Privacy-Preserving Rule Induction Using CKKS

Abstract

Rule-based learning involves using specific rules to categorize or identify datasets. This study introduces a new approach called homomorphic encryption-based rule induction (HORI) algorithm, designed specifically for scenarios where data confidentiality is critical. This method is constructed using CKKS homomorphic encryption to make it work with encrypted data, thereby enhancing the privacy of both training data and the input for inference. To overcome the inefficiency caused by utilizing homomorphic encryption, we utilize the modified Gini impurity index (MGI) (Bădulescu, 2020) for training, simplify training with a single variable, and speed up inference by combining all relevant rules into a single ciphertext. Comparative analysis shows that the training algorithm of this method is, on average, 1,019 times slower, and the inference algorithm is approximately 935 times slower than their counterparts working with plaintext data. However, these performance metrics are still considered efficient compared to traditional algorithms based on homomorphic encryption, which often have latency factors ranging from 2,000 to 10,000 times. Additionally, when compared to recent work by Zorapaci and Ayşe Özel (2021) incorporating differential privacy, the proposed method demonstrates a superior accuracy improvement of over 10%.

Keywords

• Rule induction
• fully homomorphic encryption
• CKKS
• privacy-preserving machine learning