Dataset distillation for memorized data: Soft labels can leak held-out teacher knowledge

TL;DR

This paper demonstrates that neural network students trained with soft labels from teachers can memorize and transfer specific facts, including held-out data, even without generalization, highlighting risks of information leakage in dataset distillation.

Contribution

It reveals that soft labels can leak memorized teacher knowledge to students, enabling perfect recall of held-out data, and analyzes how this depends on label smoothing temperature.

Findings

Students can memorize held-out data with soft labels.

Memorization persists across architectures and datasets.

Temperature controls the extent of information leakage.

Abstract

Dataset distillation aims to compress training data into fewer examples via a teacher, from which a student can learn effectively. While its success is often attributed to structure in the data, modern neural networks also memorize specific facts, but if and how such memorized information is can transferred in distillation settings remains less understood. In this work, we show that students trained on soft labels from teachers can achieve non-trivial accuracy on held-out memorized data they never directly observed. This effect persists on structured data when the teacher has not generalized.To analyze it in isolation, we consider finite random i.i.d. datasets where generalization is a priori impossible and a successful teacher fit implies pure memorization. Still, students can learn non-trivial information about the held-out data, in some cases up to perfect accuracy. In those settings, enough soft labels are available to recover the teacher functionally - the student matches the teacher's predictions on all possible inputs, including the held-out memorized data. We show that these phenomena strongly depend on the temperature with which the logits are smoothed, but persist across varying network capacities, architectures and dataset compositions.