About
I’m a Principal Research Scientist at Fujitsu Research of Europe. My work sits at the intersection of AI security and cybersecurity—from computer vision in the early years to today’s language‑centric and distributed AI systems. I design evaluation methods, open datasets, and practical defenses for LLM/LMM safety, agentic RAG systems, and adversarial robustness. Recent projects include inference‑time safety for LLMs, disparity auditing via single‑token probes (TFDP), and secure orchestration for tool‑using agents. I care about results that hold up under scrutiny: measured improvements, reproducible benchmarks, and deployments that make systems safer in practice. Selected work appears at EMNLP 2025 and ACL 2025, with prior contributions at BMVC, ECCV, and ACML, alongside a granted US patent (12183118) and open datasets for red‑teaming and disparity evaluation.
Research Focus
Trustworthy Multimodal Agentic Systems
I build and evaluate agentic systems that reason over text, code, and images—with an emphasis on reliability, safety, and red‑teaming. This includes benchmark design, safety instrumentation, and failure analysis for complex reasoning pipelines.
Decentralized & Collaborative AI
I focus on orchestrated-decentralized frameworks for peer-to-peer LLM federation, enabling collaborative learning and knowledge exchange without raw data sharing—bridging the gap between privacy and performance in distributed systems.
Large Language/Multimodal Model Security
I work on verification and assessment methods for LLMs/LMMs—guardrails, inference‑time safety, and privacy‑aware training in distributed settings—grounded in measured risk and reproducible evaluation.
AI‑Centric Cybersecurity
I bridge conventional security with AI‑specific attack surfaces across pipelines: orchestrator red teaming, retrieval poisoning defenses, and secure toolchains with useful telemetry for operations.
Adversarial Robustness and Metric Learning
Earlier work focused on adversarial robustness in metric learning, with impact on production image retrieval systems. I continue to use those lessons when stress‑testing modern multimodal pipelines.
Algorithmic Disparity Assessment
I study disparities in LLM behavior, including TFDP’s single‑token masked probing and multi‑scale alignment, to quantify differences precisely and efficiently.
Computer Vision Security
Work on biometric systems and physical‑world attacks/defenses laid the groundwork for my approach to robustness and safety in current multimodal systems.
Training Data Attribution
I study attribution and provenance in generative models to help answer practical questions about data influence and copyright in diffusion‑based systems.
Recent Highlights
Selected Publications
Open Datasets & Benchmarks
Collaboration
I enjoy working on grounded problems in AI security—LLM/LMM safety, agentic systems, and applied cybersecurity. If you have a project that needs careful evaluation, targeted red‑teaming, or practical defenses, feel free to reach out.
Key Research Areas for Collaboration
- Large Language/Multimodal Model Safety and Security
- Trustworthy Multimodal Agentic Systems
- Adversarial Robustness in Deep Learning
- Automated Red-teaming and Safety Benchmarking
- AI System Disparity Assessment and Mitigation