A Community Effort to Protect Genomic Data Sharing, Collaboration and Outsourcing

Shuang Wang, Xiaoqian Jiang, Haixu Tang, XiaoFeng Wang, Diyue Bu, Knox Carey, Stephanie O. M. Dyke, Dov Fox, Chao Jiang, Kristin Lauter, Bradley Malin, Heidi J. Sofia, Amalio Telenti, Lei Wang, Wenhao Wang, Lucila Ohno‐MachadoSSRN 2017 · DOI

TL;DR

This paper details the outcomes of the third Critical Assessment of Data Privacy and Protection (CADPP) competition, a collaborative effort to advance privacy-preserving techniques for genomic data. The competition focused on three critical challenges: securing data sharing via Beacons, enabling collaborative discovery across institutions, and outsourcing data to cloud services, yielding significant improvements in privacy and efficiency.

What Problem Does This Paper Address?

The increasing utility of human genomic data in healthcare and research brings immense potential but also significant privacy and security concerns. Genomic data can reveal highly sensitive information, including predispositions to diseases and familial relationships, making it vulnerable to re-identification even when anonymized. Sharing this data, especially on a large scale or with third-party cloud services, necessitates robust privacy-preserving methods to prevent unauthorized disclosure and protect individual and familial privacy. Existing technical solutions often lack direct comparisons in real-world scenarios, making it difficult to assess their practical capabilities and limitations.

Methods

The paper describes the three tracks of the CADPP 2016 competition, each designed to address a specific privacy challenge:

  • Track 1: Beacon Services: Focused on mitigating re-identification risks in genomic data sharing Beacons. Beacons allow users to query for the presence of specific genetic variants. The challenge was to develop methods that allow for a high number of queries while minimizing the risk of an attacker inferring an individual's presence in the dataset. The winning strategy involved a "strategic flipping" method, framed as an optimization problem to balance utility and privacy risk.
  • Track 2: Collaborative Discovery: Aimed at enabling two institutions to jointly analyze genomic data without directly sharing sensitive patient information. This track leveraged secure multiparty computation (SMC) to identify similar genomes across institutions. The core problem was efficiently computing similarity (Levenshtein distance) between genomic sequences in a privacy-preserving manner. The winning solution employed a "reference-based partition strategy" to approximate edit distances.
  • Track 3: Outsourcing to Public Clouds: Focused on securely outsourcing genomic data storage and computation to untrusted cloud environments. This track specifically addressed the challenge of performing genetic testing on homomorphically encrypted data, ensuring that the cloud provider learns nothing about the data, queries, or access patterns. The winning approach utilized "permutation-based cuckoo hashing" to improve the performance of homomorphic comparisons.

Datasets

  • Track 1: A Beacon database of 500 genomes from the 1000 Genomes Project was used, with allele frequencies derived from this project.
  • Track 2: ZNF717 gene sequences (~3470 bps) were used to query databases of 500 patients. This dataset was chosen for its manageable size to allow for evaluation within a few minutes.
  • Track 3: The Charcot-Marie-Tooth disease type 21 was used as a case study. The evaluation involved scenarios with VCF files containing 10,000 and 100,000 records, and up to 50 VCF files.

Key Insights and Ideas

  • Community-driven innovation is crucial: The CADPP competition format effectively brings together diverse expertise (biomedical informatics, computer security, ethics, law) to tackle complex, real-world genomic privacy problems.
  • Balancing privacy and utility is a persistent challenge: While significant progress was made, all tracks highlighted the trade-off between protecting sensitive genomic information and maintaining the utility of the data for analysis.
  • Specific optimizations are key: Highly optimized solutions tailored to specific tracks (e.g., strategic flipping for Beacons, reference-based partitioning for SMC, and permutation-based cuckoo hashing for homomorphic encryption) demonstrated substantial performance gains.
  • Genomic data privacy is a multi-faceted problem: The competition successfully addressed distinct challenges, from interactive querying (Beacons) to collaborative analysis and cloud outsourcing, underscoring the need for a range of privacy-preserving techniques.
  • Real-world applicability is improving: The competition's focus on practical scenarios and benchmarks is helping to bridge the gap between theoretical privacy methods and their adoption in actual biomedical research.

Key Results

  • Track 1: The winning strategy could answer approximately 160,000 queries with an error rate of 0.115, significantly outperforming the baseline and demonstrating a practical approach to securing Beacon services.
  • Track 2: The IBM team achieved the best performance with a runtime under 12 seconds and high accuracy for identifying similar patients, showcasing the feasibility of SMC for cross-institutional genomic analysis.
  • Track 3: The Microsoft team's solution demonstrated impressive speed for homomorphic computation, decryption, and data transfer, enabling genetic testing on 100,000 records in under 4 seconds.
  • Overall, the competition saw multiple orders of magnitude performance improvement in computational runtime and memory requirements compared to previous years, indicating substantial progress in the field.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while impressive, the solutions are often highly optimized for specific competition tracks and may not be universally applicable.
  • For SMC (Track 2), the competition primarily focused on a two-party scenario, and extending solutions to more parties could introduce significant overhead.
  • The paper notes that potential risks due to implementation bugs were not fully assessed.
  • Challenges remain in securely computing exact edit distances over long sequences without approximation.
  • Advanced secure analysis tasks like regression model learning or variant calling over encrypted data were not considered.
  • Future Work:
    • The authors aim to develop more extendable and flexible foundations for tackling emerging genomic privacy challenges.
  • They plan to transition competition outcomes into practice by creating accessible interfaces for integration into existing data-sharing portals.
  • Future competition tasks will focus on more practical biomedical problems, such as machine learning over encrypted data and hardware-based solutions for whole-genome analysis.
  • There is a continued effort to engage geneticists to improve the design and relevance of future competitions.

Summary updated

BibTeX
@misc{wang2017community,
  author = {Shuang Wang and Xiaoqian Jiang and Haixu Tang and XiaoFeng Wang and Diyue Bu and Knox Carey and Stephanie O. M. Dyke and Dov Fox and Chao Jiang and Kristin Lauter and Bradley Malin and Heidi J. Sofia and Amalio Telenti and Lei Wang and Wenhao Wang and Lucila Ohno‐Machado},
  title = {{A Community Effort to Protect Genomic Data Sharing, Collaboration and Outsourcing}},
  year = {2017},
  doi = {10.2139/ssrn.3157968},
  howpublished = {SSRN},
}
Figure