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

2025-06-17

AI Summary

This work investigates whether teacher models’ sample-specific memorization in dataset distillation can be inadvertently leaked to student models via soft labels. Method: We employ temperature-scaled logit smoothing, comparative evaluation across networks of varying capacity, and functional equivalence analysis. Contribution/Results: We establish— for the first time—that even when a teacher model is trained exclusively on unstructured, i.i.d. data and overfits completely (i.e., perfectly memorizes training samples), its soft labels alone enable a student model to achieve 100% prediction accuracy on held-out memorized samples. This phenomenon is independent of data structure or task semantics, and persists robustly across diverse network capacities and architectures. Our findings demonstrate that soft-label distillation can functionally replicate the teacher’s behavior, challenging conventional assumptions about the informational limits of knowledge distillation. This reveals critical implications for model memorization, knowledge leakage, and data privacy in deep learning systems.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Computer Vision: Diffusion Models for Vision
• Natural Language Processing: Sentence-level Semantics, Textual Inference, etc.

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Security and Privacy: Data transparency and provenance

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.

Problem

• How soft labels transfer memorized teacher knowledge to students
• Impact of temperature on accuracy in dataset distillation
• Student learning from held-out data via teacher soft labels

Innovation

• Soft labels transfer memorized teacher knowledge
• Students learn held-out data via distillation
• Temperature affects soft label effectiveness