Machine unlearning for class removal through SISA-based deep neural network architectures

Abstract

As the adoption of image generation models and other AI systems accelerates, concerns around user data privacy and consent have become increasingly critical. With the slowdown in growth of publicly available data, major tech companies are anticipated to rely more on proprietary and private user data for training their models. This raises ethical and legal questions, particularly when users request data deletion after it has already influenced a trained model—a process that is both technically challenging and resource-intensive. The emerging field of machine unlearning addresses this issue by developing techniques to remove specific data from trained models. Among these, the SISA (Sharded, Isolated, Sliced, and Aggregated) framework has shown promise as a scalable and privacy-aware unlearning solution. In this research, we focus on leveraging a modified SISA framework to enable class-level data removal in Convolutional Neural Network (CNN) architectures. We evaluate how well the modified SISA framework supports effective class unlearning without requiring full model retraining, and analyze the trade-offs in model performance, accuracy, and privacy. Through experiments across various image datasets and CNN architectures, we aim to demonstrate the practical viability of SISA-based class unlearning for real-world applications, offering insights into its strengths, limitations, and potential for deployment in privacy-sensitive AI systems. The code for this research is publicly available at https://github.com/SiamFS/sisa-class-unlearning.