The field of machine learning is moving towards developing more efficient and principled defenses against adversarial attacks, with a focus on improving feature representations and uncertainty calibration. Researchers are exploring new frameworks that strategically prepend an empirical risk minimimization phase to conventional adversarial training, resulting in more effective robustness acquisition and improved clean accuracy. Additionally, there is a growing emphasis on reliable perturbation-based explanations, with studies investigating the relationship between uncertainty calibration and explanation quality. Another area of focus is controlling the patterns a model learns to prevent reliance on irrelevant or misleading features, with proposed solutions including robust feature attribution methods that optimize for explanation robustness and mitigation of shortcut learning. Noteworthy papers include:

• Ignition Phase, which introduces Adversarial Evolution Training, a simple yet powerful framework that achieves comparable or superior robustness more rapidly and improves clean accuracy.
• Model Guidance via Robust Feature Attribution, which proposes a simplified objective that consistently reduces test-time misclassifications by 20% compared to state-of-the-art methods.