Modern AI systems are getting more powerful and more centralized at the same time. That creates real problems: privacy risks, single points of failure, and data that organizations cannot share. I build distributed AI frameworks that let multiple parties train models together without exposing raw data. My work covers healthcare, cybersecurity, and environmental monitoring.
Hospitals, banks, IoT networks, and mobile devices all hold sensitive data that AI needs. The problem is that privacy laws, competition, and security risks make it impossible to pool that data in one place. I build tools that let AI models train across separate data silos without moving or exposing any raw data.
I also study the security of decentralized systems themselves. Smart contracts and federated learning protocols have real vulnerabilities that attackers exploit. I build detection and defense tools grounded in representation learning and multimodal analysis to protect these systems.
Both research areas share a core principle: test under realistic adversarial conditions, not just clean benchmarks. Systems built with a clear threat model in mind are more secure and generalize better. My results consistently show this.
Applying federated learning to large language and vision-language models, where communication costs and data heterogeneity are far greater than in standard FL.
Building tighter, composable privacy accounting methods that give real guarantees for deployed federated systems without the accuracy loss of standard DP-SGD.
Combining static analysis, fuzzing, and learned representations to build audit pipelines that catch new vulnerability classes as Solidity and the EVM evolve.
Using federated multi-source remote sensing for climate and biodiversity monitoring, enabling global environmental models without centralizing sovereign satellite data.