هجمات PLAA على مستوى البلاستيك تجنب NIDS بنسبة 92.78٪ من النجاح

92.78% of adversarial network traffic generated by PLAA evades detection by current NIDS models — and it does so without turning the malicious payload into garbage. That number comes from a new packet-level attack algorithm designed specifically for network traffic, not borrowed from computer vision.

Why CV-Style Attacks Fail on Network Traffic

Most adversarial attacks on NIDS treat network flows like images: perturb a flow-level feature vector and call it done. The result is traffic that either doesn't route properly (invalid packets) or loses the original attack semantics — the malicious intent gets scrambled. PLAA's authors argue this is a fundamental mismatch. Network traffic has structural constraints: packet headers must be valid, payloads must be executable, and the sequence matters. Perturbing a flow-level feature ignores all of that.

PLAA Builds Traffic Packet by Packet While Preserving Semantics

PLAA works incrementally at the packet level. Instead of starting with a target flow feature and working backward, it generates packets one by one. At each stage a semantic monitor checks whether the generated traffic still carries the original attack's intent — say, a SQL injection or a DoS payload. If the semantics degrade, the generation path is adjusted. This is the key difference: the attack respects the traffic's structure, so the resulting adversarial packets are both valid and meaningful.

The algorithm was evaluated against current NIDS models on three datasets: CIC-UNSW-NB15, CIC-DDoS2019, and CIC-IDS-2017. Across all three, PLAA averaged a 92.78% evasion success rate. That is not cherry-picked on one dataset; the performance holds across different attack types and traffic profiles.

What This Means for Network Defenders

NIDS operators have long assumed that adversarial attacks from the CV world were a theoretical concern — too brittle to work in practice. PLAA shows that is false. A packet-level generation strategy that respects protocol constraints makes evasion practical. I expect to see red teams incorporating this approach, and defenders will need to move beyond flow-level feature analysis. Packet-level inspection and adversarial training on realistic traffic are no longer optional.

PLAA exposes a critical blind spot in NIDS architectures that rely on flow-level features — defenders will need to rethink how they analyze traffic at the packet level to close this gap.

Source: PLAA: Packet-level Adversarial Attacks in Network Traffic Detection
Domain: arxiv.org