Twin Auto-Encoder Model for Learning Separable Representation in Cyberattack Detection