Introduction

We generate genetic information for research, clinical care and personal curiosity at exponential rates. Sharing these genetic datasets is vital for accelerating the pace of biomedical discoveries and for fully realizing the promises of the genetic revolution. However, one of the key issues of broad dissemination of genetic data is finding an adequate balance that ensures data privacy. I will present several strategies to breach genetic privacy using open internet tools, including a systematic analysis of the strategy that implicated the Golden State Killer. Our analyses show that these strategies can identify major parts of the US population from their allegedly anonymous genetic information by anyone in the world. I will conclude my talk with practical suggestions to reconcile genetic privacy with the need to share genetic information.

Recent advances in biotechnology and medicine allow us to collect an immense amount of functional genomics data at the personalized and population level. This surge in data gives rise to a paradigm shift in biology and medicine towards data intensive discoveries. While this provides the perfect opportunity to study human genetics and disease, it also presents daunting challenges in maintaining the genomic privacy of study participants and patients. I will discuss the scalable tools we have developed to overcome privacy concerns associated with sharing and analyzing functional genomics data. These tools are based on privacy risk assessment through linking attacks, machine learning, and data sanitization grounded in privacy and utility.

Electronic health records (EHR) are often siloed across a network of hospitals, presenting a challenge for researchers who wish to perform aggregate count queries on these records—e.g. How many patients have diabetes? While many works in the computational biology community—and the CS community at large—have addressed this problem (by using sketching algorithms like HyperLogLog, or deliberately injecting noise such as in differential privacy), all existing practical approaches to computing an aggregated count across more than two parties reveal some amount of private information. It has remained an open question whether there is a practical approximation algorithm which provably does not reveal any private information other than the aggregated value itself. Here, we resolve this fundamental open question by presenting what is to our knowledge the first practical algorithm to compute an approximate federated aggregate count without revealing any private data not contained in the count itself. In addition to being theoretically sound, the algorithm is practical; the code we provide handles up to 8 parties using a single CPU thread in less than 1 minute and with an expected approximation error of around 6%. Furthermore, the protocol is parallelizable across cores, which can further reduce computation time enabling interactive usage.

Large-scale biomedical data are being collected by many research and medical institutions as well as commercial entities. However, due to privacy concerns and strict data-sharing regulations, these data are often not shared and remain in silos. This limitation hinders biomedical research as access to large and diverse datasets is becoming increasingly important for discovering new scientific and clinical insights. In this talk, I will introduce a new framework for privacy-preserving federated analysis of biomedical data based on multiparty homomorphic encryption. I will explain how our framework overcomes privacy concerns by leveraging state-of-the-art cryptographic techniques to ensure end-to-end data confidentiality and to enable fine-grained access control to the analysis results. Contrary to alternative approaches, our solution does not introduce noise in the computation for privacy protection, and it enables a large number of entities to collaborate while locally keeping their private data. I will demonstrate our framework’s applicability by replicating, in a federated and privacy-preserving manner, a range of essential biomedical analysis tasks, including genome-wide association studies, Kaplan-Meier survival analysis, and principal component analysis. Our framework has the potential to accelerate biomedical research by enabling privacy-preserving access to siloed data and unlocking new collaborative studies.

The amount of human genomic data generated and shared continues to grow in a variety of sectors, including health care, biomedical research and the direct-to-consumer (or recreational genomics) industry. At the same time, so too do the legal and technical challenges associated with the protection of personal rights. In this presentation, I will discuss some of the recent advancements in policy in the US and abroad, as well as new computational mechanisms for the collection, analysis, and disclosure of genomic data, and the knowledge derived from it. Along the way, I will illustrate the relationship, as well as tension, that between social and technical components of the problem and offer illustrations of where they are working together to create new data governance solutions.

Policies and technology to address privacy concerns often fail to recognize that multiple factors are involved in someone’s decision to share data or keep these data private. While policies that accommodate the majority are reasonable, it may also be possible to obtain from individuals or whole groups their preferences for privacy and to honor these preferences in real settings. I will discuss my experience with different types of data sharing systems, stakeholders, and technology in the context of sharing consented and “de-identified” data in various contexts.