arXiv subjects

Hewu Li

Publications and source records attributed to Hewu Li.

At least 37 records · Page 2Linked to original sources

arXiv2025

The Pervasive Blind Spot: Benchmarking VLM Inference Risks on Everyday Personal Videos

The proliferation of Vision-Language Models (VLMs) introduces profound privacy risks from personal videos. This paper addresses the critical yet unexplored inferential privacy threat, the risk of inferring sensitive personal attributes over the data. To address this gap, we crowdsourced a dataset of 508 everyday personal videos from 58 individuals. We then conducted a benchmark study evaluating VLM inference capabilities against human performance. Our findings reveal three critical insights: (1) VLMs possess superhuman inferential capabilities, significantly outperforming human evaluators, leveraging a shift from object recognition to behavioral inference from temporal streams. (2) Inferential risk is strongly correlated with factors such as video characteristics and prompting strategies. (3) VLM-driven explanation towards the inference is unreliable, as we revealed a disconnect between the model-generated explanations and evidential impact, identifying ubiquitous objects as misleading confounders.

cs.HC↗

arXiv2025

Speculating LLMs' Chinese Training Data Pollution from Their Tokens

Tokens are basic elements in the datasets for LLM training. It is well-known that many tokens representing Chinese phrases in the vocabulary of GPT (4o/4o-mini/o1/o3/4.5/4.1/o4-mini) are indicating contents like pornography or online gambling. Based on this observation, our goal is to locate Polluted Chinese (PoC) tokens in LLMs and study the relationship between PoC tokens' existence and training data. (1) We give a formal definition and taxonomy of PoC tokens based on the GPT's vocabulary. (2) We build a PoC token detector via fine-tuning an LLM to label PoC tokens in vocabularies by considering each token's both semantics and related contents from the search engines. (3) We study the speculation on the training data pollution via PoC tokens' appearances (token ID). Experiments on GPT and other 23 LLMs indicate that tokens widely exist while GPT's vocabulary behaves the worst: more than 23% long Chinese tokens (i.e., a token with more than two Chinese characters) are either porn or online gambling. We validate the accuracy of our speculation method on famous pre-training datasets like C4 and Pile. Then, considering GPT-4o, we speculate that the ratio of "Yui Hatano" related webpages in GPT-4o's training data is around 0.5%.

cs.CL↗

arXiv2025

Position: Human-Robot Interaction in Embodied Intelligence Demands a Shift From Static Privacy Controls to Dynamic Learning

The reasoning capabilities of embodied agents introduce a critical, under-explored inferential privacy challenge, where the risk of an agent generate sensitive conclusions from ambient data. This capability creates a fundamental tension between an agent's utility and user privacy, rendering traditional static controls ineffective. To address this, this position paper proposes a framework that reframes privacy as a dynamic learning problem grounded in theory of Contextual Integrity (CI). Our approach enables agents to proactively learn and adapt to individual privacy norms through interaction, outlining a research agenda to develop embodied agents that are both capable and function as trustworthy safeguards of user privacy.

cs.HC↗

arXiv2025

Conflect: Designing Reflective Thinking-Based Contextual Privacy Policy for Mobile Applications

Privacy policies are lengthy and complex, leading to user neglect. While contextual privacy policies (CPPs) present information at the point of risk, they may lack engagement and disrupt tasks. We propose Conflect, an interactive CPP for mobile apps, guided by a reflective thinking framework. Through three workshops with experienced designers and researchers, we constructed the design space of reflective thinking-based CPP design, and identified the disconnect between context and action as the most critical problem. Based on participants' feedback, we designed Conflect to use sidebar alerts, allowing users to reflect on contextualized risks and fostering their control. Our system contextually detects privacy risks, extracts policy segments, and automatically generates risk descriptions with 94.0% policy extraction accuracy on CPP4APP dataset and a 4.35s latency. A user study (N=28) demonstrated that Conflect improves user understanding, trust, and satisfaction while lowering cognitive load compared to CPPs, privacy policies and privacy labels.

cs.HC↗

arXiv2025

PrivWeb: Unobtrusive and Content-aware Privacy Protection For Web Agents

While web agents gained popularity by automating web interactions, their requirement for interface access introduces significant privacy risks that are understudied, particularly from users' perspective. Through a formative study (N=15), we found users frequently misunderstand agents' data practices, and desired unobtrusive, transparent data management. To achieve this, we designed and implemented PrivWeb, a trusted add-on on web agents that utilizes a localized LLM to anonymize private information on interfaces according to user preferences. It features privacy categorization schema and adaptive notifications that selectively pauses tasks for user control over information collection for highly sensitive information, while offering non-disruptive options for less sensitive information, minimizing human oversight. The user study (N=14) across travel, information retrieval, shopping, and entertainment tasks compared PrivWeb with baselines without notification and without control for private information access, where PrivWeb reduced perceived privacy risks with no associated increase in cognitive effort, and resulted in higher overall satisfaction.

cs.HC↗

arXiv2025

Through Their Eyes: User Perceptions on Sensitive Attribute Inference of Social Media Videos by Visual Language Models

The rapid advancement of Visual Language Models (VLMs) has enabled sophisticated analysis of visual content, leading to concerns about the inference of sensitive user attributes and subsequent privacy risks. While technical capabilities of VLMs are increasingly studied, users' understanding, perceptions, and reactions to these inferences remain less explored, especially concerning videos uploaded on the social media. This paper addresses this gap through a semi-structured interview (N=17), investigating user perspectives on VLM-driven sensitive attribute inference from their visual data. Findings reveal that users perceive VLMs as capable of inferring a range of attributes, including location, demographics, and socioeconomic indicators, often with unsettling accuracy. Key concerns include unauthorized identification, misuse of personal information, pervasive surveillance, and harm from inaccurate inferences. Participants reported employing various mitigation strategies, though with skepticism about their ultimate effectiveness against advanced AI. Users also articulate clear expectations for platforms and regulators, emphasizing the need for enhanced transparency, user control, and proactive privacy safeguards. These insights are crucial for guiding the development of responsible AI systems, effective privacy-enhancing technologies, and informed policymaking that aligns with user expectations and societal values.

cs.HC↗

arXiv2025

Understanding Users' Privacy Perceptions Towards LLM's RAG-based Memory

Large Language Models (LLMs) are increasingly integrating memory functionalities to provide personalized and context-aware interactions. However, user understanding, practices and expectations regarding these memory systems are not yet well understood. This paper presents a thematic analysis of semi-structured interviews with 18 users to explore their mental models of LLM's Retrieval Augmented Generation (RAG)-based memory, current usage practices, perceived benefits and drawbacks, privacy concerns and expectations for future memory systems. Our findings reveal diverse and often incomplete mental models of how memory operates. While users appreciate the potential for enhanced personalization and efficiency, significant concerns exist regarding privacy, control and the accuracy of remembered information. Users express a desire for granular control over memory generation, management, usage and updating, including clear mechanisms for reviewing, editing, deleting and categorizing memories, as well as transparent insight into how memories and inferred information are used. We discuss design implications for creating more user-centric, transparent, and trustworthy LLM memory systems.

cs.HC↗

arXiv2025

Towards Aligning Personalized Conversational Recommendation Agents with Users' Privacy Preferences

The proliferation of AI agents, with their complex and context-dependent actions, renders conventional privacy paradigms obsolete. This position paper argues that the current model of privacy management, rooted in a user's unilateral control over a passive tool, is inherently mismatched with the dynamic and interactive nature of AI agents. We contend that ensuring effective privacy protection necessitates that the agents proactively align with users' privacy preferences instead of passively waiting for the user to control. To ground this shift, and using personalized conversational recommendation agents as a case, we propose a conceptual framework built on Contextual Integrity (CI) theory and Privacy Calculus theory. This synthesis first reframes automatically controlling users' privacy as an alignment problem, where AI agents initially did not know users' preferences, and would learn their privacy preferences through implicit or explicit feedback. Upon receiving the preference feedback, the agents used alignment and Pareto optimization for aligning preferences and balancing privacy and utility. We introduced formulations and instantiations, potential applications, as well as five challenges.

cs.HC↗

arXiv2025

Evaluating the Efficacy of Large Language Models for Generating Fine-Grained Visual Privacy Policies in Homes

The proliferation of visual sensors in smart home environments, particularly through wearable devices like smart glasses, introduces profound privacy challenges. Existing privacy controls are often static and coarse-grained, failing to accommodate the dynamic and socially nuanced nature of home environments. This paper investigates the viability of using Large Language Models (LLMs) as the core of a dynamic and adaptive privacy policy engine. We propose a conceptual framework where visual data is classified using a multi-dimensional schema that considers data sensitivity, spatial context, and social presence. An LLM then reasons over this contextual information to enforce fine-grained privacy rules, such as selective object obfuscation, in real-time. Through a comparative evaluation of state-of-the-art Vision Language Models (including GPT-4o and the Qwen-VL series) in simulated home settings , our findings show the feasibility of this approach. The LLM-based engine achieved a top machine-evaluated appropriateness score of 3.99 out of 5, and the policies generated by the models received a top human-evaluated score of 4.00 out of 5.

cs.HC↗

arXiv2025

From Patient Burdens to User Agency: Designing for Real-Time Protection Support in Online Health Consultations

Online medical consultation platforms, while convenient, are undermined by significant privacy risks that erode user trust. We first conducted in-depth semi-structured interviews with 12 users to understand their perceptions of security and privacy landscapes on online medical consultation platforms, as well as their practices, challenges and expectation. Our analysis reveals a critical disconnect between users' desires for anonymity and control, and platform realities that offload the responsibility of ``privacy labor''. To bridge this gap, we present SafeShare, an interaction technique that leverages localized LLM to redact consultations in real-time. SafeShare balances utility and privacy through selectively anonymize private information. A technical evaluation of SafeShare's core PII detection module on 3 dataset demonstrates high efficacy, achieving 89.64\% accuracy with Qwen3-4B on IMCS21 dataset.

cs.HC↗

arXiv2025

The Manipulative Power of Voice Characteristics: Investigating Deceptive Patterns in Mandarin Chinese Female Synthetic Speech

Pervasive voice interaction enables deceptive patterns through subtle voice characteristics, yet empirical investigation into this manipulation lags behind, especially within major non-English language contexts. Addressing this gap, our study presents the first systematic investigation into voice characteristic-based dark patterns employing female synthetic voices in Mandarin Chinese. This focus is crucial given the prevalence of female personas in commercial assistants and the prosodic significance in the Chinese language. Guided by the conceptual framework identifying key influencing factors, we systematically evaluate effectiveness variations by manipulating voice characteristics (five characteristics, three intensities) across different scenarios (shopping vs. question-answering) with different commercial aims. A preliminary study (N=24) validated the experimental materials and the main study (N=36) revealed significant behavioral manipulation (up to +2027.6%). Crucially, the analysis showed that effectiveness varied significantly with voice characteristics and scenario, mediated by user perception (of tone, intonation, timbre) and user demographics (individual preferences, though limited demographic impact). These interconnected findings offer evidence-based insights for ethical design.

cs.HC↗

arXiv2025

Understanding the Dark Side of LLMs' Intrinsic Self-Correction

Intrinsic self-correction was proposed to improve LLMs' responses via feedback prompts solely based on their inherent capability. However, recent works show that LLMs' intrinsic self-correction fails without oracle labels as feedback prompts. In this paper, we aim to interpret LLMs' intrinsic self-correction for different tasks, especially for those failure cases. By including one simple task and three complex tasks with state-of-the-art (SOTA) LLMs like ChatGPT families (o1, 4o, 3.5-turbo) and Llama families (2-7B, 3-8B, and 3.1-8B), we design three interpretation methods to reveal the dark side of LLMs' intrinsic self-correction. We identify intrinsic self-correction can (1) cause LLMs to waver both intermedia and final answers and lead to prompt bias on simple factual questions; (2) introduce human-like cognitive bias on complex tasks. In light of our findings, we also provide two simple yet effective strategies for alleviation: question repeating and supervised fine-tuning with a few samples. We open-source our work at https://x-isc.info/.

cs.CL↗

arXiv2025

Characterizing Unintended Consequences in Human-GUI Agent Collaboration for Web Browsing

The proliferation of Large Language Model (LLM)-based Graphical User Interface (GUI) agents in web browsing scenarios present complex unintended consequences (UCs). This paper characterizes three UCs from three perspectives: phenomena, influence and mitigation, drawing on social media analysis (N=221 posts) and semi-structured interviews (N=14). Key phenomenon for UCs include agents' deficiencies in comprehending instructions and planning tasks, challenges in executing accurate GUI interactions and adapting to dynamic interfaces, the generation of unreliable or misaligned outputs, and shortcomings in error handling and feedback processing. These phenomena manifest as influences from unanticipated actions and user frustration, to privacy violations and security vulnerabilities, and further to eroded trust and wider ethical concerns. Our analysis also identifies user-initiated mitigation, such as technical adjustments and manual oversight, and provides implications for designing future LLM-based GUI agents that are robust, user-centric, and transparent, fostering a crucial balance between automation and human oversight.

cs.HC↗

arXiv2025

An Engorgio Prompt Makes Large Language Model Babble on

Auto-regressive large language models (LLMs) have yielded impressive performance in many real-world tasks. However, the new paradigm of these LLMs also exposes novel threats. In this paper, we explore their vulnerability to inference cost attacks, where a malicious user crafts Engorgio prompts to intentionally increase the computation cost and latency of the inference process. We design Engorgio, a novel methodology, to efficiently generate adversarial Engorgio prompts to affect the target LLM's service availability. Engorgio has the following two technical contributions. (1) We employ a parameterized distribution to track LLMs' prediction trajectory. (2) Targeting the auto-regressive nature of LLMs' inference process, we propose novel loss functions to stably suppress the appearance of the token, whose occurrence will interrupt the LLM's generation process. We conduct extensive experiments on 13 open-sourced LLMs with parameters ranging from 125M to 30B. The results show that Engorgio prompts can successfully induce LLMs to generate abnormally long outputs (i.e., roughly 2-13$\times$ longer to reach 90%+ of the output length limit) in a white-box scenario and our real-world experiment demonstrates Engergio's threat to LLM service with limited computing resources. The code is released at: https://github.com/jianshuod/Engorgio-prompt.

cs.CR↗

arXiv2025

Actual Achieved Gain and Optimal Perceived Gain: Modeling Human Take-over Decisions Towards Automated Vehicles' Suggestions

Driver decision quality in take-overs is critical for effective human-Autonomous Driving System (ADS) collaboration. However, current research lacks detailed analysis of its variations. This paper introduces two metrics--Actual Achieved Gain (AAG) and Optimal Perceived Gain (OPG)--to assess decision quality, with OPG representing optimal decisions and AAG reflecting actual outcomes. Both are calculated as weighted averages of perceived gains and losses, influenced by ADS accuracy. Study 1 (N=315) used a 21-point Thurstone scale to measure perceived gains and losses-key components of AAG and OPG-across typical tasks: route selection, overtaking, and collision avoidance. Studies 2 (N=54) and 3 (N=54) modeled decision quality under varying ADS accuracy and decision time. Results show with sufficient time (>3.5s), AAG converges towards OPG, indicating rational decision-making, while limited time leads to intuitive and deterministic choices. Study 3 also linked AAG-OPG deviations to irrational behaviors. An intervention study (N=8) and a pilot (N=4) employing voice alarms and multi-modal alarms based on these deviations demonstrated AAG's potential to improve decision quality.

cs.HC↗

arXiv2025

Adanonymizer: Interactively Navigating and Balancing the Duality of Privacy and Output Performance in Human-LLM Interaction

Current Large Language Models (LLMs) cannot support users to precisely balance privacy protection and output performance during individual consultations. We introduce Adanonymizer, an anonymization plug-in that allows users to control this balance by navigating a trade-off curve. A survey (N=221) revealed a privacy paradox, where users frequently disclosed sensitive information despite acknowledging privacy risks. The study further demonstrated that privacy risks were not significantly correlated with model output performance, highlighting the potential to navigate this trade-off. Adanonymizer normalizes privacy and utility ratings by type and automates the pseudonymization of sensitive terms based on user preferences, significantly reducing user effort. Its 2D color palette interface visualizes the privacy-utility trade-off, allowing users to adjust the balance by manipulating a point. An evaluation (N=36) compared Adanonymizer with ablation methods and differential privacy techniques, where Adanonymizer significantly reduced modification time, achieved better perceived model performance and overall user preference.

cs.HC↗

arXiv2024

STARVERI: Efficient and Accurate Verification for Risk-Avoidance Routing in LEO Satellite Networks

Emerging satellite Internet constellations such as SpaceX's Starlink will deploy thousands of broadband satellites and construct Low-Earth Orbit(LEO) satellite networks(LSNs) in space, significantly expanding the boundaries of today's terrestrial Internet. However, due to the unique global LEO dynamics, satellite routers will inevitably pass through uncontrolled areas, suffering from security threats. It should be important for satellite network operators(SNOs) to enable verifiable risk-avoidance routing to identify path anomalies. In this paper, we present STARVERI, a novel network path verification framework tailored for emerging LSNs. STARVERI addresses the limitations of existing crypto-based and delay-based verification approaches and accomplishes efficient and accurate path verification by: (i) adopting a dynamic relay selection mechanism deployed in SNO's operation center to judiciously select verifiable relays for each communication pair over LSNs; and (ii) incorporating a lightweight path verification algorithm to dynamically verify each segment path split by distributed relays. We build an LSN simulator based on real constellation information and the results demonstrate that STARVERI can significantly improve the path verification accuracy and achieve lower router overhead compared with existing approaches.

cs.NI↗

arXiv2024

COSMIC: Compress Satellite Images Efficiently via Diffusion Compensation

With the rapidly increasing number of satellites in space and their enhanced capabilities, the amount of earth observation images collected by satellites is exceeding the transmission limits of satellite-to-ground links. Although existing learned image compression solutions achieve remarkable performance by using a sophisticated encoder to extract fruitful features as compression and using a decoder to reconstruct, it is still hard to directly deploy those complex encoders on current satellites' embedded GPUs with limited computing capability and power supply to compress images in orbit. In this paper, we propose COSMIC, a simple yet effective learned compression solution to transmit satellite images. We first design a lightweight encoder (i.e. reducing FLOPs by 2.6~5x) on satellite to achieve a high image compression ratio to save satellite-to-ground links. Then, for reconstructions on the ground, to deal with the feature extraction ability degradation due to simplifying encoders, we propose a diffusion-based model to compensate image details when decoding. Our insight is that satellite's earth observation photos are not just images but indeed multi-modal data with a nature of Text-to-Image pairing since they are collected with rich sensor data (e.g. coordinates, timestamp, etc.) that can be used as the condition for diffusion generation. Extensive experiments show that COSMIC outperforms state-of-the-art baselines on both perceptual and distortion metrics.

eess.IV↗