Inference: Capturing Provenance Information with Minimal Intrusion

Commercial and government decisions are driven by data. Provenance is the record of how data and processes were created, modified and used. It is used to support quality assessments for data, provide traceability, identify possible system intrusions, etc. Unfortunately, all of the uses of provenance require that provenance information be captured by each system within a system of systems. This "capture problem" is costly and does not scale. To date, only applications that have a high value to scientists have been provenance capture-enabled [9, 17]. Instead, we seek to build observation points external to any pre-built system that will create partial, or inferred, provenance that can be reused across any system that uses the same architectural components.

In order to facilitate the adoption of provenance within enterprise systems built from a heterogeneous software stack that is unique to each organization, the Infer-Proven-ence project is researching the underlying feasibility and creating a toolbox of techniques that will reduce the number of applications that must be provenance-enabled. Unlike a provenance-enabled application that can report observed provenance, inferred provenance has a probability of being what actually happened. Depending on the overall architecture, different provenance inference techniques need to be available. An inference technique that works within a database and its limited set of transformations will not work over streaming data. This work is establishing the theoretical underpinnings for two different provenance-capture inference mechanisms that work within common architectures. It will create implementations of each technique that can be evaluated within real-world scenarios.

The Infer-Proven-ence approach shall be evaluated across two distinct architectures: one for stream processing, and one for data analytics. While architectures exist that combine all of these components, we intentionally split them into the smallest representative unit with respect to data flow and application-type. With this in mind, Infer-Proven-ence will be evaluated across two distinct architectures: a stream processing of sensor data architecture; and a data analytic architecture. Evaluation will consider: ability to correctly infer provenance; accuracy of inferred provenance; cost of implementation within the given architecture, scalability of approach and the utility of the inferred provenance for a use case specific to each problem domain. For the first technique, we will work with partners at Roke Manor Research and their autonomous vehicle program in which data from disparate sensors is streamed through a set of micro-processors and driving decsions are made. Provenance within this use case will be used to highlight anomalies and likely sources of decision errors. For the second technique, we will work within a data analytic architecture in which source data is transformed and manipulated during the process of analysis. Provenance within this use case will be used to reproduce the analytic results

In addition to the real-world evaluation, we shall work closely with UK's Software Sustainability Institute, which promotes sustainable software technologies in order to build software that can be transitioned and reused by others. SSI shall assist in ensuring that Infer-Proven-ence is generalizable and relevant to any discipline based only on the architecture required by that discipline. Finally, Infer-Proven-ence will produce a roadmap for further research, taking stock of the work done and identifying future opportunities.

Infer-Proven-ence also builds partnerships across several institutions including Southampton's Cyber Security Research Centre, the University of Massachusetts Amherst, the Software Sustainability Institute and Roke Manor Research in order to investigate provenance inference in real-world situations.

National Importance
Infer-Proven-ence will facilitate adoption of provenance technologies in non-computational disciplines, such as those within government and commercial organizations that rely on heterogeneous software and lack the computational infrastructure the existing provenance solutions integrate with. By doing this, we will expand the ability to capture provenance for later usage, while minimizing the system impact. This pervasive provenance capture will facilitate: reduced audit costs, reproducible research, highlighting best practices, flagging data anomalies and supporting understanding and trust of data.
These issues are of high importance to academics, industry and the public. In 2009, leading provenance researchers wrote a prospective article extolling all of the uses and future research requirements for provenance after it is widely adopted "10 years in the future", with highlights from the financial sector through scientific research [7]. Today, almost 10 years in the future, very little of those visions have become a reality, mostly because of a slow adoption of provenance by commercial and government organizations. While these organizations are interested in the use and benefits of provenance [6], the high cost of implementing provenance capture within their environments has been prohibitive. The use cases within [6] for using provenance span across themes such as assisting with: context and understanding; curation and reuse; identification of good practice; integrity; interoperability; linking entities; quality; reproducibility; uncertainty. However, there must exist provenance information to work over in order to provide any benefit. While some high-value projects with a very specific focus and minimal technology heterogeneity have end-to-end provenance support [9, 17, 18], it is rare. By providing a mechanism that provides high-quality provenance, without the costly system creation and maintenance costs, we will enable commercial and government agencies to actually use provenance.
Moreover, new technologies, such as autonomous vehicles, require provenance for both debugging, and to establish liability, and have new requirements for capture and storage. In order to facilitate adoption of these new technologies, the provenance component needs to be addressed. Outside of debug mode, there is limited storage in the "automotive black box" should a later investigation be necessary. As such, we cannot spend large quantities of processing power capturing provenance, or large amounts of storage space. Instead, industry needs help identifying the "sweet spot" in which some provenance is captured and stored, and the rest is inferred to reproduce what happened.

Academic Impact
The UK has a strong presence in research in data provenance tools with established teams at, among others, Southampton, KCL, Manchester, Edinburgh, and Newcastle, and track record of EPSRC-funded projects. Infer-Proven-ence will lower the barrier to entry for using provenance in systems that would like provenance information, but find implementation of full provenance capture to be too costly. This will directly benefit other provenance researchers in that it will provide more provenance information to develop their research over. This project also has significant promise to the scientific communities. Many scientific communities, e.g. biologists, geologists, chemists, astrophysicists, have eagerly adopted the use of provenance. However, they are often constrained to particular toolsets if they desire provenance capture. This project will make provenance capture in the scientific domain a less costly and hopefully more commonplace occurrence. This will then feedback to the provenance community by providing the ability to have more provenance across a wider range of systems for their research and experimentation.