PRISMA: Data-Distribution-Aware Approach to Machine Unlearning

Abstract

Preserving data subjects’ privacy has become increasingly important in the era of advanced artificial intelligence (AI). Ensuring lawfulness, fairness, and transparency in data processing requires giving data subjects control over how their personal data is used, including the right to revoke consent. However, removing data from the training set and retraining the entire model is often prohibitively expensive for real-world applications. Machine unlearning (MU) provides a promising approach to address this issue while safeguarding data privacy. One notable method, SISA (Sharded, Isolated, Sliced, and Aggregated), enables efficient removal of individual data points but can disrupt data distribution within shards, leading to performance degradation. To address these limitations, this study introduces PRISMA, a framework built on SISA that improves robustness, monitoring, and efficiency. PRISMA preserves class distribution within shards, integrates validation-based evaluation with extended metrics, applies label shift detection to support adaptive retraining, and enhances efficiency through optimized hashing and consent-aware data mapping. Designed to be explainable through transparent shard-level outcomes, PRISMA improves robustness, scalability, and deployment readiness. Experimental results across five benchmark datasets demonstrate that PRISMA consistently outperforms SISA in both accuracy and retraining efficiency. PRISMA achieves accuracy improvements of 3.0%-4.5% on Purchase, maintains Top-1 accuracy of 83.5%-84.0% on CIFAR-100 across 5 to 50 shards, in contrast to SISA’s degradation from 83.0% to 72.0%, and outperforms SISA on Mini-ImageNet across all shard configurations (78.1% vs. 64.0% at 10 shards; 64.2% vs. 55.0% at 20 shards). With respect to computational efficiency, PRISMA reduces per-compute-slice retraining time to 20-50 seconds on Purchase and 20-30 seconds on MNIST, compared to several hundred seconds required by SISA, and completes SVHN updates in hundreds of seconds versus thousands of seconds under SISA. These findings demonstrate that PRISMA provides a reliable, scalable, and practically deployable solution for machine unlearning, effectively satisfying real-world privacy compliance requirements without compromising model performance.