The article introduces a new approach, RQP-SGD, for privacy-preserving machine learning. The rise of IoT devices necessitates efficient and secure data processing. Traditional machine learning models can be impractical for large models, hence the need for models with quantized discrete weights that also preserve the privacy of the underlying dataset. RQP-SGD combines differentially private stochastic gradient descent (DP-SGD) with randomized quantization, providing a measurable privacy guarantee in machine learning. The study demonstrates its efficacy over deterministic quantization and shows its practical effectiveness through experiments conducted on two datasets.
Publication date: 9 Feb 2024
Project Page: https://arxiv.org/abs/2402.06606
Paper: https://arxiv.org/pdf/2402.06606
