Scalable Malware Image Classification Using Progressive Resizing in Deep Neural Networks

Malware image classification converts binary executables to grayscale images, which allows CNNs to learn structural patterns using visual textures. Nonetheless, downsizing all the samples to a fixed resolution changes spatial properties and harms the ability to separate classes especially when dealing with heterogeneous malware data. To help solve this, we will develop a Progressive Deep Convolutional Neural Network (P-DCNN) that incorporates progressive resizing in training. Malware images are grouped into Executable Size Bands (ESB) which can be explained to be the nearest square sizes calculated using original binary sizes. The model is initially trained on smaller bands to identify rough global structural features and gradually narrowed down on larger bands to identify smaller textures and spatial details. This form of staged learning enhances convergence stability, intrinsic texture distribution, and feature generalization. Also, the band-wise intensity normalization employs the same scale of grayscale at different resolutions. Experiments on the Malimg and Microsoft BIG 2015 Dataset (MMD) exhibit a higher level of accuracy and strength than the traditional fixed-resolution CNN models, which demonstrates the usefulness of progressive resizing to scalable and real-world malware classification.