Privacy, AI & Cybersecurity

Decision reliability under privacy, trust, and adversarial constraints.

Focus

My AI and cybersecurity work treats reliability as a decision problem. The goal is not only to build accurate models, but to decide whether a model, privacy control, authentication rule, or robustness claim is trustworthy enough for deployment.

Core Themes

• Privacy-preserving learning: homomorphic encryption, federated or distributed learning, gradient leakage, encryption ratios, utility loss, and operational cost.
• Adversarial robustness: stress-testing segmentation and perception models under plausible attack patterns and transferability risk.
• Authentication and trust: RF fingerprinting, known/unknown device decisions, thresholding, rogue transmitter detection, and adversarially generated negative examples.
• AI-system evaluation: model behavior evaluation, prompt optimization, retrieval/tool-use patterns, local model serving, and monitoring.
• Deployment decisions: choosing privacy, robustness, and model controls as governed parameters rather than ad hoc safeguards.

Representative Work

• Secure distributed learning for connected and autonomous vehicles using leveled homomorphic encryption.
• Practical homomorphic encryption for federated learning.
• Adversarial patch transferability across real-time autonomous-vehicle segmentation models.
• RF fingerprinting authentication for IoT networks using Siamese networks.
• Adversarial-resilient RF fingerprinting with CNN-GAN rogue transmitter detection.

Evidence On This Site

• AI for causal inference lectures
• AI-assisted software systems agenda