Supplementary information files for "RULER: Representation-level verification of machine unlearning"

<p dir="ltr">Supplementary information files for the paper "RULER: Representation-Level Verification of Machine Unlearning" by Georgina Cosma (Loughborough University) and Axel Finke (Newcastle University), accepted at the 2nd Workshop on Machine Unlearning and Privacy Preservation (WIPE-OUT 2026), co-located with ECML PKDD 2026, Naples, Italy, and published in Springer Communications in Computer and Information Science (CCIS).</p><p dir="ltr">The paper introduces RULER, a set of representation-level metrics for verifying machine unlearning.

The oracle-comparative metric M2 measures whether forget-set records occupy the same representational position as in a model retrained without them; the oracle-free metric M4 detects residual memorisation from the unlearned model's internal similarity structure alone.<br></p><p dir="ltr">This document contains the additional multi-seed experiments, robustness checks, per-dataset breakdowns, and architecture details referenced in the paper: the experimental settings summary, output-level and per-dataset MIA results, sensitivity and calibration of M2, robustness to mini-batch training, hyperparameter and forget-set sampling sensitivity, per-dataset representation shift and M4 distributions, the Bad Teacher comparison, M4 on large-scale tabular, image, and LFW face-identity settings, the image and Bio_ClinicalBERT architectures, and reproducibility details (Sections S1 to S15, Tables S1 to S13, Figures S1 to S6).<br><br>© The Author(s), CC BY-NC 4.0</p>