Lalith Medury

Ph.D. Candidate, Computer Science and Software Engineering · Auburn University
bzm0092@auburn.edu · linkedin.com/in/lmedury

About Me

I am a Ph.D. candidate in Computer Science and Software Engineering at Auburn University, advised by Dr. Farah Kandah in the Cyber Forensics and Emerging Technologies Lab, with an expected graduation in December 2026. My dissertation is titled Privacy against device identification attacks through adversarial traffic injection. Before joining Auburn, I earned a B.Tech. in Computer Science and Engineering from Jawaharlal Nehru Technological University in Hyderabad, India.

What I Do

I work on privacy-preserving and generative machine learning for IoT and networked systems, with an emphasis on device identification, traffic obfuscation, adversarial robustness, and information theory. My work spans:

• Privacy-preserving machine learning methods for IoT device identification and resistance to network-traffic profiling attacks
• Generative traffic-injection and obfuscation techniques that preserve network utility while reducing device-identification accuracy
• GAN- and LLM-driven traffic generation and information-theoretic privacy objectives for network traffic obfuscation
• Full deployment pipelines covering traffic capture, feature engineering, protocol validation, packet crafting, and wireless injection using Python, PyTorch, Scapy, tcpdump, OpenWRT, and monitor-mode Wi-Fi adapters

Alongside research, I work as a Graduate Assistant and Data Analyst at the Auburn University RFID Lab, where I built Python pipelines for ingesting and analyzing over one million RFID and barcode scans, achieving a 400× speedup, and where I manage CI/CD pipelines and production deployments. I previously worked as a software engineer on the Algorand Name Service, a decentralized name service supporting over 17,500 users.

Research

My research agenda, Trustworthy Autonomous AI for Secure Networked and Software Systems, addresses a central question: how can AI systems detect and reduce behavioral leakage while adapting safely to changing systems and threats? Networked systems increasingly reveal what they are doing even when encryption protects the actual content—packet sizes, directions, timing, and communication patterns can expose device states, user actions, and internal workflows to passive adversaries.

My prior work progresses through three stages: determining what an adversary can infer from observable traffic, identifying limitations in existing privacy mechanisms, and building defenses that mimic different device behaviors to confuse adversaries without disrupting legitimate communication.

• GoNP extracts graph-based behavior from UDP application traffic without relying on network addresses, substantially outperforming conventional machine learning baselines under IP-spoofing conditions
• UMIoT, a dynamic per-device multiclassifier, combines behavioral clustering with confidence-based new-device detection and incorporates new devices without retraining a large global model
• ProTOF, a protocol-aware traffic-mixing framework, reduced identification accuracy from 91–95% down to 31–38% across evaluated systems
• DYGANO and GANFUSION use WGAN-GP-based generative models to inject protocol-compliant synthetic traffic until adversarial accuracy falls below a target privacy threshold, achieving 98.5% and 99.6% packet correctness respectively in live testbed evaluation
• CAMEO, my current work, uses neural device embeddings and CTGAN-based synthetic traffic to select behaviorally dissimilar traffic patterns for adaptive mixing

My ongoing work treats packet sequences as a "language" for Transformer-based modeling and formalizes leakage as mutual information between traffic patterns and device identities. I also use Fano's inequality to establish classifier-agnostic lower bounds on adversarial device-identification error.

Looking forward, my research agenda expands into three connected thrusts: measuring exploitable behavioral leakage across AI, HPC, and cyber-physical systems; creating adaptive AI agents that reduce leakage while preserving system performance; and ensuring these agents remain safe and reliable as threats evolve.

Teaching

Selected Publications

• Medury, L., Robinson, L. & Kandah, F. (2025). Unmasking IoT Devices: A Dynamic and Adaptive Classification Approach. IEEE Internet of Things Journal, 12(22), 46878–46888.
• Medury, L., Robinson, L. & Kandah, F. (2025). GANFUSION: GAN-Fused Synthetic Injection for Obfuscating Network Traffic Analysis. IEEE 13th Conference on Communications and Network Security (CNS).
• Medury, L., Robinson, L. & Kandah, F. (2025). A Dynamic GAN-Based Obfuscation Approach Against Profiling Attacks. IEEE 50th Conference on Local Computer Networks (LCN).
• Medury, L., Kothapalli, U. K. & Kandah, F. (2025). Towards Enhancing Device Anonymity and Classification Resistance in IoT. IEEE International Conference on Communications Workshops (ICC Workshops).
• Medury, L. & Kandah, F. (2024). GoNP: Graph of Network Patterns for Device Identification using UDP Application Layer Protocols. IEEE 49th Conference on Local Computer Networks (LCN).
• Medury, L. & Kandah, F. (2024). B2-C2: Blockchain-based Flow Control Consistency for Multi-Controller SDN Architecture. IEEE International Conference on Consumer Electronics.

Additional journal publications appear in IEEE Access, the Journal of King Saud University Computer and Information Sciences, and the IEEE Internet of Things Journal. I currently have manuscripts under review at the IEEE Internet of Things Journal and IEEE Transactions on Consumer Electronics.

Service and Awards

I serve as a peer reviewer for Artificial Intelligence Review (Springer Nature), IEEE Access (11 completed reviews), and The Journal of Supercomputing (Springer). I am a recipient of the Charles E. Gavin Fellowship Award and won Best AI Application built using Cloudflare at Auburn Hacks.

Academic Positions