Rule learning is the use of rules to identify or classify data. Rule learning can be broadly divided into two types: decision tree and rule induction. Decision Tree is a popular and widely used rule learning method due to its advantages of high explan...
Rule learning is the use of rules to identify or classify data. Rule learning can be broadly divided into two types: decision tree and rule induction. Decision Tree is a popular and widely used rule learning method due to its advantages of high explanatory power because it can be viewed visually and the model is easy to interpret. However, decision trees also have disadvantages. They are The irrelevant condition problem. Irrelevant conditions that are not related to the model can make it difficult to interpret the model and reduce accuracy. Rule induction, in contrast, does not have this problem. On the other hand, data privacy is becoming increasingly important in systems that perform training and inference based on data. In particular, data privacy is important in fields such as finance and healthcare, where the data itself contains a lot of individual information, and data leakage can lead to privacy violations or criminal issues. Data privacy is also necessary when the input itself should not be exposed, and when the model itself has a financial value. In this study, we propose homomorphic rule induction to use rule induction in environments where data privacy is important. To this end, we consider the fol- lowing three issues that arise from using homomorphic ciphers: efficient rule generation using modified-gini, efficient training using only one variable, and efficient inference by collecting all rules in one ciphertext. The proposed training method can be trained on encrypted data without decryption, providing high data piracy. The learning time of homomorphic rule induction is on average 1,120 times slower for ciphertext training than for plaintext training, and the inference time is on average 64 times slower for ciphertext inference than for plaintext inference. This is a very efficient implementation compared to homomorphic algorithms, which are typically over 2,000 times slower than plaintext algorithms. In addition, when comparing the accuracy of encrypted data and unencrypted plaintext data using the homomorphic rule induction method, the average difference is 0.018, which shows that the accuracy of plaintext and ciphertext is not significantly different, indicating that homomorphic rule induction can be used in practice.