Do Denoising Autoencoders Improve Image-Based Static Malware Detection Under Synthetic Byte-Level Obfuscation and UPX Packing? A Controlled Evaluation

Image-based static malware detectors can be disturbed by obfuscation, and a denoising autoencoder (DAE) in front of the classifier is often proposed to undo it. We built a controlled evaluation on 17,000 files from the TRAPMINE collection. Six synthetic byte-level transformations and the real packer UPX were applied to both classes, so every file had a known clean version. We compared a DAE with no defence, training on obfuscated examples (augmentation), a blur filter, and an exact inverse, using two CNN classifiers, three random splits, paired bootstrap intervals, and an equivalence margin of ±0.02 in the area under the receiver operating characteristic curve (AUC). Without a defence, the AUC fell from 0.859 to 0.628 for zlib-packed files and from 0.858 to 0.686 for UPX-packed files (CNN-1). Augmentation recovered most of the UPX loss (0.842) but not the zlib loss (0.716). A DAE with a matched classifier was above augmentation by only 0.006 and 0.008 AUC on average over five transformations; the clearest gain was for encryption, a known invertible change (up to +0.029). File size and entropy alone reached an AUC of 0.813, and the CNNs kept 0.795 and 0.824 inside groups of equal size and entropy. For this image construction, dataset, and set of transformations, a DAE is not a general remedy; training on obfuscated examples is a simpler baseline of similar effect.