Shuning Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Shuning Zhang has a total of 122 co-authors. Here are the top co-authors by number of shared publications:
- Xin Yi
Tsinghua University, Beijing, China
27 shared publications - Hewu Li
Tsinghua University, Beijing, China
22 shared publications - Ying Ma
The University of Melbourne, Melbourne, Australia
8 shared publications - Yongquan Hu
National University of Singapore, Singapore, Singapore
7 shared publications - Shixuan Li
Tsinghua University, Beijing, China
6 shared publications - Yan Kong
Tsinghua University, Beijing, China
5 shared publications
Latest Publications
- Open Access
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 2000-2014
Investigating Novice Researchers' Perceptions of Research Privacy Within LLM-Assisted Workflows
Large Language Model (LLM)-assisted scholarly workflows introduce critical privacy and intellectual property risks. As a uniquely vulnerable cohort driven by publication pressure, and often weaker institutional support, novice researchers rely heavily on public LLMs, and must navigate high-stakes privacy-publication trade-offs. To investigate these concerns, we conducted semi-structured interviews with 44 researchers across diverse disciplines. We found that the fear of idea leakage paradoxically prompted some participants to increase their reliance on LLMs, as researchers utilize them to accelerate publication. Participants' beliefs--that their ideas lack unique value and their inputs are safely diluted in massive datasets--are contradicted by technical findings in the prior literature. From the interviews, we identified participants using five types of mitigation, including input fragmentation and adversarial probing, though participants also largely perceived these measures as ineffective. We outline recommendations that include implementing institution-level sandboxed environments, scenario-based privacy education, and verifiable data-deletion audits to improve transparency.
- 14 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 2000-2014
Open AccessInvestigating Novice Researchers' Perceptions of Research Privacy Within LLM-Assisted Workflows
Large Language Model (LLM)-assisted scholarly workflows introduce critical privacy and intellectual property risks. As a uniquely vulnerable cohort driven by publication pressure, and often weaker institutional support, novice researchers rely heavily on public LLMs, and must navigate high-stakes privacy-publication trade-offs. To investigate these concerns, we conducted semi-structured interviews with 44 researchers across diverse disciplines. We found that the fear of idea leakage paradoxically prompted some participants to increase their reliance on LLMs, as researchers utilize them to accelerate publication. Participants' beliefs--that their ideas lack unique value and their inputs are safely diluted in massive datasets--are contradicted by technical findings in the prior literature. From the interviews, we identified participants using five types of mitigation, including input fragmentation and adversarial probing, though participants also largely perceived these measures as ineffective. We outline recommendations that include implementing institution-level sandboxed environments, scenario-based privacy education, and verifiable data-deletion audits to improve transparency.
- Open Access
- 6 October 2026
- Published by Elsevier in SSRN Electronic Journal
PriVis: Proactive and Explainable Visualization of Visual Privacy Inferences on AI-Powered AR Glasses
With the rapid advancement of AI-powered augmented reality (AR) glasses, application providers can use Vision-Language Models (VLMs) to derive sensitive user profiles through latent privacy inferences. However, techniques to mitigate such risks remain underexplored. We present PriVis, an explainable visualization framework for AI-powered AR glasses designed to enhance user awareness and support informed control over privacy inferences derived from environmental objects. PriVis identifies source objects, infers private attributes, and presents explanations that quantify inference likelihood and the contribution of individual objects. A user study (N=24) evaluated PriVis against alternative conditions lacking explainable visualization or user control. Results show that PriVis outperformed baselines in terms of privacy awareness and control.
- 6 October 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessPriVis: Proactive and Explainable Visualization of Visual Privacy Inferences on AI-Powered AR Glasses
With the rapid advancement of AI-powered augmented reality (AR) glasses, application providers can use Vision-Language Models (VLMs) to derive sensitive user profiles through latent privacy inferences. However, techniques to mitigate such risks remain underexplored. We present PriVis, an explainable visualization framework for AI-powered AR glasses designed to enhance user awareness and support informed control over privacy inferences derived from environmental objects. PriVis identifies source objects, infers private attributes, and presents explanations that quantify inference likelihood and the contribution of individual objects. A user study (N=24) evaluated PriVis against alternative conditions lacking explainable visualization or user control. Results show that PriVis outperformed baselines in terms of privacy awareness and control.
- Open Access
- 15 August 2026
- Published by Elsevier in SSRN Electronic Journal
Designing Proactive Memory Privacy Visualization and Management in LLM-based Chatbots
The integration of long-term memory into wearable Large Language Model (LLM)-based agents enhances personalization, but introduces inferential privacy risks, where sensitive user attributes are deduced from non-sensitive interaction patterns. Users are frequently unaware of these risks, and existing memory systems often fail to provide necessary transparency or control. We propose MemoAnalyzer, an add-on to LLM-based chatbots' interfaces that visualizes the sensitivity of inferred private information within memories and tracks contributing chat histories. MemoAnalyzer uses a prompt-based method to infer and highlight sensitive data. It mapped background color temperature and transparency to inferred sensitivity and confidence respectively, enabling users to easily identify and modify inferences. A 5-day evaluation (N=72) compared MemoAnalyzer with ChatGPT memory setting and two research baselines, one with cluster operations aggregating chats and one with drag-based chat-level operations. Results showed that MemoAnalyzer improved privacy awareness and protection without significantly compromising interaction utility.
- 15 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessDesigning Proactive Memory Privacy Visualization and Management in LLM-based Chatbots
The integration of long-term memory into wearable Large Language Model (LLM)-based agents enhances personalization, but introduces inferential privacy risks, where sensitive user attributes are deduced from non-sensitive interaction patterns. Users are frequently unaware of these risks, and existing memory systems often fail to provide necessary transparency or control. We propose MemoAnalyzer, an add-on to LLM-based chatbots' interfaces that visualizes the sensitivity of inferred private information within memories and tracks contributing chat histories. MemoAnalyzer uses a prompt-based method to infer and highlight sensitive data. It mapped background color temperature and transparency to inferred sensitivity and confidence respectively, enabling users to easily identify and modify inferences. A 5-day evaluation (N=72) compared MemoAnalyzer with ChatGPT memory setting and two research baselines, one with cluster operations aggregating chats and one with drag-based chat-level operations. Results showed that MemoAnalyzer improved privacy awareness and protection without significantly compromising interaction utility.
- Open Access
- 15 June 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies
- Vol. 10  (2) , 1-48
Beyond Screen Time: Inferring Everyday Life Context from Diverse Smartphone Data
Smartphones serve as ubiquitous sensors of daily life, generating rich data streams. However, existing research often focuses narrowly on usage metrics from small or homogeneous samples, limiting our understanding of how phone data reflects broader real-world contexts. To address this gap, this paper introduces a hierarchical context inference framework designed to systematically map raw and multimodal data into two deep contextual dimensions: the Informational-Cognitive Context and the Physical-Situational Context. We empirically validated this framework through a large-scale, 14-day in-the-wild study involving a diverse cohort ( N = 539) spanning 15 occupations and 31 provincial regions in China. Applying our framework to this data reveals generalizable phenomena: the information environment frequently mixes multiple topics, shows high emotional fluctuation, and exhibits task-driven volatility rhythms. Physically, users own diverse Bluetooth peripherals but show focused usage, and our models reveal stable and cyclical phone interaction patterns across distinct operational states. Integrating these dimensions yields nuanced user personas that differ across occupational, hobby, and personality groups. This work contributes a validated framework for understanding the interplay between users' digital and physical worlds, supported by the release of a comprehensive multimodal dataset.
- 15 June 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies
- Vol. 10  (2) , 1-48
Open AccessBeyond Screen Time: Inferring Everyday Life Context from Diverse Smartphone Data
Smartphones serve as ubiquitous sensors of daily life, generating rich data streams. However, existing research often focuses narrowly on usage metrics from small or homogeneous samples, limiting our understanding of how phone data reflects broader real-world contexts. To address this gap, this paper introduces a hierarchical context inference framework designed to systematically map raw and multimodal data into two deep contextual dimensions: the Informational-Cognitive Context and the Physical-Situational Context. We empirically validated this framework through a large-scale, 14-day in-the-wild study involving a diverse cohort ( N = 539) spanning 15 occupations and 31 provincial regions in China. Applying our framework to this data reveals generalizable phenomena: the information environment frequently mixes multiple topics, shows high emotional fluctuation, and exhibits task-driven volatility rhythms. Physically, users own diverse Bluetooth peripherals but show focused usage, and our models reveal stable and cyclical phone interaction patterns across distinct operational states. Integrating these dimensions yields nuanced user personas that differ across occupational, hobby, and personality groups. This work contributes a validated framework for understanding the interplay between users' digital and physical worlds, supported by the release of a comprehensive multimodal dataset.
- Open Access
- 15 June 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies
- Vol. 10  (2) , 1-38
The Pervasive Blind Spot: Benchmarking VLM Inference Risks on Everyday Personal Videos
Applying Vision-Language Models (VLMs) to pervasive personal videos introduces profound privacy risks. This paper addresses the critical yet unexplored inferential privacy threat, specifically the risk of inferring sensitive personal attributes from seemingly benign data. To address this gap, we crowdsourced a dataset of 508 everyday personal videos from 58 individuals, and benchmarked VLM inference capabilities against human performance. Our findings reveal three key insights: (1) VLMs surpass recruited human evaluators in inferential accuracy, analyzing temporal behavioral patterns rather than relying solely on object recognition. (2) Inferential risk is strongly correlated with specific video characteristics and prompting strategies. (3) VLM-driven explanation towards the inference is unreliable, as we observe a disconnect between the model's reasoning and evidential impact, where ubiquitous objects often serve as misleading confounders.
- 15 June 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies
- Vol. 10  (2) , 1-38
Open AccessThe Pervasive Blind Spot: Benchmarking VLM Inference Risks on Everyday Personal Videos
Applying Vision-Language Models (VLMs) to pervasive personal videos introduces profound privacy risks. This paper addresses the critical yet unexplored inferential privacy threat, specifically the risk of inferring sensitive personal attributes from seemingly benign data. To address this gap, we crowdsourced a dataset of 508 everyday personal videos from 58 individuals, and benchmarked VLM inference capabilities against human performance. Our findings reveal three key insights: (1) VLMs surpass recruited human evaluators in inferential accuracy, analyzing temporal behavioral patterns rather than relying solely on object recognition. (2) Inferential risk is strongly correlated with specific video characteristics and prompting strategies. (3) VLM-driven explanation towards the inference is unreliable, as we observe a disconnect between the model's reasoning and evidential impact, where ubiquitous objects often serve as misleading confounders.
- Open Access
- 30 April 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Human-Computer Interaction
- Vol. 10  (2) , 1
Corrigendum: Exploring Collaboration Patterns and Strategies in Human-AI Co-creation through the Lens of Agency: A Scoping Review of the Top-tier HCI Literature
This is a corrigendum for the article "Exploring Collaboration Patterns and Strategies in Human-AI Co-creation through the Lens of Agency: A Scoping Review of the Top-tier HCI Literature" published in Proc. ACM Hum-Comput. Interact. 9, 7, Article CSCW413 (November 2025), 43 pages.
- 30 April 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Human-Computer Interaction
- Vol. 10  (2) , 1
Open AccessCorrigendum: Exploring Collaboration Patterns and Strategies in Human-AI Co-creation through the Lens of Agency: A Scoping Review of the Top-tier HCI Literature
This is a corrigendum for the article "Exploring Collaboration Patterns and Strategies in Human-AI Co-creation through the Lens of Agency: A Scoping Review of the Top-tier HCI Literature" published in Proc. ACM Hum-Comput. Interact. 9, 7, Article CSCW413 (November 2025), 43 pages.
- Open Access
- 13 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-19
Characterizing Unintended Consequences of GUI Agents For Web Browsing
- 13 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-19
Open AccessCharacterizing Unintended Consequences of GUI Agents For Web Browsing
- Open Access
- 13 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-22
VisGuardian: A Lightweight Group-based Visual Privacy Control Technique For Smart Glasses in Home Environments
Always-on sensing of AI applications on AR glasses makes traditional permission techniques inefficient for context-dependent private visual data within home environments. Home presents a challenging privacy context due to massive sensitive objects and the intimate nature of daily routines. We propose VisGuardian, a fine-grained content-based visual permission technique for AR glasses. VisGuardian features a group-based control mechanism that enables users to efficiently manage permissions for multiple private objects. VisGuardian detects objects using YOLO and adopts a pre-classified schema to group them. By selecting a single object, users can obscure groups of related objects based on criteria including privacy sensitivity, object category, or spatial proximity. A technical evaluation shows VisGuardian achieves mAP50 of 0.6704 with only 14.0 ms latency and a 1.7% increase in battery consumption per hour. Furthermore, a user study (N=24) comparing VisGuardian to slider-based and object-based baselines found it to be significantly faster for setting permissions and was preferred by users for its efficiency, effectiveness, and ease of use.
- 13 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-22
Open AccessVisGuardian: A Lightweight Group-based Visual Privacy Control Technique For Smart Glasses in Home Environments
Always-on sensing of AI applications on AR glasses makes traditional permission techniques inefficient for context-dependent private visual data within home environments. Home presents a challenging privacy context due to massive sensitive objects and the intimate nature of daily routines. We propose VisGuardian, a fine-grained content-based visual permission technique for AR glasses. VisGuardian features a group-based control mechanism that enables users to efficiently manage permissions for multiple private objects. VisGuardian detects objects using YOLO and adopts a pre-classified schema to group them. By selecting a single object, users can obscure groups of related objects based on criteria including privacy sensitivity, object category, or spatial proximity. A technical evaluation shows VisGuardian achieves mAP50 of 0.6704 with only 14.0 ms latency and a 1.7% increase in battery consumption per hour. Furthermore, a user study (N=24) comparing VisGuardian to slider-based and object-based baselines found it to be significantly faster for setting permissions and was preferred by users for its efficiency, effectiveness, and ease of use.
- Open Access
- 13 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-22
Request a Note: How the Request Function Shapes X's Community Notes System
- 13 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-22
Open AccessRequest a Note: How the Request Function Shapes X's Community Notes System
- Open Access
- 13 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-29
PrivWeb: Unobtrusive and Content-aware Privacy Protection For Web Agents
- 13 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-29
Open AccessPrivWeb: Unobtrusive and Content-aware Privacy Protection For Web Agents