Divide and Conquer: Introducing Partial Multi-Variant Execution

Abstract

After several decades of defensive research against the exploitation of memory errors, a wide range of techniques has been proposed, yet no silver bullet has been found. Multi-Variant eXecution (MVX) is one promising proposal for defending against a wide range of known and potentially unknown attacks. MVX systems run multiple program variants in parallel on the same inputs while monitoring their behavior and deduplicating their outputs. By constructing these program variants using automated software diversity techniques, we can ensure that the variants behave identically under normal operating conditions but diverge when attacked. The MVX system detects these divergences and reacts appropriately.State-of-the-art MVX systems have several fundamental problems that inhibit their real-world adoption. First, they often require full source code availability to construct variants and eliminate non-deterministic program behavior. Second, they incur significant resource overhead that linearly increases with the number of variants running in parallel.We propose Partial Multi-Variant eXecution (PMVX), a technique that can mitigate these problems by limiting the scope of MVX to certain well-delineated parts of a target application and by running the rest of the application in Single-Variant eXecution (SVX) mode. PMVX relaxes the source code availability requirement of traditional MVX systems and yields substantially reduced resource consumption while maintaining the strong security guarantees of these systems. However, PMVX implementations must address the non-trivial problem of ensuring all variants are in equivalent states whenever they switch from SVX to MVX mode.We designed and implemented a proof-of-concept PMVX system called FORTDIVIDE that solves this state-equivalency problem using state migration and resynchronization. We thoroughly evaluated the security and performance of our system as a whole, and of our state migration and synchro-nization mechanisms in isolation. We conclude that PMVX has great potential but needs to be applied with the utmost care since the added overhead of state resynchronization can quickly outweigh the benefits of running in SVX mode.