EfficientNet-based Convolutional Neural Networks for Malware Classification
Vasundhara Acharya, Vinayakumar Ravi, Nazeeruddin Mohammad · 2021
Malicious software is posing a great challenge to Cybersecurity by causing huge damage to enterprises and users. Traditional signature-based techniques have found it challenging to detect malware as it has changed and evolved over time. Machine learning techniques have become popular for malware detection but they are based on heuristic feature engineering, which is expensive and requires domain expertise. In the recent years, deep learning, are being actively investigated for this crucial malware detection application to achieve robust solutions with higher accuracy. To assist in achieving this goal, the EfficientNet-B1 model pretrained on the ImageNet dataset with the last layer adapted to attain malware family classification is presented. The malware samples to be inputted are represented as byteplot grayscale images. The experimental results on a dataset comprising of 10,868 samples from 9 different malware families showed that the proposed approach can be effectively used to classify malware families. Classification accuracy of 98.57% on the held-out test set was obtained by the proposed model which outperformed the other pretrained deep learning models.