Effective and privacy-preserving learning for multimedia retrieval

Speaker: Pengfei Zhang
Seminar Date: Mon, 14/12/2020 - 10:30
Venue: 78-631; Data Science MM lab
Host: A/Prof Helen Huang

Seminar Type:  PhD Confirmation Seminar

Abstract: 

In the light of the unprecedented amount of multimedia data with high-dimensional representations nowadays, there exists an urgent demand for effective and efficient searching techniques to analyse and model complex data contents, such as images, texts and videos. To handle it, deep hashing methods have been widely studied and shown promising performance on large-scale retrieval. However, there are two problems need to be considered. First, existing deep hashing methods are usually assumed in fully supervised settings, where human-annotated information is provided. Nevertheless, requiring large-scale data with manual annotations are not realistic in the real world. Second, deep models with remarkable retrieval accuracy pose great threats to data privacy, as they may be used nefariously to conduct a search that violates privacy of individuals. 

To tackle these, we focus our current study on performing effective and privacy-preserving learning for retrieval. We target first at designing a deep hashing model, which can reliably capture the inherent semantic structures of data and accordingly learn discriminative binary representations. Further, we plan to propose a privacy-preserving algorithm for retrieval, which  implements a proactive protection mechanism by transferring original clean data into adversarial data with quasi-imperceptible perturbations before releasing them. Consequently, malicious parties who capture the released data items can not accordingly find out expected information. Our future work involves investigating challenging real-world applications of the proposed algorithm, where data from more sources are involved, including wildlife images, videos, audios.

Biography: 

Pengfei Zhang received his Bachelor and Master degrees in Computer Science from Shandong University, China, in 2015 and 2018, respectively. Currently, he is working towards his PhD degree (supervisors: Dr. Helen Huang and Dr. Hongzhi Yin) at The School of Information Technology and Electrical Engineering, The University of Queensland, Australia. His research interests are multimedia retrieval, optimization, and privacy protection.