Zero-Day Malware Detection through Unsupervised Deep Learning
Mohammad Mousa, Ayman Mohammad Bahaa-Eldin, Mohamed Sobh, Ayman Taha · 2023
This paper investigates the problem of zero-day malicious software (Malware) detection through unsupervised deep learning. We built a sequence-to-sequence auto-encoder model for learning the behavior of normal software by monitoring the sequence of operating system kernel API calls. After training is completed, the model is used to identify malicious software activities through assigning an anomaly score for each testing sample. The model achieves 90% AUC score compared to 86% AUC score for the previous work done over the same dataset used for performance assessment.