RoME: Robust Mixture of Low-Rank Experts against Multiple Adversarial Perturbations

Official implementation of our ECCV 2026 paper:

RoME: Robust Mixture of Low-Rank Experts against Multiple Adversarial Perturbations

Woo Jae Kim¹, Kyle Min², Suhyeon Ha¹, Joonsung Jeon¹, Sung-eui Yoon¹

¹KAIST, ²Oracle

---

Overview

Multi-perturbation adversarial training (MAT) seeks robustness against multiple adversarial threats at once, but suffers from robustness trade-offs across threats. RoME routes each threat through a distinct model pathway via a Mixture of Experts, using three components:

• Robust low-rank experts — experts are low-rank additive updates over a shared backbone, so the backbone captures threat-common features while experts learn threat-specific ones.
• Threat-distinguishing dual-scale gating — gating fuses local patch-level and global image-level features to better tell threats apart.
• Threat-guided gating diversification — a regularizer that pushes different threats to distinct expert combinations, avoiding threat-agnostic routing.

RoME is modular (plugs into MAT recipes such as MAX and RANDOM), and because threat labels are used only at training time, it also generalizes to unseen threats at inference.

Key results

• Union robustness + natural accuracy. RoME+MAX improves union robustness and natural accuracy over baselines across CIFAR-10, ImageNet-100, and ImageNet-1K datasets.
• Unseen threats. Outperforms prior MAT on common corruptions, $\ell_0$, perceptual, patch, spatial, and semantic attacks, as well as an adaptive gating-misrouting attack.
• Architecture-agnostic. Consistent gains on both Transformer and CNN-based models.

See the paper for full tables and analysis.

Code

🚧 Code will be released upon publication!

Citation

@inproceedings{kim2026rome,
  title     = {RoME: Robust Mixture of Low-Rank Experts against Multiple Adversarial Perturbations},
  author    = {Kim, Woo Jae and Min, Kyle and Ha, Suhyeon and Jeon, Joonsung and Yoon, Sung-eui},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}

Acknowledgements

Low-rank experts build on LoRA. We thank the authors of the MAT baselines and evaluation suites whose code and protocols we build upon.