[2023-03] Anonymous-GAN Framework: A General Purpose Privacy Law Compliant Image Generation Framework using GPT Embedding and Reinforcement Learning - A Case Study in Dry Eye Disease
운영자
날짜 2023.08.28
조회수 131
[2023-03]
Anonymous-GAN Framework: A General Purpose Privacy Law Compliant Image Generation Framework using GPT Embedding and Reinforcement Learning - A Case Study in Dry Eye Disease Bongsang Kim Moon Soul Graduate School of Future Strategy, KAIST Jooyoung Jeon Moon Soul Graduate School of Future Strategy, KAIST
Donghui Lim Moon Soul Graduate School of Future Strategy, KAIST Jiho Cha Moon Soul Graduate School of Future Strategy, KAIST Abstract Achieving accurate and objective diagnoses of Dry Eye Disease, a prevalent ocular condition, using Deep Learning remains a challenge due to the scarcity of training data in real-world medical contexts. While Generative Adversarial Networks (GANs) offer potential for data augmentation, medical privacy law necessitates that GAN-produced images be rigorously de-identified, eliminating any traceability to original sources. In this perspective, we introduce the Anonymous-GAN Framework, a pioneering solution engineered to meet both data augmentation needs and stringent privacy regulations. Distinctively, our framework refrains from altering the core operations of GANs, ensuring its seamless integration in medical applications. Central to our method is the utilization of the GPT's high-dimensional embedding, facilitating a robust de-identification process. Concurrently, reinforcement learning optimizes the seed values in GAN image creation, pinpointing the most conducive segments for efficient de-identification. This approach curtails reliance on intuitive assessments, bolstering the accuracy and reliability of the de-identification task. Our case study substantiates the framework's proficiency with eye imagery, particularly highlighting Dry Eye Disease. Its modular design ensures adaptability with diverse GAN configurations without impinging on performance or mandating extraneous modifications. This adaptability is paramount in medical settings, where a harmonious blend of data propagation, enhancement, and privacy is essential.