Reproduce “Malware Detection by Eating a Whole EXE” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Edward Raff, Jon Barker, Jared Sylvester, Robert Brandon, Bryan Catanzaro, Charles Nicholas
MalConv, a convolutional neural network, effectively classifies entire executable files as malicious or benign directly from raw byte sequences, achieving robust generalization and performance while processing inputs exceeding two million time steps. The model’s performance scales with larger datasets, improving generalizability to diverse malware samples.
Key insights
- Malware detection from raw byte sequences of entire executables is feasible and yields high performance and strong generalization, challenging previous limitations of deep learning for extremely long sequence processing.
- Global temporal max-pooling is crucial for handling variable-length binaries and identifying sparsely distributed, highly discriminative malicious features, which are often diluted by averaging operations.
- Batch normalization consistently fails to enable learning in this context, likely due to the highly multi-modal and non-Gaussian distributions of byte activations within binary executables, a key finding for the deep learning community.
- The model’s decisions are interpretable through a sparse-CAM technique, which highlights specific regions within the binary, revealing that MalConv utilizes diverse sections (e.g., .rsrc, .text) beyond just the PE-Header for classification.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.