Back to all papers

EffiAuL: Influence-guided approximate unlearning for privacy-preserving medical image analysis.

October 21, 2026pubmed logopapers

Authors

Parvez MZ,Islam R,Zheng L,Islam MZ

Affiliations (2)

  • School of Computing, Mathematics and Engineering, Charles Sturt University, NSW, Australia.
  • AI and Cyber Futures Centre, Charles Sturt University, Australia.

Abstract

Machine unlearning is increasingly relevant to privacy-aware medical artificial intelligence when trained models must respond to subsequent data-removal requests. This study proposes EffiAuL, an influence-guided approximate-unlearning framework designed to balance forgetting behaviour, predictive-utility preservation, and computational cost. EffiAuL combines sample-level gradient-norm influence estimation, adaptive strategy selection, selective refinement of transfer-learned parameters, retained-data divergence control, and a validation-derived utility criterion with fallback to full retraining. The primary binary chest X-ray dataset comprised 2,330 Normal and COVID-19 images, partitioned using a fixed stratified 70/15/15 train-validation-test split and evaluated across five independent seeds. EffiAuL was compared with full retraining, SCRUB, SISA, IF-Diag, and a no-deletion FineTune-Control using predictive, forgetting-specific, membership-inference, distributional, computational, and stability measures. Additional validation used a four-class chest X-ray dataset containing 4,852 images. EffiAuL achieved a weighted F1-score of 0.9966 ± 0.0013 (95% CI: 0.9950-0.9982), compared with 0.9960 ± 0.0016 (95% CI: 0.9941-0.9979) for full retraining. The paired difference was not statistically significant (<i>p</i> = 0.374; Cohen's <i>d</i> <sub><i>z</i></sub> = 0.45), supporting utility preservation rather than superiority. Both achieved an AUC of 1.000. EffiAuL remained close to full retraining in prediction space (mean distance 0.0023), while the forget-set membership rate changed from 0.42 at baseline to 0.41 after EffiAuL, compared with 0.36 after full retraining. On the four-class dataset, weighted F1-scores were 0.8774 for EffiAuL and 0.8766 for full retraining. However, EffiAuL incurred greater end-to-end runtime because of influence-estimation and adaptive-selection overhead. EffiAuL preserves predictive utility close to full retraining while producing measurable forgetting-related behavioural changes. The findings support multidimensional evaluation of approximate unlearning rather than reliance on predictive performance or a single forgetting metric. EffiAuL provides empirical evidence of approximate unlearning and should not be interpreted as formally certified deletion or evidence of regulatory compliance.

Topics

Journal Article

Ready to Sharpen Your Edge?

Subscribe to join 11k+ peers who rely on RadAISlice. Get the essential weekly briefing that empowers you to navigate the future of radiology.

We respect your privacy. Unsubscribe at any time.