Criptografia Completamente Homomórfica com Concrete ML: Análise de desempenho com o Scikit-learn

Abstract

With the growing adoption of cloud computing for data processing, including sensitive information, information security in machine learning pipelines has become crucial. Traditional cryptography protects data at rest and in transit, but the need for plaintext processing during model execution represents a privacy vulnerability. Fully Homomorphic Encryption (FHE) has emerged as a promising solution, allowing computations directly on encrypted data to mitigate this risk. However, its application in ML algorithms faces practical challenges, especially in relation to the potential impact on the accuracy and effectiveness of the models. To investigate the feasibility of maintaining performance under FHE, this study performs a comparative analysis between traditional Scikit-Learn models and their encrypted implementations in Concrete ML. Using public health data related to COVID-19, the performance of Logistic Regression and Decision Trees was evaluated in terms of accuracy, precision, recall and AUC-ROC. The results show that while Decision Trees maintained statistically equivalent performance in their encrypted version, Logistic Regression showed an 8% reduction in accuracy, but with exceptionally high recall (>99%). Statistical tests carried out confirmed these differences, suggesting that FHE currently offers better value for money for non-linear models. The study concludes that, although there are trade-offs between privacy and performance, FHE is already proving viable for specific use cases where data confidentiality is a priority, especially when combined with hardware optimizations and adaptive quantization techniques.