Deep Learning Classification Methods Applied to Tabular Cybersecurity Benchmarks

This research recasts the network attack dataset from UNSW-NB15 as an intrusion detection problem in image space. Using one-hot-encodings, resulting grayscale thumbnails provide a quarter-million examples for deep learning algorithms. Applying MobileNetV2’s convolutional neural architecture, work demonstrates 97% accuracy distinguishing normal and traffic. Further class refinements to 9 individual families (exploits, worms, shellcodes) show overall 54% accuracy. feature importance rank, random forest solution on subsets shows most important source-destination factors least ones mainly obscure protocols. It further extends classification other cybersecurity benchmarks such malware signatures extracted binary headers, with 80% detect computer viruses portable executable files (headers only). Both novel datasets are available community Kaggle.