Shravya Kanchi
PhD Candidate
Computer Science Department, Virginia Tech
shravya@vt.edu
Google Scholar
GitHub
Resume
AI Security Researcher
I am a final-year PhD candidate at Virginia Tech, Blacksburg, advised by Dr. Daphne Yao, working at the intersection of security and generative AI. My research focuses on applying generative AI to improve real-world cybersecurity outcomes, spanning LLM safety, synthetic data generation, and adversarial ML. I build evaluation-driven systems that improve security robustness in practice and have published in IEEE S&P, Asia CCS, and ACSAC. My work includes hands-on experience with LLMs, GANs, and diffusion models for security-specific data generation and for mitigating toxicity in chatbot customization pipelines.
I am currently seeking industry research or postdoctoral roles starting in Summer 2026 and welcome inquiries via email.
Updates
Research
Research Themes
- Synthetic data pipelines for robust security ML under data scarcity or sensitivity.
- LLM safety and toxicity mitigation for real-world conversational deployments.
- Automated and adversarial security testing frameworks for evolving software and ML threats.
- Evaluation pipelines emphasizing measurable security outcomes at deployment scale.
Representative Projects
Improving Security ML Under Data Scarcity Using Generative AI
Problem: Security ML classifiers are often limited by data scarcity and imbalance, rather than algorithmic design.
Approach: Applied Generative AI-based synthetic data augmentation to improve classifier generalization. Evaluated multiple state-of-the-art GenAI methods and introduced Nimai, a controlled synthetic data generation scheme for security tasks.
Evaluation: Conducted empirical evaluation across 7 diverse security tasks, including severely low-data settings.
Outcome: Achieved up to 32.6% performance improvement in security classifiers.
Reducing Toxicity in LLM Fine-tuning Pipelines
Problem: Fine-tuning LLMs on untrusted conversational data can introduce toxic or unsafe behavior.
Approach: Developed TuneShield, a defense framework for safe fine-tuning of LLaMA-based chat models, combining LLM-based toxicity detection, targeted synthetic data generation ("healing data"), and alignment using PEFT-based Direct Preference Optimization (DPO).
Evaluation: Evaluated against toxicity injection, adaptive adversarial, and jailbreak attacks during conversational fine-tuning.
Outcome: Reduced toxic outputs to near-zero levels while preserving conversational quality.
Agent-based Automation for Security Test Generation
Problem: Manual vulnerability testing for third-party libraries does not scale for software supply-chain security.
Approach: Built a self-correcting multi-agent LLM system for automated JUnit security test generation and execution.
Evaluation: Evaluated across 50 Java client codebases, using LLM-based validation to determine whether generated tests successfully trigger known vulnerabilities.
Outcome: Achieved automated end-to-end vulnerability test generation and validation using self-correcting LLM agents.
Experience
Virginia Polytechnic Institute And State University
Graduate Research Assistant
- Built evaluation-driven AI security systems spanning LLM safety and alignment, synthetic data generation, and adversarial ML, with publications at IEEE S&P, Asia CCS, and ACSAC.
- Designed alignment-aware fine-tuning pipelines for LLaMA-based models using PEFT and Direct Preference Optimization (DPO) to mitigate toxicity injection and preserve conversational quality.
- Developed multi-agent LLM systems that generate, execute, and self-correct JUnit security tests, enabling automated vulnerability validation across 50 Java client codebases in software supply-chain security.
- Created controlled synthetic data generation frameworks that improved ML-based security classifiers by up to 32.6% under severe data scarcity.
JP Morgan Chase & IIIT Hyderabad
Applied Research Collaboration
- Developed the first named-entity-labeled corpus for SEBI regulations, covering 7,500+ sub-regulations and defining 7 domain-specific entity types for financial compliance analysis.
- Built an overlapping Named Entity Recognition system achieving 87.47% precision, enabling large-scale regulatory text understanding for governance and compliance workflows.
Publications
Shravya Kanchi, Neal Mangaokar, Aravind Cheruvu, Sifat Muhammad Abdullah, Shirin Nilizadeh, Atul Prakash, Bimal Viswanath
ACM AsiaCCS 2026, Bangalore, India, 2026.
PDF Code and dataset
Aravind Cheruvu*, Shravya Kanchi*, Sifat Muhammad Abdullah, Nicholas Kong, Daphne Yao, Murtuza Jadliwala, Bimal Viswanath
arXiv preprint, July 2025.
Sifat Muhammad Abdullah, Aravind Cheruvu, Shravya Kanchi, Taejoong Chung, Peng Gao, Murtuza Jadliwala and Bimal Viswanath
IEEE S&P (Oakland) 2024, San Francisco, CA, May 2024.
PDF Code and dataset Video
Aravind Cheruvu, Connor Weeks, Sifat Muhammad Abdullah, Shravya Kanchi, Daphne Yao, and Bimal Viswanath
ACSAC 2023, Austin, Texas, December 2023.
PDF Code and dataset Video
Sathvik Sanjeev Buggana, Deepti Saravanan, Shravya Kanchi, Ujwal Narayan, Shivam Mangale, Lini T. Thomas, Kamalakar Karlapalem, Natraj Raman
WWW Workshop, Lyon, France, April 2022
Shravya Kanchi, Kamalakar Karlapalem
CODASPY, Online, April 2021
PDF Code
Skills
AI Security & Generative AI
AI Security, LLM Safety & Alignment, Generative AI, Adversarial Machine Learning, Security Evaluation & Benchmarking
Model Fine-Tuning & Alignment
Direct Preference Optimization (DPO), PEFT (LoRA), Supervised Fine-Tuning (SFT), Instruction Tuning, Prompt Engineering
Synthetic Data & Robustness
Synthetic Data Generation, Tabular Data Augmentation, Concept Drift, Robustness Analysis, Data Poisoning & Toxicity Injection
Agent-Based Systems
LLM Agents, Multi-Agent Systems, Self-Correcting Agents, Automated Security Test Generation, Agentic Reasoning
Security Domains
Malware & Phishing Detection, Malicious URL Detection, Network Intrusion Detection (IDS), Software Supply-Chain Security, Access Control Systems
ML Frameworks & Tooling
PyTorch, Hugging Face Transformers, LangChain, Evaluation Pipelines
Education
-
Virginia Polytechnic Institute And State University (Virginia Tech)
Ph.D. in Computer Science Aug 2021 - PresentGPA: 3.9/4 -
International Institute of Information Technology (IIIT), Hyderabad
Masters by Thesis in Computer Science Jul 2018 - May 2021GPA: 9/10 -
Indian Institute of Information Technology (IIIT), Sricity
Bachelor of Technology (Honors) in Computer Science Jul 2014 - May 2018GPA: 8.79/10