About Me

I am a PhD Candidate at the University of Twente, Netherlands, where I am part of the Semantics, Cybersecurity & Services (SCS) group and work on developing attacks and defences on Large Language Models (LLMs).

I completed my B.Tech. (Hons.) and MS by Research in Computer Science and Engineering from IIIT Hyderabad, where I was part of the SyPy Group and advised by Dr. Ankit Gangwal at the Center for Security, Theory & Algorithmic Research (CSTAR).

Previously, I was an Applied AI Researcher at ServiceNow AI Research, where I worked towards improving multi-task performance of LLMs under low resource constraints and benchmarking content moderation in agentic systems.

News

• Sept 2026 Awarded the EMNLP D&I Grant/Award!
• Sept 2026 Selected as a Student Volunteer at EMNLP 2026!
• Aug 2026 Our paper "DARTS: Decoder-Aware Representation Surgery via Tuning for Model Merging" has been accepted at EMNLP 2026 Main Conference! See you at Budapest!
• Jul 2026 Started my PhD journey at the University of Twente!
• Jul 2026 Our paper "Merge Now, Regret Later: The Hidden Cost of Model Merging Is Adversarial Transferability" has been accepted at RAID 2026. See you at Lancaster, UK!
• Feb 2026 Our work "GENIE: Watermarking Graph Neural Networks for Link Prediction" has been accepted at TMLR!
• Jan 2026 Joined ServiceNow AI Research as an Applied AI Researcher!
• Sep 2025 Presented my work at Research Fest 2025 @ IIIT Hyderabad!
• Aug 2025 Poster accepted at AsiaCCS 2025. See you in Hanoi!
• Jun 2024 Our paper "GENIE" is now on arXiv.

Research Interests

My interests lie broadly in the security and safety of machine learning systems.

• Adversarial Machine Learning
• Model Ownership/Watermarking
• Explainable AI (XAI) & Graph Representation Learning
• Security and Safety of Large Language Models (LLMs)
• Privacy in AI

Selected Publications

DARTS: Decoder-Aware Representation Surgery via Tuning for Model Merging

Aaryan Ajay Sharma*, Sai Nishanth Padala*, and Seganrasan Subramanian

Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026)

[arXiv]

Merge Now, Regret Later: The Hidden Cost of Model Merging Is Adversarial Transferability

Mauro Conti^†, Ankit Gangwal, and Aaryan Ajay Sharma*

Proceedings of the 29th International Symposium on Research in Attacks, Intrusions and Defenses (RAID 2026)

[arXiv]

GENIE: Watermarking Graph Neural Networks for Link Prediction

VSP Bachina*, Aaryan Ajay Sharma*, Ankit Gangwal, and Charu Sharma

Transactions on Machine Learning Research (TMLR)

[arXiv] [OpenReview] [Interactive DWT]