Secure and Privacy-preserving Genome-wide and Phenome-wide Association Studies via Intel Software Guard Extensions (SGX)

随着生物医学研究中收集的数据量（如人类基因组数据）的快速增长，

With the rapid growth of the data volume (e.g., human genomic data) collected in biomedical research, data protection, in particular for patients’ privacy in secondary uses of these data, has attracted much attention recently. Today, a vast majority of sensitive biomedical data, including individual human genomic data and their associated health metadata, are shared only through controlled-access databases (e.g. dbGaP) and biomedical researchers are required to sign a user agreement before getting access to these data. Security research has already produced a suite of techniques that can serve the general purpose of privacy-preserving computation; their direct applications are, however, too expensive (in terms of resource consumption) for real-world biomedical applications. An alternative solution is hardware-assisted Trusted Execution Environment (TEE) solutions developed or being developed by both hardware vendors (Intel, AMD, ARM) and the open-source research community. A prominent example is Intel’s Software Guard Extension (SGX), which is available as a feature in Intel's mainstream CPUs (i.e., Skylake and Kaby Lake). In this project, we plan to explore potential applications of TEE to two popular genome computation tasks involving sensitive biomedical data, i.e., the genome-wide and phenome-wide association studies. For GWAS, a secondary research user may collect genomic sequences (in encrypted form) with (cases) or without (controls) a disease phenotype from multiple data owners, on which association tests or advanced GWAS algorithms can be conducted within the SGX enclave. Similarly, for PheWAS, a user may collect phenotype data from individuals whose genomes containing (case) or not containing (control) one or more specific variations. We will address two issues when developing these approaches: 1) we will customize GWAS/PheWAS algorithms for efficient execution in the TEE with limited resources (e.g, memory, I/O, etc), and 2) we will develop new genome computing outsourcing and data sharing platforms suing the SGX techniques, and further understand and mitigate its potential side-channel risks with regards to GWAS/PheWAS computing tasks. The proposed research will lead to a practical solution for secure GWAS and PheWAS in three application scenarios: 1) secure outsourcing: a research institution collects matched genomic and phenotypic data from a large cohort of case and control individuals, and outsources the storage of these data and potential repeated GWAS and PheWAS computation to a public or commercial cloud; 2) secure collaboration: a consortium of researchers across multiple institutions attempt to collaborate on a large GWAS/PheWAS study using the data collected by each participating institution; and 3) secure data sharing: researchers want to share their data with a broad biomedical research community so that potential data users may conduct a secondary GWAS/PheWAS analysis.