Documented Data Dispersal

We define the Documented Data Dispersal (DDD) problem as determining, given inputs, outputs, and an application connecting them, whether or not a certain input was used to produce a certain output. There exist many possible ways to address this problem. All of the solutions can be characterized by the level of program analysis, instrumentation, and data analysis that they perform. These parameters determine the tradeoffs between risk and performance that a particular solution makes. We first evaluate similarity metrics as a method of addressing the DDD problem. This method cannot detect all potential data dispersal but has characteristics that make it ideally suited for some domains, such as plagiarism detectors. Next, we evaluate binary rewriting by instrumenting common functions that perform data movement, thereby tracking memory buffers that contain data from a particular input file. This method is very fast; it introduces almost no overhead. Unfortunately, it also misses significant amounts of data movement. Finally, we evaluate a binary interpreter as a means to dynamically identify instrumentation points. This method incurs significantly more overhead than the first two but promises to mitigate more risk as well Analyzing the parameters described above is a powerful tool that can be used to determine the risk/performance tradeoff that a method makes. Using this tool, we argue that future development of methods to address the DDD problem should start from a position of no risk and terrible performance and explore the performance enhancements that can be made while minimizing the risk. This bottom-up approach appears to have promise.