Facilitating Trust on Data through Provenance

Research on trusted computing focuses mainly on the security and integrity of the execution environment, from hardware components to software services. However, this is only one facet of the computation, the other being the data. If our goal is to produce trusted results, a trustworthy execution environment is not enough: we also need trustworthy data. Provenance of data plays a pivotal role in ascertaining trustworthiness of data. In our work, we explore how to use state-ofthe-art systems techniques to capture and reconstruct provenance, thus enabling us to build trust on both newly generated and existing data. 1.1 Motivation Provenance is a record that describes the sources and agents involved in producing a piece of data [6]. This record can be analyzed e.g. to understand if data conforms designated standards or to calculate a level of trust on the data in order to assist decision making. Thus, knowing the provenance of data can play a central role in the trust we put on them. On the other hand, not having any provenance information on our data could undermine the benefits of using a trustworthy execution environment: if we cannot trust the data we process, we will also be unable to trust the produced results. 1.2 Capturing Provenance Through Dynamic Instrumentation We have developed a new system called DataTracker [7] which uses Dynamic Taint Analysis (DTA) to capture high-fidelity provenance from unmodified programs. DataTracker is based on Intel Pin Dynamic Binary Instrumentation framework and a modified version of the libdft [4] library which provides a reusable framework for Dynamic Taint Analysis. The architecture of DataTracker is depicted in Fig. 1a. Its main components are a Pin tool and a converter written in Python. The former generates provenance information in raw format which are converted to the W3C PROV format [6] by the latter. After converting to PROV, existing tools can be used to further process and visualize the provenance. Fig. 1b shows the provenance graph produced for a simple grep-like utility. DataTracker attributes the output to only two of the four input files, which is 1 Source code available on: http://github.com/m000/dtracker 2 https://software.intel.com/articles/pintool Raw Provenance Information PROV file (Turtle RDF) Application Image Linux Kernel Application Process Raw to PROV Converter file A file B dtracker libdft.a C++ STL Pin VM Pin Instrumentation API Code Cache Syscall/Event Dispatcher