AGAT: adversarial training with sample selection via enhanced local stability
Start here
The paper proposes the selective adversarial training framework AGAT: a lightweight score ALS measures the change in logits between an input and its label-preserving augmented variant, selecting neighborhood-unstable samples, and adversarial supervision is applied to only 20-30% of samples per batch; on MNIST, CIFAR-10, CIFAR-100 and Tiny-ImageNet-200, robust accuracy is comparable or higher, and training time is lower than all adversarial training baselines.
Generated by AI from reports · Updated 17 days ago
Attention and developments
Attention, hourly
Now 20, 0% in 24 hours·Peak 20 (Nov 6 03:00)
Related research
This event's paper
- Augmentation-Guided Adversarial Training via Local Stability-Based Sample Selection
arXiv · 2026-10-23 · Deep reading
Progress timeline
- arXiv: Adversarial Attacks, Backdoors & PoisoningData augmentation-guided adversarial training based on local-stability sample selection (AGAT)
The paper proposes a selective adversarial training framework, AGAT, which applies adversarial supervision to only 20-30% of training samples per batch, with comparable or higher robust accuracy. It uses a lightweight score, ALS, to measure the normalized change in logits between an input and its label-preserving augmented variant, selecting neighborhood-unstable samples. On MNIST, CIFAR-10, CIFAR-100 and Tiny-ImageNet-200, its training time is lower than all adversarial training baselines.