Evaluating Sisa-Based Machine Unlearning Across Diverse Modalities: Tabular, Visual and Auditory Data

Retraining a model entirely from scratch to delete specific information is highly inefficient in terms of both time and cost. This study comprehensively evaluates the training framework known as Sharded, Isolated, Sliced, and Aggregated (SISA) as a practical solution to this problem. Researchers tested SISA’s unlearning performance across a wide range of data types. These included tabular data from the Purchase and Adult datasets, visual data from the DR Grading dataset, and auditory data from the ESC fifty dataset. The core principle of the method involves partitioning data into small, independent sections. This allows for the retraining of only the relevant component when a deletion request is made. The study’s key findings indicate that SISA’s success is largely dependent on the characteristics of the data. The approach demonstrates exceptional efficiency with large, well-structured datasets, significantly reducing retraining time while maintaining high model accuracy. However, its performance shows fluctuations when applied to more complex or irregular datasets. Ultimately, this work presents the SISA method as a scalable and viable strategy for data privacy in modern artificial intelligence systems, offering a crucial balance between model accuracy, operational efficiency, and legal compliance.