Researcher in Federated Learning and Trustworthy AI (PhD)
https://victorobarafor.com/
- Federated learning under non-IID data
- Training instability and failure modes
- Robust aggregation and optimization
- Personalization under heterogeneity
- Geometry of distributed updates
Federated learning systems are typically studied under idealized conditions.
In practice, heterogeneity, drift, and conflicting client updates introduce instability that standard methods do not address.
My work focuses on a central question:
When and why does federated learning fail under real-world conditions?
The goal is to characterize these failure modes and design methods that remain stable in realistic distributed environments.
This work is organized around three directions:
Robust Federated Learning (Non-IID)
Aggregation under distribution shift and client heterogeneity.
Personalization under Heterogeneity
Client-specific adaptation and its effect on global performance.
Geometry and Instability in Federated LoRA
Training dynamics through update alignment, conflict, and early predictors of failure.
-
robust-federated-learning-noniid
Robust aggregation under heterogeneous client distributions -
federated-personalization-depth
Client-specific adaptation in federated systems -
federated-lora-geometry
Geometry and instability in federated LoRA
Many approaches optimize for average-case performance.
In realistic settings, systems fail due to:
- conflicting updates
- distributional imbalance
- unstable optimization dynamics
Understanding these behaviors requires focusing on failure modes, not just performance.
- Stability-aware aggregation
- Early indicators of training collapse
- Scaling under increasing heterogeneity
- Geometry-informed optimization
- Website: https://victorobarafor.com/