Temporal Analysis of Distribution Shifts in Malware Classification for Digital Forensics

In recent years, malware diversity and complexity have increased substantially, so the detection and classification of malware families have become one of the key objectives of information security. Machine learning (ML)-based approaches have been proposed to tackle this problem. However, most of these approaches focus on achieving high classification performance scores in static scenarios, without taking into account a key feature of malware: it is constantly evolving. This leads to ML models being outdated and performing poorly after only a few months, leaving stakeholders exposed to potential security risks. With this work, our aim is to highlight the issues that may arise when applying ML-based classification to malware data. We propose a three-step approach to carry out a forensics exploration of model failures. In particular, in the first step, we evaluate and compare the concept drift generated by models trained using a rolling windows approach for selecting the training dataset. In the second step, we evaluate model drift based on the amount of temporal information used in the training dataset. Finally, we perform an in-depth misclassification and feature analysis to emphasize the interpretation of the results and to highlight drift causes. We conclude that caution is warranted when training ML models for malware analysis, as concept drift and clear performance drops were observed even for models trained on larger datasets. Based on our results, it may be more beneficial to train models on fewer but recent data and re-train them after a few months in order to maintain performance