Work Experience

University of California, Irvine

Visiting Researcher, California, USA, 2022

Provide information-theoretic upper bounds on the amount of information that the local model update reveals about any single user’s dataset in a federated learning system.
Conduct various state-of-the-art model pruning schemes in Federated Learning under various privacy attacks. These experimental evaluations quantify the exact amount of privacy loss in concrete settings and demonstrate alignment with theoretical bounds.
Design a defence mechanism by applying a personalised defence mask and adapting the defence pruning rate, so as to jointly optimise model accuracy, privacy, and efficiency for low-resource environments.

Research Assistant, Madrid, Spain, 2020

Apply Federated Learning in centralised training tasks, ensuring maintained accuracy.
Design robust aggregation methods that take advantage of reputation models and superior voting elections to fight against poisoning attacks.
Conduct various state-of-the-art Byzantine-robust aggregations and poisoning attacks in Federated Learning to evaluate and compare the performance of our algorithms.
Implement the proposed method into a browser extension and validated its performance through real users’ tasks.
Integrate the algorithms into Acuratio’s Multicloud Federated Learning Platform, enabling training on real user data within an industrial setting.

Research Assistant, Shenzhen, China, 2017

Design and execute a gait experiment at the Luohu Geriatric Hospital in Shenzhen, utilising wearable sensors and camera technology to gather data.
Employ a real-time approach for detecting the 2D pose of multiple individuals in gait experiment videos, resulting in the generation of over 14,000 images.
Implement various machine learning algorithms to classify the actions of elderly individuals, subsequently comparing their performance.