Zhengdao Li

Zhengdao Li

Auto-generated profile, based on publicly available information

Zhengdao Li

School of Artificial Intelligence, Guangzhou University, Guangzhou, China

Research Focus & Metrics

Academic Career Length

5Years (2022-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
18 September 2023 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Verified
Journal Article
Open access
YLYuejin Li
Yuejin Li
Mi WangMi Wang
Mi Wang
Kai HwangKai Hwang
Kai Hwang
ZLZhengdao Li
Zhengdao Li
TJTongkai Ji
Tongkai Ji
This paper proposes a new satellite-based framework for global-scale remote sensing that is integrated with on-orbit cloud computing and AI ( artificial intelligence ) services. These spaced-based services cover the entire earth surfaces using massive LEO satellite constellation. Global-scale sensing of earth resources must be supported by massive number of LEO satellites equipped with cloud/AI computing services in real time. New satellite computer architectural features are presented along with some satellite constellation deployment topologies. We design satellite-based computers to support on-orbit remote sensing and AI scene analysis. This demands real-time performance without transmitting the sensed data back to earth for delayed processing. Notable space data services include on-orbit data sensing of large areas, machine learning from earth resources data, earth scene/event analysis, geomorphology observation, smart city management, disaster relief, global healthcare IoT, and environmental ecology protection, etc. We attempt to achieve high-efficiency earth resources utilization along with green energy, low cost, and robustness in real-life services.

20 July 2022 in IEEE Transactions on Cloud Computing
Verified
Journal Article
Closed access
KSKefan Shuai
Kefan Shuai
YMYiming Miao
Yiming Miao
Kai HwangKai Hwang
Kai Hwang
ZLZhengdao Li
Zhengdao Li
In the Big Data era, resource-constrained mobile devices generate an overwhelmingly large amount of data with complex tasks that demand distributed execution. Offloading computation-intensive tasks to nearby edge clouds is promising to solve this problem. However, mobile end devices cannot handle heterogeneous or delay-sensitive tasks. These end devices are also energy constrained with weak adaptability to environment changes. To address and tackle these problems, we present a two-module transfer reinforcement learning (TRL) framework for adaptive task offloading. A domain adaptation module is used to align heterogeneous characteristics of mobile devices. The TRL makes offloading decisions with a deep reinforcement learning (DRL) module. We evaluate the performance of TRL through real-world experiments on edge clouds. Our experiment results show that TRL reduces the task processing time by a factor of 20% from using three well known DRL methods. Our method achieved (15.4$\sim$40)% reduction in task drop rate over these methods. With domain adaptation, the TRL results in (50$\sim$80)% reduction in model convergence time. These advantages in using the TRL framework make it appealing in real-life edge computing applications.

24 June 2022 in IEEE Transactions on Cloud Computing
Verified
Journal Article
Closed access
YLYuejin Li
Yuejin Li
Kai HwangKai Hwang
Kai Hwang
KSKefan Shuai
Kefan Shuai
ZLZhengdao Li
Zhengdao Li
Albert ZomayaAlbert Zomaya
Albert Zomaya
Distributed cloud/edge resources are needed to execute pervasive artificial intelligence tasks, collectively. The AI workload and data sets have variable multitasking granularity, privacy constraints, and communication latency concerns. This paper presents a novel federated cloud/edge (FCE) framework, illustrated by distributed medical image processing across multiple hospital sites. This federated cloud system appeals to train many machine learning models efficiently with workload balancing and reduced communication overheads. We tested the FCE model on a multi-cloud platform recently built at the Chinese University of Hong Kong in Shenzhen. We claim three distinct advantages in using the FCE system. First, our federated cloud system results in 41.3% reduction in total AI processing time in large-scale ML/DL experiments. Second, high machine model accuracy was achieved at 87% level in telemedicine experiments. The virtual graph helps reduce internode traffic latencies to avoid ML inference slowdowns. Third, the system can tolerate multiple cloud failures to enter a graceful degradation mode in case of node failures. The scalable performance gains in AI processing speed, model accuracy, and fault tolerance make our federated clouds a truly viable approach to solving massive AI multitasking problems in pervasive AI applications.

14 April 2025 in IEEE Transactions on Information Forensics and Security
Verified
Journal Article
Closed access
WMWanlun Ma
Wanlun Ma
DWDerui Wang
Derui Wang
YSYiliao Song
Yiliao Song
MXMinhui Xue
Minhui Xue
VWVincent Sheng Wen
Vincent Sheng Wen
YXYang Xiang
Yang Xiang
Model inversion (MI) attacks, for which effective defense strategies are still lacking, pose significant risks to privacy by reconstructing private training data through access to well-trained classifiers. Addressing this concern, this study introduces TrapNet, designed to defend against advanced MI attacks while maintaining good model utility. TrapNet intentionally injects trapdoors into the classification manifold of the protected target model. In this way, TrapNet can effectively mislead MI attack optimization. Specifically, TrapNet leverages a conditional GAN (cGAN) trained on the private dataset to generate diverse and realistic trapdoor samples. In addition, we propose a graph-matching self-obfuscation strategy and an entropy regularization technique to optimize trapdoor injection while preserving model utility. Compared to the existing defense, TrapNet can provide universal protection to all target classes without access to any auxiliary public data. Extensive experiments on CelebA, VGG-Face, and VGG-Face2 datasets demonstrate TrapNet’s superior performance over existing defenses, including the most advanced NetGuard and BiDO, against state-of-the-art model inversion attacks, i.e., PLG-MI, LOMMA, and Plug&Play.

4 February 2026 in ACM Computing Surveys
Verified
Journal Article
Open access
PYPei-Gen Ye
Pei-Gen Ye
HRHuali Ren
Huali Ren
ZLZhengdao Li
Zhengdao Li
AYAnli Yan
Anli Yan
HYHongyang Yan
Hongyang Yan
JLJin Li
Jin Li
State-of-the-art watermarking and fingerprinting techniques for Large Language Models (LLMs) are explored, with our analysis spanning a wide array of methodologies designed to protect the intellectual property of LLMs. The review of watermarking techniques is based on embedding watermarks during the training, logits generation, and token sampling phases. Meanwhile, we investigate the application of watermarking technology in multimodal LLMs and potential attacks on watermarks. Moreover, our examination of fingerprinting techniques revealed the ingenuity behind methods used to identify LLMs. We discussed the development of fingerprints based on model behavior and using deep learning models to learn thresholds for fingerprint comparison. Our survey has underscored the importance of advancing security measures for LLMs, especially in light of the increasing sophistication of adversarial attacks. As LLMs continue to play a pivotal role in advancing AI technologies, developing and refining security measures that safeguard their intellectual property and ensure their ethical deployment is imperative.

12 August 2025 in ACM Transactions on Multimedia Computing, Communications, and Applications
Verified
Journal Article
Open access
JLJiaxing Li
Jiaxing Li
YTYu-An Tan
Yu-An Tan
JYJie Yang
Jie Yang
ZLZhengdao Li
Zhengdao Li
HYHeng Ye
Heng Ye
YLYuanzhang Li
Yuanzhang Li
Many edge computing applications based on computer vision have harnessed the power of deep learning. As an emerging deep learning model for vision, Vision Transformer models have recently achieved record-breaking performance in various vision tasks. But many recent studies on the robustness of the Vision Transformer have shown that the Vision Transformer is still vulnerable to adversarial attacks and is easily affected by adversarial attacks, causing the model to misclassify the input. In this work, we ask an intriguing question: “Can Adversarial Perturbations against Vision Transformers be detected with model explanations?” Driven by this question, we observe that benign samples and adversarial examples have different attribution maps after applying the Grad-CAM interpretability method on the Vision Transformer model. We demonstrate that an adversarial example is a Feature Shift of the input data, which leads to an Attention Deviation of the visual model. We propose a framework for capturing the Attention Deviation of vision models to defend against adversarial attacks. Furthermore, experiments show that our model achieves expectative results.