Li Zhu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Li Zhu has a total of 174 co-authors. Here are the top co-authors by number of shared publications:
- Yaxiong Wang
Hefei University of Technology, Hefei, China
9 shared publications - Fei Guo
School of Software Engineering, Xian Jiaotong University, Xian, China
8 shared publications - Zhedong Zheng
University of Macau, Macau, China
7 shared publications Fei Guo
Xi’an Jiaotong University
6 shared publications- Xueming Qian
Xi’an Jiaotong University
6 shared publications Han Qi
Shanghai Artificial Intelligence Laboratory
5 shared publications
Latest Publications
- Open Access
- 11 November 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (22) , 11157
A Context-Aware ZTNA Framework with eBPF/XDP Policy Enforcement and Machine-Learning-Based Trust Evaluation for Campus Networks
University networks host heterogeneous users, endpoints, access locations, and services, so access decisions cannot rely solely on identity information. This work proposes the Context-Aware Dynamic Trust Zero Trust Network Access (CADT-ZTNA) framework, integrating log-based risk classification with an eBPF/XDP-assisted policy enforcement design. We analyzed 113,964 campus security records collected over 30 days: 94,012 actor logs fed the supervised classifier, while 19,952 vendor logs were set aside for separate analysis. Using a chronological train–validation–test split, XGBoost reached a test ROC-AUC of 0.9808, AP of 0.8740 and F1 of 0.8892. Random Forest delivered marginally better ROC-AUC (0.9834) and AP (0.8887). Ablation results highlight the role of temporal behavioral features: removing the temporal feature group reduced ROC-AUC from 0.9808 to 0.9056 and AP from 0.8740 to 0.1869. Removing only user_time_gap reduced AP to 0.6849 and F1 to 0.7676. As 99.15% of positive samples are dormant-account login events, the reported classification performance reflects discrimination against these predefined operational security labels, rather than detection of independently verified real-world malicious activity.
- 11 November 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (22) , 11157
Open AccessA Context-Aware ZTNA Framework with eBPF/XDP Policy Enforcement and Machine-Learning-Based Trust Evaluation for Campus Networks
University networks host heterogeneous users, endpoints, access locations, and services, so access decisions cannot rely solely on identity information. This work proposes the Context-Aware Dynamic Trust Zero Trust Network Access (CADT-ZTNA) framework, integrating log-based risk classification with an eBPF/XDP-assisted policy enforcement design. We analyzed 113,964 campus security records collected over 30 days: 94,012 actor logs fed the supervised classifier, while 19,952 vendor logs were set aside for separate analysis. Using a chronological train–validation–test split, XGBoost reached a test ROC-AUC of 0.9808, AP of 0.8740 and F1 of 0.8892. Random Forest delivered marginally better ROC-AUC (0.9834) and AP (0.8887). Ablation results highlight the role of temporal behavioral features: removing the temporal feature group reduced ROC-AUC from 0.9808 to 0.9056 and AP from 0.8740 to 0.1869. Removing only user_time_gap reduced AP to 0.6849 and F1 to 0.7676. As 99.15% of positive samples are dormant-account login events, the reported classification performance reflects discrimination against these predefined operational security labels, rather than detection of independently verified real-world malicious activity.
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 17584-17591
Who You Are and Where You Belong: Learning Individual and Group Features for Social Media Video Popularity Prediction
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 17584-17591
Open AccessWho You Are and Where You Belong: Learning Individual and Group Features for Social Media Video Popularity Prediction
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Theory
- Vol. 72 (11) , 8680-8708
Graph Feedback Bandits on Similar Arms: With and Without Graph Structures
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Theory
- Vol. 72 (11) , 8680-8708
Graph Feedback Bandits on Similar Arms: With and Without Graph Structures
- 5 October 2026
- Published by Springer Nature
- p. 348-366
Before Thinking, Learn to Decide: Proactive Routing for Efficient Visual Reasoning
- 5 October 2026
- Published by Springer Nature
- p. 348-366
Before Thinking, Learn to Decide: Proactive Routing for Efficient Visual Reasoning
- 19 July 2026
- Published by Springer Nature
- p. 168-179
DFDN: Domain Feature Disentanglement Network for Cross-Domain Few-Shot Action Recognition
- 19 July 2026
- Published by Springer Nature
- p. 168-179
DFDN: Domain Feature Disentanglement Network for Cross-Domain Few-Shot Action Recognition
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Visualization and Computer Graphics
- Vol. 32 (7) , 5415-5426
Look, Compare and Draw: Differential Query Transformer for Automatic Oil Painting
This work introduces a new approach to automatic oil painting that emphasizes the creation of dynamic and expressive brushstrokes. A pivotal challenge lies in mitigating the duplicate and common-place strokes, which often lead to less aesthetic outcomes. Inspired by the human painting process, i.e., observing, comparing, and drawing, we incorporate differential image analysis into a neural oil painting model, allowing the model to effectively concentrate on the incremental impact of successive brushstrokes. To operationalize this concept, we propose the Differential Query Transformer (DQ-Transformer), a new architecture that leverages differentially derived image representations enriched with positional encoding to guide the stroke prediction process. This integration enables the model to maintain heightened sensitivity to local details, resulting in more refined and nuanced stroke generation. Furthermore, we incorporate adversarial training into our framework, enhancing the accuracy of stroke prediction and thereby improving the overall realism and fidelity of the synthesized paintings. Extensive qualitative evaluations, complemented by a controlled user study, validate that our DQ-Transformer surpasses existing methods in both visual realism and artistic authenticity, typically achieving these results with fewer strokes.
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Visualization and Computer Graphics
- Vol. 32 (7) , 5415-5426
Look, Compare and Draw: Differential Query Transformer for Automatic Oil Painting
This work introduces a new approach to automatic oil painting that emphasizes the creation of dynamic and expressive brushstrokes. A pivotal challenge lies in mitigating the duplicate and common-place strokes, which often lead to less aesthetic outcomes. Inspired by the human painting process, i.e., observing, comparing, and drawing, we incorporate differential image analysis into a neural oil painting model, allowing the model to effectively concentrate on the incremental impact of successive brushstrokes. To operationalize this concept, we propose the Differential Query Transformer (DQ-Transformer), a new architecture that leverages differentially derived image representations enriched with positional encoding to guide the stroke prediction process. This integration enables the model to maintain heightened sensitivity to local details, resulting in more refined and nuanced stroke generation. Furthermore, we incorporate adversarial training into our framework, enhancing the accuracy of stroke prediction and thereby improving the overall realism and fidelity of the synthesized paintings. Extensive qualitative evaluations, complemented by a controlled user study, validate that our DQ-Transformer surpasses existing methods in both visual realism and artistic authenticity, typically achieving these results with fewer strokes.
- Open Access
- 26 June 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (13) , 6401
EFIB-Net: Information Bottleneck-Guided Multi-Resolution Attention Network for Robust ECG Denoising
Wearable electrocardiogram (ECG) monitoring enables continuous cardiovascular assessment, yet signals acquired in ambulatory environments are inevitably corrupted by baseline wander, electrode motion artifacts, and muscle interference, which obscure diagnostically critical waveform features. Existing deep learning denoisers rely on heuristic attention mechanisms and time-domain-only losses, lacking principled control over what information the network retains or discards. To address this limitation, we propose EFIB-Net, an information bottleneck-guided multi-resolution network for robust ECG denoising. The framework introduces two complementary components: an efficient frequency-guided attention module that derives temporal attention weights directly from the energy distribution of parallel multi-resolution convolutional branches, requiring only four learnable parameters while providing physically interpretable feature selection that naturally highlights QRS complexes, and a variational information bottleneck constraint at the encoder–decoder bottleneck that forces the latent representation to retain only reconstruction-relevant information and discard noise, guided by a spectral–temporal composite loss. To the best of our knowledge, we are among the first to explicitly introduce the information bottleneck principle into deep-learning-based ECG signal denoising. Experiments on the MIT-BIH Arrhythmia Database show that EFIB-Net outperforms ten traditional and deep learning baselines across four standard metrics—signal-to-noise ratio (SNR), root mean square error, percentage root-mean-square difference, and correlation coefficient; at an input SNR of −5 dB it reaches dB output SNR, surpassing the strongest attention-based competitor by 1.77 dB () while using only 0.45 M parameters and 10.8 ms inference latency per segment; downstream evaluation further demonstrates that the denoised signals achieve 99.18% R-peak detection sensitivity and 91.26% heartbeat classification F1-score, both within approximately one percentage point of the clean-signal upper bound, making it practical for real-time cardiac monitoring on resource-constrained wearable devices. Zero-shot cross-database evaluation on the QT Database further confirms generalizability, with only 0.54 dB degradation without retraining.- 26 June 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (13) , 6401
Open AccessEFIB-Net: Information Bottleneck-Guided Multi-Resolution Attention Network for Robust ECG Denoising
Wearable electrocardiogram (ECG) monitoring enables continuous cardiovascular assessment, yet signals acquired in ambulatory environments are inevitably corrupted by baseline wander, electrode motion artifacts, and muscle interference, which obscure diagnostically critical waveform features. Existing deep learning denoisers rely on heuristic attention mechanisms and time-domain-only losses, lacking principled control over what information the network retains or discards. To address this limitation, we propose EFIB-Net, an information bottleneck-guided multi-resolution network for robust ECG denoising. The framework introduces two complementary components: an efficient frequency-guided attention module that derives temporal attention weights directly from the energy distribution of parallel multi-resolution convolutional branches, requiring only four learnable parameters while providing physically interpretable feature selection that naturally highlights QRS complexes, and a variational information bottleneck constraint at the encoder–decoder bottleneck that forces the latent representation to retain only reconstruction-relevant information and discard noise, guided by a spectral–temporal composite loss. To the best of our knowledge, we are among the first to explicitly introduce the information bottleneck principle into deep-learning-based ECG signal denoising. Experiments on the MIT-BIH Arrhythmia Database show that EFIB-Net outperforms ten traditional and deep learning baselines across four standard metrics—signal-to-noise ratio (SNR), root mean square error, percentage root-mean-square difference, and correlation coefficient; at an input SNR of −5 dB it reaches dB output SNR, surpassing the strongest attention-based competitor by 1.77 dB () while using only 0.45 M parameters and 10.8 ms inference latency per segment; downstream evaluation further demonstrates that the denoised signals achieve 99.18% R-peak detection sensitivity and 91.26% heartbeat classification F1-score, both within approximately one percentage point of the clean-signal upper bound, making it practical for real-time cardiac monitoring on resource-constrained wearable devices. Zero-shot cross-database evaluation on the QT Database further confirms generalizability, with only 0.54 dB degradation without retraining. - 30 March 2026
- Published by Springer Nature in International Journal of Machine Learning and Cybernetics
- Vol. 17 (5) , 239
Meme selection via multiple-choice with masked language models in multimodal dialogue
- 30 March 2026
- Published by Springer Nature in International Journal of Machine Learning and Cybernetics
- Vol. 17 (5) , 239
Meme selection via multiple-choice with masked language models in multimodal dialogue
- 17 March 2026
- Published by Elsevier in Pattern Recognition
- Vol. 179
Minimizing the pretraining gap: Domain-aligned text-based person retrieval
- 17 March 2026
- Published by Elsevier in Pattern Recognition
- Vol. 179
Minimizing the pretraining gap: Domain-aligned text-based person retrieval
- 12 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 73 (6) , 4259-4270
AdapSNE: Adaptive Fireworks-Optimized and Entropy-Guided Dataset Sampling for Edge DNN Training
Training deep neural networks (DNNs) on edge devices faces challenges due to the large-scale datasets required, which are costly for edge devices, especially in large language model (LLM) tasks. To address this, a DNN-free method called Near-Memory Sampling (NMS) has been introduced. NMS reduces dimensionality and performs exemplar sampling in the reduced space, avoiding architectural bias and improving generalization. However, NMS has two limitations: 1) The mismatch between the search method and the non-monotonic property of the perplexity error function leads to the emergence of outliers; 2) Key parameter (i.e., target perplexity) is selected empirically, introducing arbitrariness and leading to uneven sampling. These two issues lead to representative bias of exemplars, resulting in degraded accuracy. To overcome these, we propose AdapSNE, which integrates the Fireworks Algorithm (FWA) for efficient non-monotonic search to avoid outliers and uses entropy-guided optimization for uniform sampling, ensuring representative training samples. To reduce the cost of iterative computations, we design an accelerator with custom dataflow and time-multiplexing mechanisms. Experimental results show that AdapSNE outperforms state-of-the-art methods, including both DNN-based (DQAS) and DNN-free (NMS) approaches, across small-scale image datasets, large-scale datasets, and the MMLU benchmark for LLM tasks.
- 12 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems I: Regular Papers
- Vol. 73 (6) , 4259-4270
AdapSNE: Adaptive Fireworks-Optimized and Entropy-Guided Dataset Sampling for Edge DNN Training
Training deep neural networks (DNNs) on edge devices faces challenges due to the large-scale datasets required, which are costly for edge devices, especially in large language model (LLM) tasks. To address this, a DNN-free method called Near-Memory Sampling (NMS) has been introduced. NMS reduces dimensionality and performs exemplar sampling in the reduced space, avoiding architectural bias and improving generalization. However, NMS has two limitations: 1) The mismatch between the search method and the non-monotonic property of the perplexity error function leads to the emergence of outliers; 2) Key parameter (i.e., target perplexity) is selected empirically, introducing arbitrariness and leading to uneven sampling. These two issues lead to representative bias of exemplars, resulting in degraded accuracy. To overcome these, we propose AdapSNE, which integrates the Fireworks Algorithm (FWA) for efficient non-monotonic search to avoid outliers and uses entropy-guided optimization for uniform sampling, ensuring representative training samples. To reduce the cost of iterative computations, we design an accelerator with custom dataflow and time-multiplexing mechanisms. Experimental results show that AdapSNE outperforms state-of-the-art methods, including both DNN-based (DQAS) and DNN-free (NMS) approaches, across small-scale image datasets, large-scale datasets, and the MMLU benchmark for LLM tasks.