Privacy-preserving LLM adaptation
Studying differentially private training and privacy-utility tradeoffs for compressed and parameter-efficient fine-tuning methods.
Ph.D. Researcher · Trustworthy AI · Privacy & Security
Pradip Kunwar
I design efficient, private, and secure machine learning systems, with a focus on LLM fine-tuning, differential privacy, and adversarial robustness for AI security.
Research focus
Privacy-preserving ML, LLM security, parameter-efficient fine-tuning, Tensor-Train LoRA, and adversarial robustness.
Ph.D.AI and Security, Tennessee Tech University (Aug 2023-Present)
Los Alamos National LabGraduate Research Intern (Dec 2024-Present)
5Recent publications / manuscripts
8+ yrsAI research and product leadership
Research agenda
Building trustworthy AI systems that are efficient, private, and robust.
Efficient fine-tuning
Developing Tensor-Train LoRA and sparse Mixture-of-Experts approaches to reduce fine-tuning cost while preserving model utility.
AI security and robustness
Evaluating membership inference, adversarial malware analysis, and secure AI behavior in high-consequence settings.
Selected publications
Recent work
Privacy Enhanced PEFT: Tensor Train Decomposition Improves Privacy Utility Tradeoffs under DP-SGD
TT-LoRA MoE: Unifying Parameter-Efficient Fine-Tuning and Sparse Mixture-of-Experts
SoK: Leveraging Transformers in Malware Analysis
MalFormer001: Multimodal Transformer Fused Attention based Malware Detector
A Survey on Adversarial Attacks for Malware Analysis
Projects
Research projects
Privacy Analysis of TTLoRA
Formal privacy audit comparing TTLoRA with standard LoRA under DP-aware training and membership inference evaluation.
TTLoRA MoE for Multi-task Learning
Sparse tensor-compressed expert architecture for efficient multi-task adaptation with reduced knowledge interference.
MalFormer001
Multimodal transformer architecture for malware detection using fused attention across image, graph, text, and audio features.
Nepali News Classifier & Summarizer
LSTM-based NLP pipeline for classification and extractive summarization of Nepali news articles.
Experience
Research and applied AI leadership
Graduate Research Intern — Los Alamos National Laboratory
Conducting research on trustworthy AI and robustness.
Graduate Research Assistant — Tennessee Tech University
Researching LLM security, privacy, membership inference attacks, and DP-aware training for compressed architectures.
Senior AI Product Manager / Solutions Lead — Fusemachines
Led AI products and deployments including OCR/NER, CV parsing, semantic scoring, and AI-integrated LMS systems.
Founder — Khozinfo.com
Built a localized digital marketing search engine and backend architecture for client data pipelines.
Achievements & Recognition
RSA Scholar — RSAC 2026
Selected for the RSA Conference Scholar program recognizing promising researchers in security.
ACM SIGHPC Rusty Lusk Scholarship — SC25
Awarded for contributions to high-performance computing and efficient large-scale AI systems.
Workshop Speaker and Organizer 2024
Organized "ChatGPT to ThreatGPT: Vulnerabilities in AI" hands-on workshop at WiCYS conference, 2024
NSF Travel Grant — IEEE SmartComp 2024
Supported participation and presentation at SmartComp conference in Osaka, Japan.
NSF Travel Grant — FLAIRS 2024
Recognized for research contributions in AI and invited to present work.
Texas Instruments Innovation Challenge Winner
Won for developing an innovative embedded systems solution during undergraduate studies.