Game Theoretic Privacy-Preserving Collaborative Data Mining

This proposal focuses on the problem setting where coalition parties, each owning a large set of data, desire to discover new knowledge when they collaborate to jointly process all the datasets; while ensuring that each individual dataset is not revealed to the other parties.

Solutions to this problem are key enablers for ensuring smooth cooperation among parties who do not necessarily trust each other fully. Example situations that reflect this need include coalition forces on military or peace-keeping missions, nations joining forces to detect and prevent terrorism activities, while not willing to reveal their actual intelligence data on national security, and organisations collaboratively analysing consumer behaviour while keeping their customers' profiles private.

The proposed research addresses challenges within the three themes of the DaISy Call, notably those aimed at secure and privacy-preserving collaborative extraction of meaning from data intensive systems comprising different parts of data owned by adaptively changing coalition partners.

We aim to develop novel techniques with guarantees of privacy that will perform data mining even when data are in encrypted form, including the specific tasks of clustering, dependency modelling, classification, regression, and association. The behaviour of such coalition parties involved in performing joint data mining will be analysed using a game theoretic framework, and various innovative collaborative data mining techniques will be developed using this framework while ensuring that privacy is preserved, even when some coalition members may collude. Our techniques will also be designed to be adaptive to and efficient when coalition membership changes.

The proposed project's results will provide solutions to the challenging problem of how to perform data mining jointly on different datasets and involving coalition parties each of which has its own interests and thus may behave selfishly, deviating from normal behaviour and colluding with a subset of other parties.

The main potential beneficiaries are the defence industry and the UK government, as well as the data mining and cloud services industry, and society at large.
It is expected that the impacts described herein will reach the stakeholder groups within 5 to 10 years after the proposed project completion; essentially because major technical results will be published and codes put into the public domain, and the techniques we develop can be implemented in software, and thus do not require additional hardware functionalities to be built into existing deployments of data-intensive systems.

The immediate beneficiary is the Ministry of Defence (MoD) in the military sector and the defence prime contractors. Indeed, our developed game theoretic collaborative data mining techniques can be deployed for strategic and tactical military operations involving different military units of varying levels of trust and even between the UK MoD and military forces of other nations when they are engaged in a coalition for international peace-keeping missions. In such a setting, the MoD can contribute its share of security sensitive military data to perform joint data mining operations with other coalition nations without risking leakage of these data to coalition members. Game theoretic collaborative data mining can unlock the potential for cooperation even among distrusting parties who may share a common goal achievable through joint collaborative data mining operations, but who are unwilling to fully disclose their own datasets to other parties as these are viewed as competitors or may leave the coalition in the future.

For similar reasons, law enforcement and national security organisations have a stake in the results developed from our research. The reason is that the developed techniques will facilitate collaborative extraction of clusters, trends, detection of undesired behaviour and profiling of individuals and events relevant to the interests of national and public safety such as terrorist and criminal activities, while keeping each party's datasets private.

Industries should position themselves for addressing the challenges posed by the increasingly widespread deployment of data outsourcing, clouds and the need for dynamic collaborative operations. In the longer term, our techniques can be applied within the UK government and commercial sectors. More precisely, large sets of data in these sectors are increasingly being hosted in outsourced locations and within the cloud infrastructure. With our techniques, these can be mined even when data are concealed or encrypted independently by different owning parties. Such a feature is ideal if any party is to ever rest assured in the idea of its private data being outsourced.

The fact that data mining can still be performed despite data outsourcing and concealment is expected to lead to the potential market success of industries which offer the services of data outsourcing and cloud computing as this means customers are more willing to subscribe to these services since data can be concealed while being hosted without sacrificing the utility for data mining purposes.

In summary, the above described impacts will offer new business opportunities for wealth creation and maintain the UK's scientific and technological competitiveness in collaborative data mining, secure data outsourcing and cloud services.