Trading inference-time compute for adversarial robustness

OpenAI Blog · 1 year ago

OpenAI researchers showed that applying more computational resources during inference—specifically through multiple verification passes—improves a model's resilience against adversarial attacks. The technique increased robustness from 16% to 51% accuracy on adversarial examples by using five verification passes instead of one. This approach suggests a path toward more reliable AI systems without requiring retraining, though it introduces a latency-accuracy tradeoff.