LOAT: Latent-Order Adversarial Training for Efficient and Transferable Robustness

Matthew Fried, Honggang Wang, Hua Fang

Keywords: adversarial training

TL;DR: The paper creates a novel teacher-student model used in adversarial training that captures latent ordering in an adaptive and efficient manner.

Abstract: Adversarial training is computationally prohibitive because projected gradient descent (PGD) is applied uniformly across all samples. Existing efficiency methods focus on “hard” examples but rely on supervised heuristics that fail to capture the emergence of robust representations. We introduce Latent-Order Adversarial Training (LOAT), a fully unsupervised framework that uncovers the latent clustering structure of adversarial dynamics. LOAT learns a multi-view clustering of samples and creates a transition matrix describing how training flows between clusters, enabling cluster-adaptive PGD budgets that allocate computation where it is most effective. The teacher’s discovered structure is transferable and amortized across student models, eliminating repeated profiling costs. On CIFAR-10, LOAT achieves up to 2× higher robustness per PGD call than standard adversarial training and, under matched compute, consistently improves robust accuracy. Ablations confirm that both the clusters and the latent order encode meaningful structure. LOAT shows that exploiting geometric emergent organization enables practical, robust adversarial training under real-world compute constraints.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 16007