If a model can operate on ciphertexts, the party holding the data never has to reveal it to the party running the computation. In principle, this offers a cryptographically rigorous foundation for privacy-preserving machine learning in healthcare, finance, and user-facing services, where exposing raw data is often the greatest risk. In practice, the survey's authors argue, integrating FHE into modern machine learning pipelines is far from straightforward.

The encryption imposes substantial constraints on arithmetic, numerical precision, and computational efficiency, and these constraints do not sit quietly at the edge of a system. They fundamentally affect how models are designed, which algorithms can be selected in the first place, and how the resulting systems perform. The review examines how a wide range of machine learning models, including traditional algorithms, deep neural networks, and emerging transformer architectures, have been adapted to operate under the restrictions that FHE imposes.

A central contribution of the paper is a unifying taxonomy that organizes the literature along key design dimensions. These include the choice of encryption scheme, strategies for handling nonlinear operations, which are notoriously difficult under encrypted arithmetic, the families of models being adapted, system-level optimizations, and the privacy objectives each approach pursues. The framework is intended to make it easier for researchers to compare methods that would otherwise be scattered across cryptography and machine learning venues.

Using this framework, the authors synthesize existing approaches and analyze the trade-offs between accuracy, efficiency, and security that recur throughout the field. They also identify recurring design patterns across the literature, offering a structured picture of how practitioners have repeatedly solved the same classes of problems when running learning systems on encrypted data. The survey does not stop at summarization.

It discusses practical limitations and deployment considerations that shape whether privacy-preserving machine learning under encryption can move from demonstration to production, and it lays out open research challenges that the authors see as defining the field's future direction. The abstract makes clear that these challenges remain open rather than solved. By bridging cryptographic techniques and modern AI system design, the authors position the work as a reference for researchers and practitioners seeking to build machine learning systems that are secure, scalable, and trustworthy.

As machine learning continues to spread into domains where data sensitivity is paramount, the survey suggests that managing the costs and constraints of encrypted computation will be central to making privacy-preserving AI practical.

Subject of Research: Machine learning over encrypted data: a survey of fully homomorphic encryption for AI systems

Article Title: Machine learning over encrypted data: a survey of fully homomorphic encryption for AI systems

Article References: Davidson, D., Jackson, M., Hussain, R., Xiong, Z., & Son, J. (2026). Machine learning over encrypted data: a survey of fully homomorphic encryption for AI systems. Artificial Intelligence Review. https://doi.org/10.1007/s10462-026-11707-2

Image Credits: AI Generated

DOI: 10.1007/s10462-026-11707-2

Keywords: Machine, learning, over, encrypted, data, survey, fully, homomorphic, encryption, systems, scientific research