Large Language Models have received significant attention due to their abilities to solve a wide range of complex tasks. However these models memorize a significant proportion of their training data, posing a serious threat when disclosed at inference time. To mitigate this unintended memorization, it is crucial to understand what elements are memorized and why. Most existing works provide a posteriori explanations, which has a limited interest in practice. To address this gap, we propose a new approach based on sliced mutual information to detect memorized samples a priori, in a classification setting. It is efficient from the early stages of training, and is readily adaptable to practical scenarios. Our method is supported by new theoretical results that we demonstrate, and requires a low computational budget. We obtain strong empirical results, paving the way for systematic inspection and protection of these vulnerable samples before memorization happens.
翻译:暂无翻译