A Windows program is a sequence of bytes with headers, code, imported functions, resources and padding. Arrange those values on a grid and the file becomes an image whose textures reflect repeated instructions, compressed regions and structural layout. The new study uses RGB representations so mature three-channel vision networks can process that structure.[1]
The task is hierarchical. The model first distinguishes benign from malicious files, then predicts a broad malware family, then one of 33 subfamilies. Four transfer-learning backbones were augmented with different attention mechanisms and evaluated under the same taxonomy and preprocessing.[1]
Reported classification levels
| Measure | Value | Context |
|---|---|---|
| Benign vs malware | 99% accuracy | The broad binary task on the study’s held-out data.Source 1 |
| Family attribution | 96–97% accuracy | Classification at the coarser malware-family level.Source 1 |
| Subfamily attribution | 91–95% accuracy | Fine-grained classification across 33 subfamilies.Source 1 |
| Best subfamily macro F1 | Up to 0.89 | Averages performance across classes rather than letting common classes dominate.Source 1 |
Why colour can carry byte structure
An RGB image groups values into three channels, matching the input expected by ResNet, DenseNet, Xception and VGG architectures trained on natural images. Transfer learning does not mean malware resembles a photograph. It means early filters that detect local edges and textures can be repurposed to identify regularities in byte layouts, then fine-tuned for the executable domain.[1]
The hierarchy answers different security questions
| Level | Operational question | Typical consequence of error |
|---|---|---|
| Binary | Is this file likely malicious? | A false negative allows execution; a false positive blocks a legitimate file.Source 1 |
| Family | Which broad code lineage or behaviour group fits? | Analysts may choose the wrong response playbook.Source 1 |
| Subfamily | Which finer variant cluster fits? | Threat intelligence, campaign linkage and forensic attribution become less precise.Source 1 |
Accuracy becomes harder as labels become narrower
Family members share large blocks of code and behaviour. Subfamilies may differ by configuration, packing, loader logic or small revisions. A model can therefore be right that a file is malicious while confusing two closely related variants. The drop from binary to subfamily performance is not a defect hidden by the paper; it is a measure of the information demanded by the label.
Macro F1 matters because malware datasets are uneven. A model can report high overall accuracy by mastering dominant classes while failing rare ones. Per-class precision, recall, F1 and confusion matrices are necessary for a security operations centre that may care most about an uncommon but severe family.[1]
The image omits what execution reveals
Static bytes do not show every runtime behaviour. Malware may unpack itself, fetch code, detect a sandbox, alter control flow or encrypt sections specifically to disrupt static analysis. The model also learns properties of the collection process: compiler versions, packing tools and dataset age can become shortcuts. A random split can place very similar samples on both sides.
The research demonstrates that one network can preserve a useful hierarchy from detection to subfamily attribution. The next test is temporal and adversarial: hold out later campaigns, unseen packers and genuinely new families, then measure calibration and analyst value. The attractive image representation is only useful if it survives the moment malware authors know it is being watched.[1]
Sources & further reading
- Multi-level malware subfamily attribution from RGB executable images using attention-augmented deep learning Scientific Reports · 16 Oct 2026 · Accessed 17 Oct 2026