Towards a Unifying Framework for Tuning Analysis Precision by Program Transformation

Static and dynamic program analyses attempt to extract useful information on program’s behaviours. Static analysis uses an abstract model of programs to reason on their runtime behaviour without actually running them, while dynamic analysis reasons on a test set of real program executions. For this reason, the precision of static analysis is limited by the presence of false positives (executions allowed by the abstract model that cannot happen at runtime), while the precision of dynamic analysis is limited by the presence of false negatives (real executions that are not in the test set). Researchers have developed many analysis techniques and tools in the attempt to increase the precision of program verification. Software protection is an interesting scenario where programs need to be protected from adversaries that use program analysis to understand their inner working and then exploit this knowledge to perform some illicit actions. Program analysis plays a dual role in program verification and software protection: in program verification we want the analysis to be as precise as possible, while in software protection we want to degrade the results of analysis as much as possible. Indeed, in software protection researchers usually recur to a special class of program transformations, called code obfuscation, to modify a program in order to make it more difficult to analyse while preserving its intended functionality. In this setting, it is interesting to study how program transformations that preserve the intended behaviour of programs can affect the precision of both static and dynamic analysis. While some works have been done in order to formalise the efficiency of code obfuscation in degrading static analysis and in the possibility of transforming programs in order to avoid or increase false positives, less attention has been posed to formalise the relation between program transformations and false negatives in dynamic analysis. In this work we are setting the scene for a formal investigation of the syntactic and semantic program features that affect the presence of false negatives in dynamic analysis. We believe that this understanding would be useful for improving the precision of existing dynamic analysis tools and in the design of program transformations that complicate the dynamic analysis