Batchencryption: Localized Federated Learning in Privacy-Preserving with Efficient Integer Vector Homomorphic Encryption
Tianying Xie, Shaohong Zhou, Qi Qi, Yantao Li · 2025
Deep Learning as a Service (DLaaS) offers an efficient solution to leverage Deep Neural Networks (DNN) across various applications. Despite the high accuracy of the DNNs, there is an urgent need for robust protocols to ensure the privacy and security of sensitive data. Allowing cloud models to process such data without proper safeguards can lead to significant privacy breaches. This concern is particularly acute during the training phase, where most models are typically trained on plaintext rather than ciphertext, posing a risk of potential privacy leaks for data owners. Although some researchers have started using homomorphically encrypted data for training, the majority still rely on a single pair of keys for encryption and decryption. This approach overlooks the importance of robustness in encryption schemes when multiple users employ unique keys. To address these issues, this paper introduces a practical localized Federated Learning (FL) method called BatchEncryption, utilizing Efficient Integer Vector Homomorphic Encryption (EIVHE) for privacypreserving training and inference phases. The BatchEncryption enhances the diversity and robustness of the model by encrypting raw datasets in blocks using different pairs of keys. Our experiments demonstrate that training neural networks with multiple pairs of keys yields higher accuracy compared to using a single pair of keys. The proposed method achieves an approximate accuracy 97 % in the MNIST data set with data encrypted by different pairs of keys.