Research Projects
In Preparation: CVPR 2027
Hawkeye
A vision-language model fine-tuned to read and reason about illegible text degraded by motion blur, low resolution, occlusion, and glare. We introduce BSTRB, a real-world blurry scene-text benchmark, and Hawkeye4B, a 4B-parameter model that rivals much larger proprietary VLMs at reading noisy in-the-wild text. Contribution: lead author.
Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies
LLM Ads
Embedding personalized advertisements within chatbot responses. We developed a system that generates targeted ads in LLM chatbot conversations and conducted a user study to assess how ad injection impacts trust and response quality. Users struggle to detect chatbot ads, and undisclosed ads are rated more favorably, raising ethical concerns. Contribution: lead author.
34th USENIX Security Symposium (2025)
Cookie Compliance
Analyzing inconsistencies in cookie consent mechanisms on websites across the globe. We developed ConsentChk, an automated system that detects and categorizes violations between a website’s cookie usage and users’ consent preferences. Contribution: measurement design, writing, and analysis of cookie consent discrepancies across 1,793 globally-popular websites.
ArXiv Preprint
AAOS Privacy
Investigating the privacy implications of Android Automotive OS (AAOS) in vehicles. We developed PriDrive, to perform static, dynamic, and network traffic inspection to evaluate the data collected by OEMs and compare it against their privacy policies. Contribution: privacy policy analyzer.
1st Cyber Security in Cars Workshop (CSCS) at CCS
AV Safety/Security
A survey paper examining security and safety challenges in autonomous vehicles (AVs). E.g., AV vulnerabilities, including surveillance risks, sensor system reliability, adversarial attacks, and regulatory concerns. Contribution: writing for surveillance and environmental safety risks.
ArXiv Preprint
Steward: NLP Web
An LLM-powered web automation tool for scalable and efficient website interaction. Steward is an end-to-end system that interprets natural language instructions to navigate websites, automate tasks, and simulate user behavior without manual coding. Contribution: lead author.
32nd USENIX Security Symposium (2023)
Eye-Shield
A novel defense system against shoulder surfing attacks on mobile devices. We designed Eye-Shield, a real-time software that makes on-screen content readable to the user but appear blurry to onlookers. Contribution: lead author.
44th IEEE Symposium on Security and Privacy (2023)
ExtPrivA
An evaluation of inconsistencies between browser extensions’ privacy disclosures and their actual data collection practices. We developed ExtPrivA to detect privacy violations by analyzing privacy policies and tracking data transfers from extensions to servers. Contribution: data collection and evaluation of 47.2k Chrome Web Store extensions finding misleading privacy disclosures.
29th ACM Conference on Computer and Communications Security (2022)
OptOutCheck
A case study analyzing the reliability of opt-out choices provided by online trackers. We developed OptOutCheck to detect inconsistencies between trackers’ stated opt-out policies and their actual data collection practices. Contribution: data collection and evaluation of 2.9k trackers.
32nd USENIX Security Symposium (2023)
Face Obfuscation Fairness
Investigating the demographic fairness of anti face recognition systems. We analyze how these metric embedding networks may exhibit disparities across different demographic groups. Contribution: coded, implemented, and tested theoretical framework proposed by lead author.
17th ACM/IEEE International Conference on Human-Robot Interaction (2022)
Confidant
Exploring privacy management in conversational social robots. We developed CONFIDANT, a privacy controller that leverages various NLP models to analyze conversational metadata. Found that robots equipped with privacy controls are perceived as more trustworthy, privacy-aware, and socially aware. Contribution: lead author.
21st Symposium Privacy Enhancing Technologies (2021)
Face-Off
Protecting user privacy by preventing face recognition. Face-Off introduces perturbations to face images, making them unrecognizable to commercial face recognition models. Contribution: coding, evaluation, and deployment, creating one of the first anti face recognition systems.
ArXiv Preprint
Hierarchical Robustness
A case study and evaluation on how deep neural networks (DNNs) are highly effective but vulnerable to adversarial inputs. Contribution: implemented hierarchical classification approach that leverages invariant features to enhance adversarial robustness without compromising accuracy.