Towards Privacy-preserving Artificial Intelligence (PAI) for Healthcare and Bioinformatics
This thesis presents innovative privacy-preserving AI solutions for healthcare and bioinformatics, including federated learning for multi-omics, data deletion from deep learning models, and on-device medical analysis with vision LLMs, advancing secure and ethical AI development.
Overview
The rapid evolution of artificial intelligence (AI), from machine learning (ML) and deep learning (DL) to large language models (LLMs) and the anticipated emergence of artificial general intelligence (AGI), has underscored privacy as a critical concern across all stages of AI development. Nowhere is this issue more pressing than in healthcare and bioinformatics, where sensitive patient data must be safeguarded while enabling AI-driven innovation. Privacy risks arise at multiple levels, including data collection, storage, transmission, model training, and inference, necessitating robust privacy-preserving solutions.
This thesis presents a comprehensive investigation into privacy-preserving artificial intelligence (PAI) for healthcare and bioinformatics, systematically addressing challenges and developing novel methodologies to mitigate risks. We introduce PPML-Omics, a pioneering federated learning framework designed for secure multi-omics analysis by designing a decentralized differentially private (DP) algorithm to ensure data confidentiality across decentralized biomedical datasets. As deep learning models may inadvertently memorize and retain sensitive information, we propose AFS, an innovative software that facilitates the efficient removal of private user data from pre-trained DL models, thereby reinforcing individuals’ rights to control their personal information. Extending privacy-preserving AI to the domain of LLMs, we present SkinGPT-4, a vision LLM-powered dermatological assistant that enables on-device medical analysis, ensuring user privacy by eliminating reliance on centralized servers for health-related interactions.
Through these contributions, this research advances the frontiers of privacy-preserving AI, establishing a foundation for secure, ethical, and practical AI applications in healthcare and bioinformatics. By addressing critical privacy vulnerabilities and proposing cutting-edge solutions, this work paves the way for a future where AI can empower medical advancements while upholding data security and individual rights.
Presenters
Brief Biography
Juexiao Zhou is a PhD candidate at King Abdullah University of Science and Technology (KAUST), Saudi Arabia, under the supervision of Professor Xin Gao. He is also the co-founder and Chief AI Scientist at DermAssure.ai, MOSS.ai, and BeautyX.ai. His research lies at the intersection of computer science and biomedicine, with a primary focus on AI-driven intelligent healthcare, bioinformatics, and ethical and trustworthy AI in healthcare. Juexiao develops cutting-edge deep learning models and large language models (LLMs) to enable disease detection, prognosis, and risk assessment across clinical settings. In the domain of bioinformatics, he builds intelligent computational frameworks to decode gene regulatory networks, predict protein structure and function, and model complex biological systems. His recent research also explores curiosity-driven AI agents as autonomous scientific researchers. He is committed to advancing ethical AI in healthcare, tackling key challenges such as data privacy, bias, fairness, security, toxicity, and the broader implications of emerging Artificial General Intelligence (AGI) in clinical practice. Juexiao has authored over 30 publications in top-tier journals and conferences, including Science Advances, Nature Machine Intelligence, Nature Computational Science, Nature Communications, The Lancet, Genome Research, Trends in Genetics, Bioinformatics, IEEE TMI, and MICCAI. His work has been featured by major media outlets such as Arab News, Radio Television Hong Kong (RTHK), and Inside Precision Medicine. He is an active member of CAAI, APBioNET, and GBD. He serves as a reviewer for leading journals and conferences, including Nature, Nature Methods, Nature Communications, Medical Image Analysis, Genome Biology, Genome Research, NeurIPS, SIGKDD, and MICCAI. He is also an editorial board member of BMC Bioinformatics, a guest editor for Biomedical Informatics, and currently serves as co-chair of the IS-HIS 2025 Symposium at ICCNS 2025 in Varna, Bulgaria.