RDNAS: Robust Dual-Branch Neural Architecture Search

ICLR 2026 Conference Submission17289 Authors

Keywords: Neural Architecture Search, Dual-Branch Architecture, Adversarial Robustness, Outlier-Aware Shapley Estimator

TL;DR: Robust dual-branch neural architecture search (RDNAS) provides a scalable and ef- fective framework for discovering architectures resilient to adversarial attacks.

Abstract: Deep neural networks have achieved remarkable success but remain highly vulnerable to adversarial perturbations, posing serious challenges in safety-critical applications. We propose **RDNAS**, a robust dual-branch neural architecture search framework that jointly optimizes standard (clean) accuracy and adversarial robustness. RDNAS introduces a dual-branch cell with separate normal and robust pathways, fused via a lightweight attention module to capture complementary representations. To guide architecture search under adversarial training, we develop **ROSE** (Robust Outlier-Aware Shapley Estimator), which integrates adversarial training into the NAS pipeline and improves operation selection through median-of-means smoothing and interquartile-range filtering, mitigating bias under noisy gradient conditions. RDNAS consistently discovers architectures that outperform both hand-crafted networks and state-of-the-art robust NAS baselines across CIFAR-10, CIFAR-100, SVHN, and Tiny-ImageNet. Notably, it achieves 52.6\% $\text{PGD}^{20}$ robustness on CIFAR-10 while maintaining strong clean accuracy. Extensive ablations validate the effectiveness of the dual-branch design, attention-based fusion, and robustness-aware search. RDNAS provides a scalable and effective framework for discovering architectures resilient to adversarial attacks.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 17289