Targeted Attacks Against the TLSH Similarity Digest Scheme

Similarity Digest Schemes are used in various applications (e.g. digital forensics, spam filtering, malware detection and malware clustering), which require them to be resistant against attacks aiming at generating (A) semantically similar inputs with very different similarity digest values, or (B) completely different inputs with very similar digest values. We show that TLSH, a widely used similarity digest function, is not robust enough against either kinds of attacks. More specifically, we propose automated methods to modify executable software binaries in a way that the modified binary has the exact same functionality as the original one, yet (A) its TLSH difference score from the original version becomes high, or (B) its TLSH digest becomes very similar to another arbitrary TLSH digest up to a complete digest collision. We evaluate our methods on a large data set containing malware binaries, and we also show that they can be used effectively to generate adversarial samples that evade detection by SIMBIoTA, a recently proposed similarity-based malware detection approach.