TA.Talal Alrawajfeh
← All research

Published · 2026

Privacy Leakage via Output Label Space and Differentially Private Continual Learning

Identifying a privacy side channel in continual learning and studying mechanisms that restore valid guarantees.

Authors
Marlon Tobaben, Talal Alrawajfeh, Marcus Klasson, Mikko A. Heikkilä, Arno Solin, Antti Honkela
Venue
Transactions on Machine Learning Research
Status
Published
Paper ↗Code ↗

The output label space can reveal private information

  1. Sensitive labelsClasses encountered in a data stream
  2. Output label spaceThe set of labels exposed by the classifier
  3. Privacy protectionProtect label release as well as model learning
When the output label space depends on sensitive data, its release must be included in the privacy analysis.

The paper identifies the output label space of a classifier as a potential privacy side channel when labels depend on sensitive data. It formalizes privacy for continual learning and evaluates mechanisms for releasing labels while preserving valid differential privacy guarantees.

Marlon Tobaben and Talal Alrawajfeh contributed equally.