Projects

A suite of memory-poisoning attacks targeting LLM-based agents with long-term memory, RAG, and tool use. Goes beyond traditional prompt injection to exploit persistent agent memory. Includes SENTINEL, a novel architectural defense, evaluated across clinical, QA, and e-commerce domains with multiple LLM backends.

Red-team framework for stealthy control of backdoors in federated learning. Achieves near-total backdoor eradication in 26 epochs across CIFAR-10, CIFAR-100, and EMNIST while maintaining benign accuracy. Published at SecureComm 2025.

Cross-modal framework for unlearning in GANs focused on human faces. Enables content moderation and privacy using only text prompts, supporting expression and multi-attribute unlearning without annotated data.

State machine-based mutational fuzzing tool for IoT devices. Uses time, context, and location-based mutation with probability-scored seed selection. Discovered buffer overflow, denial of service, and session hijacking vulnerabilities. Published at ICCIS 2018.

Analyzed over 45,000 Ethereum smart contracts for typecasting vulnerabilities and integer manipulation bugs, identifying 411 at-risk contracts. Also categorized 30+ price manipulation attacks in Web3 and implemented real-time detection using Bitquery.

Designed a methodology to detect and recover from GPS spoofing using relative distances between nearby drones and LiDAR sensors. Achieved 76% accuracy with a 2-drone model and 81% with a 3-drone model.

Implemented and extended hands-on security labs covering buffer overflows, Shellshock, Heartbleed, XSS, TCP/IP attacks, race conditions, and firewall bypass. Deployed cloud-based lab environments on AWS for university courses.