Differential Property Monitoring for Backdoor Detection

A faithful characterization of backdoors is a prerequisite for an effective automated detection. Unfortunately, as we demonstrate, formalization attempts in terms of temporal safety properties prove far from trivial and may involve several revisions. Moreover, given the complexity of the task at hand, a hapless revision of a property may not only eliminate but also introduce inaccuracies in the specification. We introduce a method called differential property monitoring that addresses this challenge by monitoring discrepancies between two versions of a property, and illustrate that this technique can also be used to analyze observations of untrusted components. We demonstrate the utility of the approach using a range of case studies – including the recently discovered xz backdoor.