Composition Challenges for Automated Software Diversity


Over the past 20 years, a variety of automated software diversity techniques have been proposed. Some techniques randomize aspects of the implementation that are left undefined by the source language specification, such as code layout, stack layout, or locations of heap-allocated objects. Others insert instrumentation or obfuscation that is transparent from an application perspective, e.g. using XOR masks to obscure data values in memory or hiding code pointers using jump tables. A common assumption is that layering these techniques improves security due to increased entropy in the resulting binary. In this paper we examine this assumption and show that it fails to hold in general. In particular, it fails in one of the strongest deployment models for software diversity—that of multiple diverse variants running together in a multi-variant execution environment (MVEE) where attacks manifest as detectable behavioral divergence. We present several examples of diversity combinations that are vulnerable to attack in an MVEE even when none of the component techniques are vulnerable in isolation. Based on these results, we present guidance on which techniques do combine well and suggestions for effective deployment of diversity in MVEEs.

Presented at the Layered Assurance Workshop (LAW), 2016



  author      = {Ben Davis and
                 Per Larsen and
                 Stijn Volckadert and
                 Simon Winwood and
                 David Melski and
                 Michael Franz and
                 Stephen Magill
  institution = {Galois, Inc.},
  title       = {Composition Challenges for Automated Software Diversity},
  year        = {2018},
  note        = {Available at \url{}}}