Title: Circumventing Principle Component Analysis-Based
Malware Detection
Abstract:
Metamorphic malware uses different but functionally equivalent code
between replicates to avoid simple signatures and the classic
dynamic defences for polymorphic code. Recently, several works have
applied machine learning to the metamorphic malware detection
problem. This presentation outlines three contributions to this
area. First, shows that some previous works relied on somewhat
unrealistic test sets when evaluating their results. Second, it
shows that the application of K-nearest neighbour (KNN)
classification can improve the results
obtained from previous principal component analysis (PCA) based
distinguishers. Finally, it gives a simple obfuscator which can fool
a previous PCA distinguisher and the improved PCA-KNN distinguisher.