MLQENABLER: Enabling Secure Machine Learning Queries over Encrypted Database in Cloud Computing

As cloud service providers increasingly offer machine learning (ML) services alongside storage, a conflict arises between data privacy and data utility. Clients want to encrypt their data to prevent unauthorized access by the cloud provider, but standard encryption renders data unusable for ML training. This paper addresses the challenge of enabling ML-based services on outsourced data without exposing the underlying raw information to the cloud.

The authors propose MLQENABLER, a framework that uses an index-aided approach to bridge the gap between security and utility. The system employs a specialized architecture called Encryption GAN (EncGAN). This framework consists of three components: a generator that transforms original data into secure index items, a discriminator that ensures these index items are indistinguishable from random noise, and a reconstructor that ensures the index items retain enough structural information for ML models to learn from them. By storing both the formally encrypted raw data (for storage security) and the secure index items (for ML utility), the system allows the cloud to train models without ever seeing the original, unencrypted data.

MLQENABLER successfully achieves four privacy goals: ciphertext privacy, index privacy, token privacy, and model privacy. By using a loss-function-controllable noise addition method, the system provides a more fine-grained approach to privacy than traditional differential privacy. Experimental results indicate that the framework achieves an acceptable level of security while incurring only a slight degradation in ML performance, effectively balancing the needs of the data-sharing economy with strict privacy requirements.

This research provides a practical pathway for businesses to leverage the massive computational power of public clouds for machine learning without sacrificing the confidentiality of their proprietary datasets. It addresses the growing regulatory pressure, such as GDPR, by ensuring that data remains encrypted while still contributing to the training of valuable ML models.