Analysis of neural network detectors for network attacks.

This article focuses on developing a novel method to generate stealthy network attacks that can evade neural network (NN)-based intrusion detection systems (NIDS) under the de facto standard threat model, where attackers can only modify network packets before transmission. Existing adversarial example generation methods fail to produce malicious packets that both compromise target machines and evade NN-based detectors due to protocol constraints and limited attacker control. The authors propose Protocol-Constraint-AWare (PCAW) adversarial examples, which respect network protocol rules while misleading detection models, and demonstrate their approach using Address Resolution Protocol (ARP) poisoning and Domain Name System (DNS) cache poisoning attacks. Their evaluation shows that PCAW adversarial examples can successfully launch stealthy attacks, though challenges remain in multi-packet scenarios requiring mitigation strategies such as dummy attacker packets and session-based adversarial value set databases. The paper also discusses limitations, potential countermeasures like DeepCloak, and situates the work within existing research on adversarial attacks in network security.