人臉影像隱私保護之多任務生成對抗網路設計

Abstract

本研究主要因應近年由於消費型鏡頭與計算硬體的蓬勃發展與進步，使得擷取與儲存大量影像資料資料變得非常容易。設想一個資料擁有者，如醫院或是政府機構，是很容易針對個人進行資訊蒐集、儲存與處理。在這樣的情況下，要如何確保資料擁有者可以精準的將每份可識別個人身份 (Identify) 的資料做到保護與個人資料隱藏 (Information Hiding)以達到隱私保護 (Privacy Protection) 的目的；同時，在針對個人資訊完成去識別化與資訊隱藏後，仍能保留相當程度之資料堪用性以利後續特定目的之運用成為機敏資料分析與處理領域的核心研究課題。所以本研究提出了Multi-Task Generative Adversarial Network : Multi-Task GAN網路架構，透過對偶推論 (Dual Inference) 與Rate-Distortion理論，來設計Multi-Task GAN幾項校正損失函數。本研究也提出了對偶Multi-Task GAN網路架構設計完整的物理意義來進行相關理論分析，並搭配多種的實驗來驗證所宣稱的效果。

This study aims to address the ease of capturing and storing large amounts of image data due to the flourishing development and progress of consumer-grade lenses and computing hardware in recent years. As a result, for data owners such as hospitals or government agencies, it has become easy to collect, store, and process personal information. In this context, how to ensure that data owners can accurately protect and hide personally identifiable information (Identity) to achieve privacy protection and information hiding while still retaining a reasonable level of data utility for sensitive data analysis and processing has become a core research issue in the field of sensitive data analysis and processing. Therefore, this study proposes a Multi-Task Generative Adversarial Network (Multi-Task GAN) network architecture, which uses dual inference and rate-distortion theory to correct several loss functions of Multi-Task GAN. The study also presents a complete physical meaning to conduct a theoretical analysis of our system(such as Mutual Information) and validates the claimed effects with relevant experiments.