Research Interests & Projects

Data poisoning attacks — where adversaries corrupt training data or model updates — pose critical threats through availability attacks (degrading accuracy) and backdoor injection (inducing targeted misbehavior). In federated learning, non-IID data, partial participation, and Sybil clients enable stealthy update-level poisoning that evades naive aggregation. My research characterizes these vulnerabilities and develops detection and mitigation strategies to improve robustness and trustworthiness of AI systems.

Federated learning enables IoT devices to collaboratively train intrusion detectors without sharing raw data. My research focuses on metric-driven client sampling — balancing resource availability, data utility, and trust — alongside multi-objective optimization under edge constraints (latency, bandwidth, memory, energy) and communication-efficient robust aggregation protocols.

AI-based IDS methods are designed for adaptive detection of intrusions across heterogeneous IoT network datasets. Key contributions include adaptive neural architecture sizing matched to dataset complexity, cross-dataset knowledge transfer for compatibility across benchmarks, and FL-aware detection to strengthen distributed IDS robustness against evolving DDoS threats.

Building AI systems that are not only accurate but also robust, fair, transparent, and secure in adversarial real-world conditions is a cross-cutting theme of my research. This covers privacy-preserving ML, fairness and accountability in distributed learning, data minimization for GDPR alignment, and interpretable outputs for security operators — underpinning all other research areas.

As AI systems increasingly operate as autonomous agents with minimal human oversight, new security challenges emerge: adversarial manipulation of agent decision-making, prompt injection through environmental inputs, tool misuse, and subverted trust in multi-agent pipelines. This is an emerging research direction motivated by LLM-driven automation in network monitoring, threat response, and incident handling.

Securing resource-constrained edge and IoT environments requires AI models that are lightweight enough for embedded hardware yet robust enough for real-time threat detection. My research develops lightweight neural architectures, model compression techniques, and cross-domain transfer learning across heterogeneous IoT ecosystems — including IIoT, smart grids, and edge computing platforms.