Academic Portfolio

Malithi Mithsara

I am a PhD student in Computer Science at Southern Illinois University Carbondale. My research focuses on human activity recognition, wearable AI, multimodal sensor-language learning, large and small language models, and trustworthy AI for healthcare and smart environments.

I am especially interested in secure and privacy-preserving sensing systems, including adversarial attacks, data poisoning detection, contrastive sensor-text alignment, and federated small language models for personalized health support.

Projects

Selected Research

Cross-modal sensor-text attack overview

Breaking Sensor-Text Alignment: Cross-Modal Attack on Contrastive Multimodal Human Activity Recognition

CVPR 2026

This work studies a cross-modal inference-time attack that perturbs text-side embeddings in contrastive multimodal HAR systems, reducing correct sensor-text cosine similarity and causing semantic drift.

Data poisoning attack diagram

Adaptive and Robust Data Poisoning Detection and Sanitization in Wearable IoT Systems using Large Language Models

IEEE Transactions on Human-Machine Systems, under review, 2026

This project explores LLM-assisted detection and sanitization of poisoned activity labels in wearable IoT systems.

LLM-HAR survey framework

A Comprehensive Survey of Large Language Models for Human Activity Recognition: Opportunities, Challenges, and Future Directions

IEEE Transactions on Computational Intelligence, under review, 2026

A survey of how LLMs can support sensor-based HAR through zero/few-shot learning, contextual reasoning, sensor knowledge learning, and deployment strategies.

Federated SLM for women's health framework

Federated Small Language Models for Personalized and Interpretable Women's Health Support on Consumer Devices

IEEE Transactions on Consumer Electronics, under review, 2026

This project uses feature-to-text conversion and federated LoRA fine-tuning of SLMs for personalized menstrual phase prediction and concise health feedback.

Zero-shot data poisoning detection pipeline

Zero-Shot Detection and Sanitization of Data Poisoning Attacks in Wearable AI Using Large Language Models

ICMLA 2024

Uses LLMs to identify and sanitize poisoned labels in wearable human activity recognition data without retraining on large datasets.

LLM-based wearable AI poisoning mitigation poster

Leveraging Large Language Models for Zero-Shot Detection and Mitigation of Data Poisoning in Wearable AI Systems

NeurIPS GenAI for Health Workshop, 2024

Workshop poster presenting LLM-based zero-shot detection and mitigation of data poisoning in wearable AI systems.

Healthcare cyber-physical systems architecture

Large Language Model Integrated Healthcare Cyber-Physical Systems Architecture

COMPSAC 2024

An LLM-integrated healthcare CPS architecture connecting distributed data, medical records, patient portals, and visualization services.

Fall detection and emergency response framework

Intelligent Fall Detection and Emergency Response for Smart Homes Using Language Models

ICMLA 2024

A smart-home fall detection system that integrates sensor-based detection with language-model-supported emergency response.

Research Output

Publications

  1. Breaking Sensor-Text Alignment: Cross-Modal Attack on Contrastive Multimodal Human Activity Recognition. CVPR 2026.
  2. Zero-Shot Detection and Sanitization of Data Poisoning Attacks in Wearable AI Using Large Language Models. ICMLA 2024.
  3. Intelligent Fall Detection and Emergency Response for Smart Homes Using Language Models. ICMLA 2024.
  4. Large Language Model Integrated Healthcare Cyber-Physical Systems Architecture. COMPSAC 2024.
  5. Poster: Leveraging Large Language Models for Zero-Shot Detection and Mitigation of Data Poisoning in Wearable AI Systems. NeurIPS GenAI for Health Workshop 2024.

Contact

Get in Touch

Email: malithi.mithsara@siu.edu

Personal: malithi.mith@gmail.com

Location: Carbondale, Illinois, USA