Adversarial Attack Research

AI Security & Robustness

Empirical research evaluating the impact of realistic domain constraints on adversarial evasion attacks against machine learning-based Network Intrusion Detection Systems (NIDS).

Python Adversarial Robustness Toolbox (ART) PyTorch / Scikit-learn CIC-IDS2017 Network Security Adversarial Machine Learning Pandas / NumPy

1. Problem Statement & Threat Model

While adversarial perturbation methods demonstrate near-perfect evasion rates in computer vision, their direct application to network intrusion detection frequently yields syntactically invalid or non-functional network traffic. In real-world environments, an attacker cannot arbitrarily mutate immutable packet headers, discrete protocols, or monotonic counters without breaking transport semantics. This research examines how enforcing strict networking domain constraints limits adversarial success and alters detection boundaries.

2. Methodology & Implementation

  • Dataset & Feature Pipelines: Processed and standardized flow-based features from the CIC-IDS2017 benchmark, isolating attack classes (such as DoS, PortScan, and Brute Force) against benign baseline traffic.
  • Baseline Detection Classifiers: Trained and calibrated ML/DL architectures on flow records to simulate enterprise-grade network anomaly classifiers.
  • Evasion Attacks via ART: Implemented gradient- and optimization-based attack formulations (including FGSM, BIM, JSMA, and Carlini & Wagner) using the Adversarial Robustness Toolbox (ART).
  • Constraint Enforcement Engine: Built masking layers that bound perturbations exclusively to non-critical, mutable flow attributes (e.g., packet inter-arrival jitter, padding bytes) while enforcing strict value ranges and preserving packet validity.
  • Robustness & Attack Degradation Analysis: Quantified the drop in evasion effectiveness (Attack Success Rate vs. Perturbation Budget $\epsilon$) when shifting from unconstrained mathematical space to realistic domain-constrained conditions.
Key Takeaway: Enforcing strict network domain constraints significantly restricts the viable search space for adversarial perturbations, proving that domain-aware feature constraints are an effective first layer of defense for ML-driven intrusion detection.

Scroll naar boven