Fingerprinting LLMs via Prompt Injection ORAL
ACL (Annual Meeting of the Association for Computational Linguistics), 2026
Proposes a prompt injection technique to uniquely fingerprint and track large language models.
Publications
* denotes equal contribution · my name in bold.
Leave My Images Alone: Preventing Multi-Modal Large Language Models from Analyzing Images via Visual Prompt Injection
ACL (Annual Meeting of the Association for Computational Linguistics), 2026
Introduces visual prompt injection methods to protect user images from unauthorized analysis by MLLMs.
Watermark-Based Detection and Attribution of AI-Generated Content
ICLR (International Conference on Learning Representations), 2026
Explores robust watermarking strategies for effectively detecting and attributing AI-generated media.
Jailbreaking Safeguarded Text-to-Image Models via Large Language Models
EACL (Conference of the European Chapter of the ACL), 2026
Leverages LLMs to construct adversarial prompts that bypass safety filters in text-to-image generation models.
A Transfer Attack to Image Watermarks
ICLR (International Conference on Learning Representations), 2025
Demonstrates cross-model transferability of attacks against image watermarks, highlighting fundamental robustness issues.
Periodic Recovery from Poisoning Attacks in Machine Learning
IEEE TDSC (IEEE Transactions on Dependable and Secure Computing), 2025
Proposes a defensive mechanism that periodically recovers ML models to mitigate the impact of data poisoning.
WebInject: Prompt Injection Attack to Web Agents
EMNLP (Empirical Methods in Natural Language Processing), 2025
Exposes prompt injection vulnerabilities in web-browsing agents navigating untrusted environments.
Certifiably Robust Image Watermark
ECCV (European Conference on Computer Vision), 2024
Develops a theoretical framework and practical approach for applying provably robust watermarks to images.
PORE: Provably Robust Recommender Systems Against Data Poisoning Attacks
USENIX Security Symposium, 2023
Introduces a recommender system with formal robustness guarantees against data poisoning and manipulation.
Screening and Diagnosis of Cardiovascular Disease Using AI-Enabled Cardiac MRI
Nature Medicine
Applies deep learning techniques to cardiac MRI for automated and accurate cardiovascular disease diagnosis.
Preprints
MalTool: Malicious Tool Attacks on LLM Agents
Investigates vulnerabilities in LLM agents when interacting with external tools, demonstrating malicious tool-based attacks.
Log-To-Leak: Prompt Injection Attacks on Tool-Using LLM Agents via Model Context Protocol
Highlights security risks in the Model Context Protocol where tool logs can be exploited for prompt injection and data leakage.
SafeText: Safe Text-to-Image Models via Aligning the Text Encoder
Aligns the text encoder to inherently reject unsafe prompts, improving the safety of text-to-image generation.
Stable Signature is Unstable: Removing Image Watermarks from Diffusion Models
Analyzes the vulnerabilities of diffusion model watermarks, showing how they can be effectively removed.