Zhi Lu

Zhi Lu

Auto-generated profile, based on publicly available information

Zhi Lu

Huazhong University of Science and Technology

Research Focus & Metrics

Academic Career Length

4Years (2023-Present)

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

Representative Publications

View All Publications
20 May 2024 in IEEE Transactions on Information Forensics and Security
Verified
Journal Article
Closed access
ZLZhi Lu
Zhi Lu
SLSongfeng Lu
Songfeng Lu
YCYongQuan Cui
YongQuan Cui
XTXueMing Tang
XueMing Tang
JWJunJun Wu
JunJun Wu
Federated Learning (FL), a distributed learning paradigm optimizing communication costs and enhancing privacy by uploading gradients instead of raw data, now confronts security challenges. It is particularly vulnerable to Byzantine poisoning attacks and potential privacy breaches via inference attacks. While homomorphic encryption and secure multi-party computation have been employed to design robust FL mechanisms, these predominantly rely on Euclidean distance or median-based metrics and often fall short in comprehensively defending against advanced poisoning attacks, such as adaptive attacks. Addressing this issue, our study introduces “Split-Aggregation”, a lightweight privacy-preserving FL solution capable of withstanding adaptive attacks. This method maintains a computational complexity of $O(dkN+k^{3})$ and a communication overhead of $O(dN)$ , performing comparably to FedAvg when $k=10$ . Here, d represents the gradient dimension, N the number of users, and k the rank chosen during random singular value decomposition. Additionally, we utilize adaptive weight coefficients to mitigate gradient descent issues in honest users caused by non-independent and identically distributed (Non-IID) data. The proposed method’s security and robustness are theoretically proven, with its complexity thoroughly analyzed. Experimental results demonstrate that at $k=10$ , this method surpasses the top-1 accuracy of current state-of-the-art robust privacy-preserving FL approaches. Moreover, opting for a smaller k significantly boosts efficiency with only marginal compromises in accuracy.

26 August 2024
Verified
Book Chapter
Closed access
PDF file
HNHewang Nie
Hewang Nie
SLSongfeng Lu
Songfeng Lu
MWMu Wang
Mu Wang
JXJue Xiao
Jue Xiao
ZLZhi Lu
Zhi Lu
ZYZepu Yi
Zepu Yi

28 August 2023 in IEEE Transactions on Artificial Intelligence
Verified
Journal Article
Closed access
ZLZhi Lu
Zhi Lu
SLSongfeng Lu
Songfeng Lu
XTXueming Tang
Xueming Tang
JWJunjun Wu
Junjun Wu
Federated Learning (FL) safeguards user privacy by uploading gradients instead of raw data. However, inference attacks can reconstruct raw data using gradients uploaded by users in FL. To mitigate this issue, researchers have combined privacy computing techniques with FL. However, these tech-niques may not ensure the Byzantine robustness of aggregation or the integrity of the aggregated outcomes. Most current robust privacy FL methods assess differences between gradients and benchmarks in the direction, allowing adversaries to poison the aggregation against the magnitude. Furthermore, these methods cannot ensure the integrity of the aggregation results. To over-come these challenges, this study proposes a novel algorithm, Robust and Verifiable Privacy Federated Learning(RVPFL), which can more effectively eliminate the poisoning attack of the opponent by measuring the direction and magnitude of the gradient in the ciphertext state. The proposed algorithm guarantees the integrity of server aggregation results while safe-guarding user privacy. In this study, comprehensive theoretical analysis and experimental validation of RVPFL are conducted to demonstrate its superiority. The proposed RVPFL algorithm solves the Byzantine robustness problem of aggregation and the integrity problem of aggregation results, which helps to research and develop more robust and effective privacy-preserving federal learning techniques.


20 March 2026 in ACM Transactions on Intelligent Systems and Technology
Verified
Journal Article
Open access
HNHewang Nie
Hewang Nie
JXJue Xiao
Jue Xiao
ZLZhi Lu
Zhi Lu
RSRenfei Shen
Renfei Shen
SLSongfeng Lu
Songfeng Lu
As Deep Neural Networks (DNNs) become integral to critical applications, protecting their Intellectual Property (IP) has become paramount. Neural network watermarking is a technique that embeds unique identifiers into models, asserting ownership and deterring unauthorized use. However, sophisticated attacks can deactivate or remove these watermarks without significantly compromising model performance, undermining current protection strategies. In this article, we introduce the first method for reactivating deactivated neural network watermarks in altered DNN models without requiring access to the original model parameters or training data. By formulating the reactivation process as an optimization problem, we employ projected gradient descent to identify new trigger inputs that restore the embedded watermark. Regularization techniques are incorporated to ensure these triggers resemble legitimate inputs, enhancing both stealth and practicality. Through experiments on various benchmark datasets and model architectures, we demonstrate the effectiveness of our method against common model alterations, including fine-tuning, pruning, and surrogate model attacks. Our work addresses a critical gap in DNN IP protection, offering a robust and practical solution for watermark reactivation. This empowers model owners to assert their rights even in the face of advanced adversarial tactics.

13 February 2026 in IEEE Transactions on Artificial Intelligence
Verified
Journal Article
Closed access
YFYuanqing Feng
Yuanqing Feng
TBTao Bai
Tao Bai
ZLZhi Lu
Zhi Lu
XTXueming Tang
Xueming Tang
SLSongfeng Lu
Songfeng Lu
Traditional federated learning requires each agent to send its local model to a central server for aggregation. However, directly transmitting the models to the server can raise privacy concerns, as the server may be able to reconstruct the original training data from the gradients of the agent models. To address this privacy issue, researchers have adopted secure aggregation methods for aggregating the global model. However, current secure aggregation schemes primarily use centralized aggregation concepts and masking techniques, which limit the efficiency of federated learning and hinder the further development of secure aggregation. In this paper, we leverage the concept of decentralization to reduce communication and computation overhead at individual nodes in secure aggregation, thereby improving the efficiency of federated learning. Additionally, we design a FedAvg aggregation algorithm based on secure multi-party computation technology and propose a dropout handling mechanism that differs from traditional secure aggregation methods. Through theoretical analysis, we show that the maximum communication complexity at individual nodes is reduced from O(pn) to $O(p √ n)$. Finally, we conduct experiments to evaluate the proposed scheme and compare it with several baselines. The experimental results show that, in a local area network (LAN) environment, the execution speed of our scheme improves by 2.7×-4.3×, and in a wide area network (WAN) environment, the execution speed improves by 3.5×-3.85×.