A Novel Differentially Private Low Rank Weight Fusion Approach

Differential privacy is widely applied in deep learning to mitigate the risk of sensitive information leakage. However, additional privacy protection measures, such as per-example gradient clipping, introduces extra computational overhead and storage requirements, which impede the exploration of differentially private deep learning methods. To address this challenge, we propose a novel differentially private low-rank weight fusion approach. Our approach utilizes a model weight composed of a shared weight and low-rank weights. The rank of the low-rank weights determines the number of sub-networks involved in the fusion process. During the training phase, we fix the shared weight and solely train the low-rank weights. After each training batch, we add noise to safeguard the privacy of the low-rank weight fusion operation, thereby achieving privacy-preserving training. This approach significantly reduces storage overhead and eliminates the need for per-example gradient computations. Additionally, the low-rank weight can be applied as a plug-in, particularly in transfer learning scenarios. We evaluate the knowledge transfer capability of our method and conduct experiments in both self-learning and transfer learning scenarios. Our results demonstrate that our approach achieves comparable accuracy to the batchensemble method without introducing additional noise.