Minimax Gaussian Mechanisms for Continual Machine Unlearning

Key points

• Machine unlearning updates a trained model after records are deleted, aiming to match exact retraining without repeating the full training procedure.1
• Using Gaussian differential privacy (GDP) and its adaptive composition rule, we show that the full sequence of released models is statistically difficult to distinguish from matched exact retraining.1
• To calibrate these mechanisms for empirical risk minimization, we derive upper bounds on the error of the Newton approximation relative to exact retraining and on how this error changes after each deletion batch.1
• Independent Gaussian noise is calibrated using bounds on the full residual at each release, whereas Gaussian random walk noise uses smaller bounds on residual increments.1

Sources (1)

• [1]Minimax Gaussian Mechanisms for Continual Machine Unlearning
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 8, 10:06 AM

  We develop Gaussian mechanisms for Newton updates under sequential deletion requests.

  Machine unlearning updates a trained model after records are deleted, aiming to match exact retraining without repeating the full training procedure.

Extractive summary: sentences quoted from the sources.