Data Poisoning Attacks to Graph Metrics Under Local Differential Privacy

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering

Final Year Thesis Oral Defense

Title: "Data Poisoning Attacks to Graph Metrics Under Local Differential 
        Privacy"

by

HAU Yiu Tong

Abstract:

Graph data are prominently featured in a large variety of applications in the 
current age of BIg Data. For the analysis of such data, graph metrics such as 
Clustering Coefficient provide key insights into the structure and properties 
of graphs. Local Differential Privacy (LDP) describes a standard by which an 
algorithm can provide a quantifiable privacy guarantee to users. When applied 
to a graph database, LDP prevents third-party observers from inferring the 
existence of an edge or a node in the graph with high confidence, while 
removing the need for a trusted central party. Although LDP frameworks for 
graphs can produce accurate estimates for graph metrics without compromising 
users' privacy, they are still susceptible to data poisoning attacks which 
manipulate graph metrics via data injection. These attacks reduce the utility 
and adaptation of graph LDP frameworks. In this work, data poisoning attacks 
for graph metrics are proposed and evaluated on real graph datasets. The 
findings of this work can be applied to improve existing graph database systems 
under LDP.


Date            : 2 May 2024 (Thursday)

Time            : 14:30 - 15:10

Venue           : Room 5510 (near lifts 25/26), HKUST

Advisor         : Prof. ZHOU Xiaofang

2nd Reader      : Prof. WONG Raymond Chi-Wing