Zhe Ma
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Zhe Ma has a total of 83 co-authors. Here are the top co-authors by number of shared publications:
- Chunyi Zhou
Zhejiang University, Hangzhou, China
2 shared publications - Jianfeng Dong
Zhejiang Gongshang University, Hangzhou, China
2 shared publications - Shouling Ji
Zhejiang University, Hangzhou, China
2 shared publications - Yong Yang
Zhejiang University, Hangzhou, China
2 shared publications - Andrey Ignatov
ETH Zurich
1 shared publication - Baolong Liu
Zhejiang Gongshang University & Zhejiang Key Lab of E-Commerce, Hangzhou, China
1 shared publication
Latest Publications
- Open Access
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 1252-1266
Customization under Fire: Plugin Poisoning in Text-to-Image Ecosystem
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 1252-1266
Open AccessCustomization under Fire: Plugin Poisoning in Text-to-Image Ecosystem
- Top Cited PapersOpen Access
- 7 May 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Software Engineering and Methodology
Navigating the Risks: A Survey of Security and Privacy Threats in LLM-Based Agents
Large language models (LLMs) are increasingly embedded into software engineering workflows as autonomous or semi-autonomous agents for code generation, test synthesis, tool orchestration, and end-to-end task automation. While these agentic capabilities promise substantial productivity and quality gains, they also introduce distinct security and privacy risks at the levels of artifacts, processes, and socio-technical deployment pipelines. These risks threaten the trustworthiness of LLM-powered software and complicate core engineering activities such as specification, design, verification, maintenance, and governance. This survey systematizes the threat landscape for LLM-based agents from a software engineering perspective. We propose a taxonomy organized by threat sources and their impacts across the software lifecycle, explicitly capturing cross-module and cross-stage attack surfaces (e.g., tool-use mediation, memory management, and environment interaction). We further distill six agent features grounded in information-flow and control-flow properties that shape risk exposure. We synthesize current mitigation techniques and their limitations for each feature, and conduct four case studies on web search, gaming, navigation, and software development to illustrate practical risks and failure modes. Building on these analyses, we outline research directions in data, methodology, and policy to guide the development of secure and trustworthy LLM-based agents.
- 7 May 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Software Engineering and Methodology
Top Cited PapersOpen AccessNavigating the Risks: A Survey of Security and Privacy Threats in LLM-Based Agents
Large language models (LLMs) are increasingly embedded into software engineering workflows as autonomous or semi-autonomous agents for code generation, test synthesis, tool orchestration, and end-to-end task automation. While these agentic capabilities promise substantial productivity and quality gains, they also introduce distinct security and privacy risks at the levels of artifacts, processes, and socio-technical deployment pipelines. These risks threaten the trustworthiness of LLM-powered software and complicate core engineering activities such as specification, design, verification, maintenance, and governance. This survey systematizes the threat landscape for LLM-based agents from a software engineering perspective. We propose a taxonomy organized by threat sources and their impacts across the software lifecycle, explicitly capturing cross-module and cross-stage attack surfaces (e.g., tool-use mediation, memory management, and environment interaction). We further distill six agent features grounded in information-flow and control-flow properties that shape risk exposure. We synthesize current mitigation techniques and their limitations for each feature, and conduct four case studies on web search, gaming, navigation, and software development to illustrate practical risks and failure modes. Building on these analyses, we outline research directions in data, methodology, and policy to guide the development of secure and trustworthy LLM-based agents.
- 8 April 2026
- Published by Elsevier in Medical Image Analysis
- Vol. 111, 104075
Fine-grained and multi-pattern anti-nuclear antibody recognition: A new dataset and framework
- 8 April 2026
- Published by Elsevier in Medical Image Analysis
- Vol. 111, 104075
Fine-grained and multi-pattern anti-nuclear antibody recognition: A new dataset and framework
- Open Access
- 26 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 144-152
Video Infringement Detection via Feature Disentanglement and Mutual Information Maximization
The self-media era provides us tremendous high quality videos. Unfortunately, frequent video copyright infringements are now seriously damaging the interests and enthusiasm of video creators. Identifying infringing videos is therefore a compelling task. Current state-of-the-art methods tend to simply feed high-dimensional mixed video features into deep neural networks and count on the networks to extract useful representations. Despite its simplicity, this paradigm heavily relies on the original entangled features and lacks constraints guaranteeing that useful task-relevant semantics are extracted from the features. In this paper, we seek to tackle the above challenges from two aspects: (1) We propose to disentangle an original high-dimensional feature into multiple sub-features, explicitly disentangling the feature into exclusive lower-dimensional components. We expect the sub-features to encode non-overlapping semantics of the original feature and remove redundant information. (2) On top of the disentangled sub-features, we further learn an auxiliary feature to enhance the sub-features. We theoretically analyzed the mutual information between the label and the disentangled features, arriving at a loss that maximizes the extraction of task-relevant information from the original feature. Extensive experiments on two large-scale benchmark datasets (i.e., SVD and VCSL) demonstrate that our method achieves 90.1% TOP-100 mAP on the large-scale SVD dataset and also sets the new state-of-the-art on the VCSL benchmark dataset. Our code and model have been released at https://github.com/yyyooooo/DMI/, hoping to contribute to the community.
- 26 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 144-152
Open AccessVideo Infringement Detection via Feature Disentanglement and Mutual Information Maximization
The self-media era provides us tremendous high quality videos. Unfortunately, frequent video copyright infringements are now seriously damaging the interests and enthusiasm of video creators. Identifying infringing videos is therefore a compelling task. Current state-of-the-art methods tend to simply feed high-dimensional mixed video features into deep neural networks and count on the networks to extract useful representations. Despite its simplicity, this paradigm heavily relies on the original entangled features and lacks constraints guaranteeing that useful task-relevant semantics are extracted from the features. In this paper, we seek to tackle the above challenges from two aspects: (1) We propose to disentangle an original high-dimensional feature into multiple sub-features, explicitly disentangling the feature into exclusive lower-dimensional components. We expect the sub-features to encode non-overlapping semantics of the original feature and remove redundant information. (2) On top of the disentangled sub-features, we further learn an auxiliary feature to enhance the sub-features. We theoretically analyzed the mutual information between the label and the disentangled features, arriving at a loss that maximizes the extraction of task-relevant information from the original feature. Extensive experiments on two large-scale benchmark datasets (i.e., SVD and VCSL) demonstrate that our method achieves 90.1% TOP-100 mAP on the large-scale SVD dataset and also sets the new state-of-the-art on the VCSL benchmark dataset. Our code and model have been released at https://github.com/yyyooooo/DMI/, hoping to contribute to the community.
- Open Access
- 18 July 2023
- Published by Association for Computing Machinery (ACM)
- p. 1273-1282
From Region to Patch: Attribute-Aware Foreground-Background Contrastive Learning for Fine-Grained Fashion Retrieval
- 18 July 2023
- Published by Association for Computing Machinery (ACM)
- p. 1273-1282
Open AccessFrom Region to Patch: Attribute-Aware Foreground-Background Contrastive Learning for Fine-Grained Fashion Retrieval
- Open Access
Realistic Bokeh Effect Rendering on Mobile GPUs, Mobile AI & AIM 2022 challenge: Report
As mobile cameras with compact optics are unable to produce a strong bokeh effect, lots of interest is now devoted to deep learning-based solutions for this task. In this Mobile AI challenge, the target was to develop an efficient end-to-end AI-based bokeh effect rendering approach that can run on modern smartphone GPUs using TensorFlow Lite. The participants were provided with a large-scale EBB! bokeh dataset consisting of 5K shallow / wide depth-of-field image pairs captured using the Canon 7D DSLR camera. The runtime of the resulting models was evaluated on the Kirin 9000's Mali GPU that provides excellent acceleration results for the majority of common deep learning ops. A detailed description of all models developed in this challenge is provided in this paper.
Open AccessRealistic Bokeh Effect Rendering on Mobile GPUs, Mobile AI & AIM 2022 challenge: Report
As mobile cameras with compact optics are unable to produce a strong bokeh effect, lots of interest is now devoted to deep learning-based solutions for this task. In this Mobile AI challenge, the target was to develop an efficient end-to-end AI-based bokeh effect rendering approach that can run on modern smartphone GPUs using TensorFlow Lite. The participants were provided with a large-scale EBB! bokeh dataset consisting of 5K shallow / wide depth-of-field image pairs captured using the Canon 7D DSLR camera. The runtime of the resulting models was evaluated on the Kirin 9000's Mali GPU that provides excellent acceleration results for the majority of common deep learning ops. A detailed description of all models developed in this challenge is provided in this paper.
- 2 April 2022
- Published by Elsevier in Annals of Oncology
- Vol. 33, S31
8MO The efficacy and safety of TQ-B3101 monotherapy in the first-line treatment in patients with ROS1-positive non-small cell lung cancer
- 2 April 2022
- Published by Elsevier in Annals of Oncology
- Vol. 33, S31
8MO The efficacy and safety of TQ-B3101 monotherapy in the first-line treatment in patients with ROS1-positive non-small cell lung cancer
- Open Access
- 1 August 2021
- Published by IOP Publishing in Journal of Physics: Conference Series
- Vol. 2004 (1) , 012008
Alleviating the unfairness of recommendation by eliminating the conformity bias
Recommendation systems play a critical role in the Internet era. In the past decade, many methods have been proposed to improve the accuracy of recommendation. However, simply increasing the accuracy is usually accompanied by the popularity bias in the recommendation results. In addition, from the perspective of item providers the popularity bias will bring unfairness in exposure of different items to users. In this paper, we propose a method to alleviate problems of unfairness and popularity bias. The main idea is based on the hypothesis that some interactions of users with popular items are not due to their preferences, but attribute to the conformity propensity, which aggravates the popularity bias problem. Thus we consider to eliminate these so called “unreliable interactions” to alleviate the unfairness. We propose a two-step preferential diffusion process between conformity users and popular items in the random walk recommendation algorithm to reduce the effect of unreliable interaction and alleviate the popularity bias and unfairness under the condition of ensuring accuracy. The experiment results on the two benchmark datasets, Movie-Lens and Netflix, show that our algorithm can improve the fairness of recommendation results and alleviate the popularity bias while achieving slightly increased recommendation accuracy.
- 1 August 2021
- Published by IOP Publishing in Journal of Physics: Conference Series
- Vol. 2004 (1) , 012008
Open AccessAlleviating the unfairness of recommendation by eliminating the conformity bias
Recommendation systems play a critical role in the Internet era. In the past decade, many methods have been proposed to improve the accuracy of recommendation. However, simply increasing the accuracy is usually accompanied by the popularity bias in the recommendation results. In addition, from the perspective of item providers the popularity bias will bring unfairness in exposure of different items to users. In this paper, we propose a method to alleviate problems of unfairness and popularity bias. The main idea is based on the hypothesis that some interactions of users with popular items are not due to their preferences, but attribute to the conformity propensity, which aggravates the popularity bias problem. Thus we consider to eliminate these so called “unreliable interactions” to alleviate the unfairness. We propose a two-step preferential diffusion process between conformity users and popular items in the random walk recommendation algorithm to reduce the effect of unreliable interaction and alleviate the popularity bias and unfairness under the condition of ensuring accuracy. The experiment results on the two benchmark datasets, Movie-Lens and Netflix, show that our algorithm can improve the fairness of recommendation results and alleviate the popularity bias while achieving slightly increased recommendation accuracy.