Adversarial robustness in standalone image recognition

Determine effective techniques to achieve adversarial robustness in standalone image recognition models that reliably prevent adversarial perturbations from causing misclassification, thereby resolving the long-standing challenge of robustness for single-task image classifiers.

Background

The paper notes that despite progress in multimodal settings, achieving adversarial robustness for standalone image classifiers has historically involved steep trade-offs in accuracy, and widely used defenses have struggled to reliably withstand strong attacks. This is highlighted as a broader field-level challenge distinct from the paper’s focus on circuit breakers for generative systems.

The authors emphasize that their circuit-breaking approach provides robustness in multimodal systems but explicitly acknowledge that robust adversarial defenses for standalone image recognition remain unresolved.