Unlearning via Propagation Suppression and Knowledge Dissipation

    by Zhiyu Chen, Jiaquan Liang, Qi Luo, Zhipeng Cai

    With the growing global emphasis on data privacy protection, particularly the enforcement of the “right to be forgotten” under the GDPR, effectively deleting private information from models has become an urgent challenge. Graph-structured data presents a particularly challenging unlearning scenario due to its non-Euclidean nature and strong relational dependencies, which are prevalent in real-world applications such as social and recommendation systems. To address this issue, graph unlearning has been introduced to eliminate the influence of deleted data on models while preserving their overall performance. The effectiveness of graph unlearning is typically evaluated by three key metrics: model performance, unlearning efficiency, and robustness against membership inference attacks, which together determine the overall quality of an unlearning method. Existing graph unlearning methods fall into exact and approximate regimes. Most studies focus on edge/node-level unlearning, and existing attempts at feature-level unlearning remain limited. Exact unlearning methods that adopt the SISA partition and retraining paradigm may inadvertently reintroduce the features intended to be unlearned during the aggregation phase, thereby leading to incomplete unlearning. Approximate methods, on the other hand, often incur excessive information loss in feature-level removal, which degrades predictive accuracy. Accordingly, we propose a graph unlearning framework specifically designed for feature-level unlearning, consisting of two main stages. In the first stage, we zero out the features of the unlearned nodes at each layer to block their propagation through the GNN, thereby reducing their influence on neighboring node representations. In the second stage, we induce misclassification of the unlearned nodes to progressively degrade model representations and learned knowledge associated with them, enabling more thorough feature-level unlearning. Experiments on multiple graph datasets and models demonstrate that our method achieves favorable overall unlearning performance in most settings, offering a balanced trade-off between accuracy, unlearning efficiency, and unlearning effectiveness.

  • Hypergraph Unlearning: A Size-Based Hyperedge Selection and Coverage Aggregation Approach

    by Jiaquan Liang, Zhiyu Chen, Qi Luo, Zhenzhen Xie, Gang Liu, Zhipeng Cai

    Graph unlearning aims to provably remove some training data from graph neural networks (GNNs) while eliminating their impact on model predictions. Although retraining the GNNs from scratch is a direct and legitimate solution, it entails substantial computational resources. To address this issue, graph unlearning methods have been proposed in the domain of graph data. However, applying existing graph unlearning methods directly to hypergraph data affects model utility. Specifically, the graph unlearning methods severely damage the higher-order structural information of hypergraphs and fail to remove all the information that needs to be unlearned. In this paper, we propose Hyperedge Size-Based Core-Sharing Decomposition, a novel hypergraph unlearning framework tailored to the structural characteristics of hypergraph data. Its contributions include a subgraph partitioning method specific to hypergraphs and an aggregation method based on node coverage. We conduct extensive experiments on seven real-world hypergraph datasets to demonstrate the unlearning efficiency and model utility. Compared to the baseline methods, our approach achieves up to 5% higher accuracy and reduces the average unlearning time by 28%. Furthermore, our node-coverage-based aggregation approach achieves up to 6% higher accuracy.


  • Lane-Change Decision of Automatic Driving Based on Reinforcement Learning Framework

    by Jiaxin Li, Gang Liu, Zhiyu Chen, Jianwei Guo

    In highway traffic environment, vehicle lane-change decisions are a critical and difficult task for an agent. The traditional deep Q-network (DQN) has been the main architecture used for this task. However, DQN still faces two major challenges in solving the lane-changing decision-making problem for autonomous driving: 1) the instability in the training process leads to frequent collisions and 2) the single reward function limits the agent’s ability to learn representative domain knowledge about highways. To address these issues, this paper proposes a multi-reward DQN for highway driving decisions, with long short-term memory (LSTM) and self-attention. We call this model LAMRDQN. It incorporates LSTM and self-attention mechanism into Q-network to reduce frequent collisions. In addition, the unitary reward is changed into multiple reward functions—each reward emphasizing a specific factor, such as speed, obeying traffic rules, and lane changing. The use of multiple rewards enables the model to capture and optimize various aspects of autonomous driving on highway tasks, thereby obtaining more representative domain knowledge. The experiment results show that, in the highway environment, the method can effectively improve the problems of frequent collisions and poor learning performance when vehicles change lanes, and has better overtaking performance than related reinforcement learning algorithms.


  • Open Access

    Semi-Supervised One-Stage Object Detection for Maize Leaf Disease

    by Jiaqi Liu, Yanxin Hu, Qianfu Su, Jianwei Guo, Zhiyu Chen, Gang Liu

    Maize is one of the most important crops globally, and accurate diagnosis of leaf diseases is crucial for ensuring increased yields. Despite the continuous progress in computer vision technology, detecting maize leaf diseases based on deep learning still relies on a large amount of manually labeled data, and the labeling process is time-consuming and labor-intensive. Moreover, the detectors currently used for identifying maize leaf diseases have relatively low accuracy in complex experimental fields. Therefore, the proposed Agronomic Teacher, an object detection algorithm that utilizes limited labeled and abundant unlabeled data, is applied to maize leaf disease recognition. In this work, a semi-supervised object detection framework is built based on a single-stage detector, integrating the Weighted Average Pseudo-labeling Assignment (WAP) strategy and AgroYOLO detector combining Agro-Backbone network with Agro-Neck network. The WAP strategy uses weight adjustments to set objectness and classification scores as evaluation criteria for pseudo-labels reliability assignment. Agro-Backbone network accurately extracts features of maize leaf diseases and obtains richer semantic information. Agro-Neck network enhances feature fusion by utilizing multi-layer features for collaborative combinations. The effectiveness of the proposed method is validated on the MaizeData and PascalVOC datasets at different annotation ratios. Compared to the baseline model, Agronomic Teacher leverages abundant unlabeled data to achieve a 6.5% increase in mAP (0.5) on the 30% labeled MaizeData. On the 30% labeled PascalVOC dataset, the mAP (0.5) improved by 8.2%, demonstrating the method’s potential for generalization.


  • Open Access

    TabFedSL: A Self-Supervised Approach to Labeling Tabular Data in Federated Learning Environments

    by Ruixiao Wang, Yanxin Hu, Zhiyu Chen, Jianwei Guo, Gang Liu

    Currently, self-supervised learning has shown effectiveness in solving data labeling issues. Its success mainly depends on having access to large, high-quality datasets with diverse features. It also relies on utilizing the spatial, temporal, and semantic structures present in the data. However, domains such as finance, healthcare, and insurance primarily utilize tabular data formats. This presents challenges for traditional data augmentation methods aimed at improving data quality. Furthermore, the privacy-sensitive nature of these domains complicates the acquisition of the extensive, high-quality datasets necessary for training effective self-supervised models. To tackle these challenges, our proposal introduces a novel framework that combines self-supervised learning with Federated Learning (FL). This approach aims to solve the problem of data-distributed training while ensuring training quality. Our framework improves upon the conventional self-supervised learning data augmentation paradigm by incorporating data labeling through the segmentation of data into subsets. Our framework adds noise by splitting subsets of data and can achieve the same level of centralized learning in a distributed environment. Moreover, we conduct experiments on various public tabular datasets to evaluate our approach. The experimental results showcase the effectiveness and generalizability of our proposed method in scenarios involving unlabeled data and distributed settings.


  • Open Access

    Lightweight One-Stage Maize Leaf Disease Detection Model with Knowledge Distillation

    by Yanxin Hu, Gang Liu, Zhiyu Chen, Jiaqi Liu, Jianwei Guo

    Maize is one of the world’s most important crops, and maize leaf diseases can have a direct impact on maize yields. Although deep learning-based detection methods have been applied to maize leaf disease detection, it is difficult to guarantee detection accuracy when using a lightweight detection model. Considering the above problems, we propose a lightweight detection algorithm based on improved YOLOv5s. First, the Faster-C3 module is proposed to replace the original CSP module in YOLOv5s, to significantly reduce the number of parameters in the feature extraction process. Second, CoordConv and improved CARAFE are introduced into the neck network, to improve the refinement of location information during feature fusion and to refine richer semantic information in the downsampling process. Finally, the channel-wise knowledge distillation method is used in model training to improve the detection accuracy without increasing the number of model parameters. In a maize leaf disease detection dataset (containing five leaf diseases and a total of 12,957 images), our proposed algorithm had 15.5% less parameters than YOLOv5s, while the mAP(0.5) and mAP(0.5:0.95) were 3.8% and 1.5% higher, respectively. The experiments demonstrated the effectiveness of the method proposed in this study and provided theoretical and technical support for the automated detection of maize leaf diseases.


  • Open Access

    A Mapless Local Path Planning Approach Using Deep Reinforcement Learning Framework

    by Yan Yin, Zhiyu Chen, Gang Liu, Jianwei Guo

    The key module for autonomous mobile robots is path planning and obstacle avoidance. Global path planning based on known maps has been effectively achieved. Local path planning in unknown dynamic environments is still very challenging due to the lack of detailed environmental information and unpredictability. This paper proposes an end-to-end local path planner n-step dueling double DQN with reward-based ϵ-greedy (RND3QN) based on a deep reinforcement learning framework, which acquires environmental data from LiDAR as input and uses a neural network to fit Q-values to output the corresponding discrete actions. The bias is reduced using n-step bootstrapping based on deep Q-network (DQN). The ϵ-greedy exploration-exploitation strategy is improved with the reward value as a measure of exploration, and an auxiliary reward function is introduced to increase the reward distribution of the sparse reward environment. Simulation experiments are conducted on the gazebo to test the algorithm’s effectiveness. The experimental data demonstrate that the average total reward value of RND3QN is higher than that of algorithms such as dueling double DQN (D3QN), and the success rates are increased by 174%, 65%, and 61% over D3QN on three stages, respectively. We experimented on the turtlebot3 waffle pi robot, and the strategies learned from the simulation can be effectively transferred to the real robot.