Explainable Federated Learning: Perform fundamental research in data security and write
peer-reviewed scientific papers. Designing new Explainable Federated Learning Technologies.
Mentors:
Yasaman Keshtkarjahromi,
Bhakti Chowkwale.
AI Security: Develop novel privacy attacks against Foundation Models (LLMs, VLMs, Vision
Transformers). Work on differentially private mechanisms to mitigate adversarial privacy risks.
[ICML ’25, AISTATS ’26].
Decentralized Foundation Models: Work on decentralized training and parameter-efficient
fine-tuning of Multimodal Foundation Models in Heterogeneous Environments.
Differentially Private and Fair Deep Learning: Investigate the disparate impact of
differential privacy on the fairness of AI models. Analyze subgroup disparities and identify algorithmic
encoding of protected patients' characteristics through the lens of shortcut learning. Develop
fairness-aware clipping and noise-addition mechanisms for DP-SGD.
Internship, Computer Vision and Algorithmic Team · LUMA Vision GmbH, Germany
2023 - 2024
Deep Learning for Ultrasound Imaging: Build an end-to-end deep learning framework for 4D
intracardiac ultrasound image segmentation. Implement a deep learning-based speckle filtering model and SVD
clutter filtering to enhance image quality. Deploy on a real-world prototype. [IEEE IUS ’24].
Federated Learning on Edge Devices: Develop a federated learning algorithm that decomposes
a large model into an ensemble of lightweight sub-models, enabling clients to train them in parallel across
multiple devices without sharing data via split learning. This approach achieves 4-8× reduction in client
memory usage while preserving privacy and incurring no additional server overhead. [IEEE TNSM ’23].
AI on Edge Devices: Develop face recognition algorithms on edge devices for evaluation in
the National Institute of Standards and Technology
(NIST) Face Recognition Technology Evaluation
Test. Achieve #1 rank in BORDER category among Vietnamese vendors at the time of submission. Develop
real-time people detection algorithms for fisheye cameras and AI-powered traffic cameras for automatic
traffic violations identification. [IEEE AVSS ’21 & ICISN ’21].
