Single AI Model Achieves Robustness Across All Threat Levels

Adversarially robust AI models typically require specialized training for specific attack budgets, leading to a proliferation of models for diverse threat environments. This paper addresses the challenge of achieving "one-for-all" robustness across a continuous spectrum of threat levels within a single model. The proposed solution is the Threat Conditional Network (TCN), which employs a representation factorization framework. TCN decomposes representation learning into a shared, threat-invariant backbone and a lightweight, threat-conditional adaptor. This adaptor conditions the single model on the perturbation level using Fourier-based embeddings and channel-wise affine modulation. By training against a distribution of perturbation budgets, TCN can adapt flexibly and seamlessly to an infinite continuum of threat levels during inference. Extensive experiments on standard datasets demonstrate that TCN matches or exceeds the performance of an entire ensemble of budget-specialized models, generalizes to unseen perturbation budgets, and maintains robust transfer under mismatched threat conditions, all with only a 4.6% parameter overhead.