Malware analysts face challenges related to increasing number of malware variants emerging every
year. Conventional classification of Windows PE32 executables into benign and malicious is no
longer sufficient and needs refinement when it comes to detecting similar functionality malware
samples belonging to the same category. Thus, it is important to explore sources of multiple dynamic
characteristics that can substantially improve similarity-based malware detection through indicators
of compromise from disk, network and memory. The goal of this thesis is to explore a way to
improve multinomial malware classification by exploiting available dynamic characteristics.
In this work dynamic features were extracted with the help of the automated malware analysis
system Cuckoo Sandbox and classified into their ten respective families with the machine learning
library Weka. It has been analysed which dynamic features contribute the most for multinomial
malware classification and what the performance gain is compared to static feature-based malware
classification. An overall classification result of 87.5% could be achieved with the best performing
dynamic features being the modified and opened registry keys, the created and modified files, the
loaded DLLs and the resolved hosts. The best performing classifier was Random Forest. This result,
however, can be improved by adding more dynamic features or combine them with selected static
features in the future.