The authors present LARAR, an adversarial training method for network intrusion detection that adds layer-by-layer vulnerability scoring and adaptive weighting to standard adversarial training. Instead of treating the whole neural network as a black box, LARAR identifies which layers are most vulnerable to adversarial perturbations (via "auxiliary classifiers" attached to intermediate layers) and focuses defense effort there. On the UNSW-NB15 intrusion-detection dataset, it achieves 95% clean accuracy and improved robustness against FGSM, PGD, and transfer attacks, while reducing computation by targeting vulnerable layers.
Main takeaways:
- Adds layer-wise vulnerability analysis to adversarial training: scores each layer's susceptibility to attacks and adapts defense accordingly.
- Uses "auxiliary classifiers" at intermediate layers to measure where adversarial perturbations propagate most.
- Achieves 95% clean accuracy and better robustness on UNSW-NB15 network intrusion data against FGSM, PGD, and transfer attacks.
- Reduces computational cost by focusing on vulnerable layers and enabling early detection of adversarial samples.
- Provides interpretable vulnerability scores for each layer, not just end-to-end robustness metrics.