The privacy-enhancing technologies (PETs) prize challenges
The PETs prize challenges are focused on accelerating the adoption and development of an emerging group of data-driven technologies known as PETs. Successful innovators will have their solutions profiled at the second Summit for Democracy, to be convened by President Joe Biden in March 2023.
PETs and federated learning
PETs are a set of novel technologies with the potential to unlock transformative insights from valuable datasets to tackle global societal challenges, while at the same time preserving citizens’ privacy.
The PETs prize challenges focus on federated learning, an approach for training machine learning models on distributed datasets without having to collect the data in a central location. However, the use of federated learning doesn’t currently deliver on end-to-end privacy protections. The challenges aim to encourage innovative federated learning solutions that leverage additional PETs to address these outstanding privacy concerns.
Use cases and their societal and economic impact
The first phase of the challenges finished in November 2022. It involved participants submitting white papers proposing privacy-preserving federated learning solutions for two different use case tracks:
- financial crime prevention (see the ‘find the technical briefs for track one’ link)
- pandemic forecasting (see the ‘find the technical briefs for track two’ link)
These two use cases were chosen for the potential impact that PETs could have on the status quo. For example, international money laundering, which undermines economic prosperity and finances organised crime including human trafficking and terrorism, costs up to US$2 trillion each year according to UN estimates.
Regarding pandemic forecasting use case, bolstering pandemic response capabilities could prevent the huge negative health, economic and socio-political effects of any future pandemic.
Financial crime prevention
Credit: weerapatkiatdumrong, iStock, Getty Images Plus via Getty Images
For the financial crime use case, innovators were provided with synthetic transaction datasets created by Swift, the global provider of secure financial messaging services, in collaboration with:
The aim is to develop solutions that are able to train machine learning models to detect anomalous transactions, while preserving the privacy of individuals’ financial information contained within the datasets.
Pandemic forecasting
Credit: baranozdemir, iStock, Getty Images Plus via Getty Images
For the pandemic forecasting use case, innovators were provided with synthetic datasets containing health and mobility data for a fictitious population undergoing an epidemic. The datasets were created by the University of Virginia Biocomplexity Institute. The aim here is to develop federated learning solutions to accurately forecast an individual’s risk of infection, while providing end-to-end privacy of their highly sensitive health and mobility data. Using such approaches to more accurately model infection risk could enable more effective and timely interventions to be made during a future epidemic or pandemic.
Challenge results to date
Six teams on the UK side of the prize challenges were chosen to receive funding to develop their solutions in phase two of the competition, which will run until February. A further four teams received prizes for their high-scoring white papers. Several other teams that submitted white papers in phase one are continuing to develop their solutions in phase two at their own cost. You can read about some of the insights gained from the proposed solutions from phase one below.
A global collaboration
Credit: franckreporter, iStock, Getty Images Plus via Getty Images
PETs could be especially valuable in allowing for global data collaboration. The PETs prize challenges are similarly a global collaborative project involving partners on both sides of the Atlantic. On the UK side, Innovate UK and CDEI are leading the delivery of the challenges.
At Innovate UK, we work to deliver the government’s vision for the UK to be a global hub for innovation by 2035. CDEI leads the UK government’s work to enable trustworthy innovation using artificial intelligence (AI) and data-driven technologies.
Through this competition we are delivering areas of strategic importance in global opportunities, future economy and responsible innovation working together with our US partners:
- The White House Office of Science and Technology Policy
- US National Institute of Standards and Technology
- US National Science Foundation
In addition, we are grateful for the involvement and support of:
- Swift
- The University of Virginia Biocomplexity Institute
- Financial Conduct Authority
- Information Commissioner’s Office
- Data and Analytics Research Environments UK
- Financial Crimes Enforcement Network
- NHS England
Spotlight on some of our Innovators!
Featurespace
Challenge track: financial crime
Dr Iker Perez, Annegret Funke, Dr David Sutton, Jason Wong from Featurespace. Credit: Featurespace
Meet the team
For nearly a decade, Featurespace has been at the forefront of modernising financial crime prevention in financial services institutions (FIs) around the world. Their next generation machine learning models provide industry leading predictive performance. Privacy preserving technologies allow collaborative learning, which Featurespace believe will be one of the vital leaps forward in the fight against financial crime.
These technologies allow FIs to use information, which was interdicted until now, providing a new source of leverage on the problem. However, widespread adoption is crucial to realising the benefits of collaborative AI.
Summary of approach
Their approach uses deep learning, the successful utilisation of which has in recent years, enabled Featurespace to push the limits of what detection systems can achieve using existing data sources.
Privacy preserving techniques, such as de-identification, local differential privacy, resampling, and k-anonymity are baked into the design for both training and inference stages. This means that all communications between nodes are privacy preserving and the privacy of citizens is never compromised.
Corvus Research
Challenge track: financial crime
Meet the team
Corvus Research is a London-based technology company developing cutting-edge machine learning, alternative data sources, and products for alpha generation within quantitative finance. The team includes experts in Bayesian methods in machine learning, graph theory, privacy, and algorithmic robustness. Corvus Research encountered the market need for privacy preserving solutions for financial datasets within their own development and research.
The PETs prize challenges are the perfect opportunity to test the generalisability of their privacy-enhancing techniques. Corvus Research is collaborating with London-based Haibrid Technologies to co-develop a generalisable, balanced privacy product that trades off accuracy, scalability, and privacy guarantees.
Summary of approach
Their solution proposes an approach to solve a general class of supervised machine learning tasks using a differentially private version of Quantised Langevin Stochastic Dynamics. This approach enables protection of client data through a strong form of differential privacy. The use of an easily tunable parameter also provides an explicit trade-off between privacy, performance and communication cost.
Diagonal
Challenge track: pandemic forecasting
Meet the team
Diagonal is a steward-owned company and a majority-female owned company. As a steward-owned company, they are legally bound to uphold our mission to build responsible technology. They specialise in spatial and sensitive data analysis.
Diagonal build tools to explore and visualise data, with an emphasis on supporting positive changes in cities. They approach projects with openness and transparency and co-create their technology with their clients.
Summary of approach
Diagonal’s approach to the pandemic forecasting use case is guided by their expertise in the innovative use of spatial data, and their mission to develop responsible technology. They are building PETs that avoid the challenges of anonymisation by using data relating to locations visited by an individual.
For this challenge, Diagonal will use a federated learning model to determine personal risk to infection, based on infection risk associated with a location and personal behaviour. Their approach favours domain oversight of the process relying on a curated selection of high-value, aggregate features.
Smaller models can work with these curated selections more effectively than with larger models that have a broader inclusion of data features, which are closer to the sensitive, underlying data. Diagonal will use homomorphic encryption to aggregate features, hiding an individual’s data from their code during aggregation.
Diagonal’s approach stands in contrast to current trends of increasingly large neural networks. Their use of curated features makes the flow of sensitive data explicit and explainable to individuals whose data is used in the model. They believe the use of explainable processes is appropriate for healthcare use cases because it provides an opportunity for informed discussions about how personal data is collected and used. Diagonal will be making the core technology developed in this challenge open source.
Faculty
Challenge track: pandemic forecasting
Faculty founder Marc Warner Credit: Faculty
Meet the team
Faculty is a founder-led company. Since the beginning, their mission has been to bring the benefits of AI to everyone. They started with the Faculty Fellowship, helping PhD students make the transition from academia into a data science career. They have worked on some of the biggest and most difficult challenges faced by major organisations, using ‘decision intelligence’ to help organisations make better decisions on the things that matter.
Summary of approach
One of the main challenges to delivering high performance machine learning is the inability to facilitate model training on sufficiently large scale, real world, datasets. In addition, patients and organisations often have understandable privacy concerns.
Faculty’s federated learning solution will collaboratively train on privatised synthetic representations of individual organisations locally held data. Thus, avoiding General Data Protection Regulation or Health Insurance Portability and Accountability Act (HIPAA) regulatory hurdles and securing sufficiently large-scale data. This solution not only strengthens data privacy through privacy guarantees but grants each contributing organisation the power to control the extent of their privacy versus performance pay offs. It sets the controls independently of both the central organisation and of each other contributing organisation without any raw or synthetic data transfer.
SPACE: fully decentralised distributed learning for trade-off of privacy, accuracy, communication complexity, and efficiency, University of Liverpool
Challenge track: both
Meet the team
The SPACE project team is from the Trustworthy Autonomous Cyber Physical Systems (ACPS) Lab, affiliated with the School of Electrical Engineering, Electronics and Computer Science (EEECS), University of Liverpool. The lab is focusing on safe AI and trustworthy solutions to address the challenges of real-world autonomous cyber-physical systems, including learning-enabled systems, wireless networks, dynamic networks, robotics, and autonomous systems.
The project team is formed of five people:
- Professor Xiaowei Huang
- Dr Yi Dong
- Dr Xingyu Zhao
- Mr Jiaxu Liu
- Mr Yi Qi
Summary of approach
Their solution works to address the issue of privacy leakage in federated learning. Federated learning can help avoid the challenges of centralised training in traditional machine learning, but it is still susceptible to privacy leakage when sensitive information is inferred locally. Proposed solutions to this problem often compromise other properties such as accuracy and communication complexity.
The SPACE team’s solution aims to use a novel hierarchical structure introducing differential privacy across the network to avoid this compromise and allow a balance of:
- privacy
- scalability
- accuracy
- communication complexity
- efficiency
Thank you to all our participants and our winners! Special thanks also to the assessors for their time and input in a competitive judging process. We are looking forward to teams submitting their developed solutions next year and testing their privacy-preserving capabilities in the red teaming phase.
A prize to build our future economy: PETs to inspire, involve and invest
At Innovate UK we have a commitment to:
- inspire: make the opportunity visible and compelling
- involve: bring relevant organisations and people together
- invest: convene the resources needed including our own
As the PETs prize challenges progress over the coming months, we continue to work to help businesses while providing technical solutions that make us all more secure.
Further information
You can connect with Sarah on LinkedIn
Follow Innovate UK on Twitter
Connect with Innovate UK on LinkedIn
Follow Innovate UK on Facebook
You can go to the new Innovate UK website
You can go to the Innovate UK EDGE website
Subscribe to our YouTube channel
Sign up for our email newsletter
Top image: Credit: carloscastilla, iStock, Getty Images Plus via Getty Images
Sarah’s work focuses on how digital ethics can make a direct and positive impact on economic growth, quality of life and the environment.
She is passionate about integrating environmental, social and governance factors into the investment process to achieve sustainable innovation.
Mahi works in the Centre for Data Ethics and Innovation’s projects team supporting the delivery of the UK-US privacy-enhancing technologies (PETs) prize challenges and the broader PETs programme.
Previously she completed a PhD in philosophy of artificial intelligence at the University of Pittsburgh.
Dee’s work is split between the Centre for Data Ethics and Innovation’s central strategy team and the delivery of the UK-US privacy-enhancing technologies prize challenges.