More about HKUST
Improving Privacy and Efficiency of Federation Knowledge Distillation
PhD Thesis Proposal Defence
Title: "Improving Privacy and Efficiency of Federation Knowledge
Distillation"
by
Mr. Sheng WAN
Abstract:
Growing concerns over data privacy and security have led to increasingly
stringent regulations on data sharing and protection. Federated Learning (FL)
has emerged as a promising paradigm that enables collaborative model training
without exchanging raw data. Federated Knowledge Distillation (FedKD) further
extends FL by transferring knowledge through logits rather than model
parameters, thereby improving communication efficiency and naturally
supporting heterogeneous model architectures. However, the transmission of
logits introduces new challenges in terms of communication efficiency and
privacy protection.
This thesis aims to improve the privacy and efficiency of FedKD while
preserving model utility under diverse federated learning scenarios. First,
we address the challenge of achieving efficient and privacy-preserving
knowledge transfer in cross-silo FL. We propose FedPDD, a privacy-preserving
double distillation framework that enables local models to learn from both
the ensemble knowledge of collaborating models and historical local
predictions. By adopting an offline training strategy and exchanging only
model outputs, FedPDD significantly reduces communication overhead while
reducing privacy risks. Differential privacy is further incorporated into the
communication process to provide formal privacy guarantees.
Second, we investigate the security of FedKD against malicious knowledge
manipulation in cross-device FL. Existing Byzantine-robust FL methods mainly
focus on parameter aggregation and are not directly applicable to knowledge
distillation. To address this limitation, we identify previously unexplored
vulnerabilities in logit-sharing FL and propose two Byzantine attack
strategies, namely the Top-k Attack and the Impersonation Attack, which
specifically target transmitted distilled knowledge. To mitigate these
threats, we develop Federated Top Guard Distillation (FedTGD), a robust
knowledge aggregation mechanism that effectively filters malicious knowledge
while maintaining model utility.
Third, we uncover a previously overlooked privacy vulnerability in
logit-sharing FL. We demonstrate that a semi-honest server can reconstruct
clients' private models from shared logits and propose the Adaptive Model
Stealing Attack (AdaMSA) to quantify this privacy risk. Furthermore, we
establish theoretical bounds on this vulnerability and introduce a
lightweight defense strategy that perturbs transmitted logits to minimize
privacy leakage while preserving training performance.
Overall, this thesis advances Federated Knowledge Distillation from two
complementary perspectives: improving communication efficiency and
strengthening privacy protection. Through efficient cross-silo knowledge
transfer, robust defense against malicious knowledge manipulation and
systematic analysis and mitigation of privacy leakage in logit-sharing FL,
this thesis provides practical algorithms and theoretical insights toward
building efficient, privacy-preserving and trustworthy FedKD systems for
real-world federated learning applications.
Date: Wednesday, 12 August 2026
Time: 4:40pm - 6:30pm
Venue: Room 3494
Lifts 25/26
Committee Members: Prof. Kai Chen (Supervisor)
Prof. Qiang Yang (Co-Supervisor)
Prof. Qian Zhang (Chairperson)
Dr. Dan Xu