How Well Can Differential Privacy Be Audited in One Run?

Steinke et al. (2024) have made strides in understanding one-run auditing, a method that improves the efficiency of auditing machine learning algorithms by allowing multiple training examples to be assessed simultaneously. Their research, published on November 13, 2025, demonstrates that one-run auditing can establish a lower bound on the true privacy parameter of the audited algorithm. However, they also highlight a key challenge: interference between the observable effects of different data elements, which can hinder the method's efficacy. To address this, they propose new conceptual approaches aimed at minimizing this interference, thereby enhancing the performance of one-run auditing. This work is particularly relevant in the context of increasing concerns over data privacy in machine learning, as it seeks to refine the tools available for ensuring that algorithms respect user privacy while maintaining their functionality.