SoK: Systematic analysis of adversarial threats against deep learning approaches for autonomous anomaly detection systems in SDN-IoT networks

Adversarial attacks can severely degrade AI-based security systems: The paper found that adversarial attacks can reduce detection accuracy in network security systems by up to 48.4%, meaning these AI systems designed to catch threats can be manipulated into missing real attacks.

Multiple attack strategies pose different risks: The research identified three categories of attacks (data-level, model-level, and hybrid), with Membership Inference attacks causing the most significant accuracy drops. This means defenders need to protect against diverse threat vectors, not just one type of attack.

Defenses exist but create practical challenges: While adversarial training can improve system robustness, the paper notes it requires "high computational overhead," making it difficult to deploy these protections in real-time network environments where speed is critical.

Current security systems lack systematic vulnerability assessment: The paper addresses a gap in existing research by providing the first systematic analysis of how adversarial attacks specifically threaten AI-based anomaly detection in IoT and software-defined networks, suggesting previous security evaluations were incomplete.

Explainable AI and adaptive defenses are needed: The authors recommend moving beyond traditional defenses toward "explainable AI-driven security frameworks" and real-time mitigation strategies, suggesting that making AI systems more transparent and adaptable is essential for trustworthy network security.