Quantifying identifiability to choose and audit $ε$ in differentially private deep learning

Differential privacy allows bounding the influence that training data records have on a machine learning model. To use differential privacy in machine learning, data scientists must choose privacy parameters $(\epsilon,\delta)$. Choosing meaningful privacy parameters is key since models trained with weak privacy parameters might result in excessive privacy leakage, while strong privacy parameters might overly degrade model utility. However, privacy parameter values are difficult to choose for two main reasons. First, the upper bound on privacy loss $(\epsilon,\delta)$ might be loose, depending on the chosen sensitivity and data distribution of practical datasets. Second, legal requirements and societal norms for anonymization often refer to individual identifiability, to which $(\epsilon,\delta)$ are only indirectly related. %Prior work has proposed membership inference adversaries to guide the choice of $(\epsilon,\delta)$. However, these adversaries are weaker than the adversary assumed by differential privacy and cannot empirically reach the upper bounds on privacy loss defined by $(\epsilon,\delta)$. Therefore, no quantification of a membership inference attack holds the exact meaning that $(\epsilon,\delta)$ does. We transform $(\epsilon,\delta)$ to a bound on the Bayesian posterior belief of the adversary assumed by differential privacy concerning the presence of any record in the training dataset. The bound holds for multidimensional queries under composition, and we show that it can be tight in practice. Furthermore, we derive an identifiability bound, which relates the adversary assumed in differential privacy to previous work on membership inference adversaries. We formulate an implementation of this differential privacy adversary that allows data scientists to audit model training and compute empirical identifiability scores and empirical $(\epsilon,\delta)$.