Poster, ICT.OPEN 2024, Utrecht, Netherlands
Poster titled “On Quantifying Data Reconstruction Risk for Multi-server Federated Learning Systems” in track CompSys Research for a Responsibly Digitalised Society.
Talk, Radboud Digital Security group Lunch Talk, Nijmegen, Netherlands
Radboud Digital Security group Lunch Talk provide a forum for sharing research results. Thanks for the invitation of Prof. Dr. H.K. Schraffenberger (Hanna) and Prof. Dr. S. Picek (Stjepan), I shared my work on hindrance of trustworthy distributed machine learning.
Talk, F+Cube 2023, Delft, Netherlands
F+cube encourages and helps female researchers in the STEM area for academic careers. In this talk, I provide my research road map and discuss how female researchers takle with difficulties related to gender.
Talk, ISSRE 2023 WDMD, Florence, Italy
Adversarial behaviors can be launched during different phrases of distributed learning, e.g., Federated Learning, Multi-discriminator GAN. This talk gives a summary on the types of attacks during training and inference phrase, and it introduced four works from us related to it.
Talk, Invited talk, Kaiserslautern, Germany
Talks on how to design effective data-free attacks and corresponding defenses based on synthetic data, in terms of goals, methods, and possible flexible parts for specifc scenarios.
Talk, DSN 2023, Porto, Portugal
Existing untargeted attacks in FL require the adversary owns either a large amount of data or eavesdropping all benign updates. This talk introduces our data-free approach to launch attack with two variants, benefiting from effectiveness and stealthiness.
Talk, PAKDD 2023, Osaka, Japan
Followed by the research of the specific heterogeneous client, Maverick, this talk analyzes its contribution and propose a distance-based client selection method with convergence guarantee.
Talk, AAAI 2022 workshop on FL, Virtual Conference
This talk discusses one important but overlooked type of heterogeneous client, Maverick, which exclusively owns specific data and we see its broad applications such as rare disease databases. And its contribution is also underestimated by existing measurements.
Talk, TPS 2020, Virtual Conference
Vision track on summary of incentives and attacks in Federated Learning. Here we explore how to evaluate the contribution of a client with the presence of both honest and malicious clients.