Privacy-preserving federated learning approach based on Hensel’s compression and differential privacy
摘要
We present a novel two-layer mechanism that combines Hensel’s Lemma with differential privacy to enhance user privacy protection in federated learning. The first layer introduces a new dimensionality reduction method, utilizing Hensel’s Lemma, which aims to minimize the dimensions of the training dataset. Hensel’s Lemma ensures uniqueness, allowing our dimensionality reduction technique to lower the dataset’s dimensions without losing any essential information. The second layer applies differential privacy on the compressed dataset from the first layer. By employing differential privacy only once before training, our approach effectively mitigates the privacy leakage issues associated with compositional effects. Our approach enables federated learning clients to train their models on privacy-preserving datasets, ensuring that sensitive data remains confidential. Our extensive experimental results demonstrate that our proposed method offers robust privacy protection while maintaining significant accuracy. Furthermore, our approach utilizes significantly fewer resources, including energy, memory, and network throughput, compared to traditional learning models. Specifically, our method achieves