EUVIMEDEUVIMED entdecken
Crossref

Reasoning Under Encryption: Assessing the Use of Large Language Models in Privacy-Preserving Drug Research

Amit Chakraborty, Raj Kumar Keshri, Rajdeep Chakraborty, Chirantana Mallick, Kasturi Barik

International Journal of Drug Delivery Technology · 08.06.2026 · 16 (47s)

DOI: 10.25258/ijddt.16.47s.91 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal of Drug Delivery Technology, online erschienen am 08.06.2026. Verfasst von Amit Chakraborty und 4 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

With the adoption of large language models (LLMs) into highly sensitive fields like healthcare, finance, and legal services, safeguarding user information in the course of model inference has become the most critical issue. Conventional deployment pipelines of LLM require access to plaintext user inputs, which means that the raw data have to be visible to the model to create an output. These solutions reveal extremely sensitive data, which may be going against privacy laws, including HIPAA and GDPR, and deceive any user. In this regard, there has emerged a hopeful cryptographic tool of homomorphic encryption (HE) since it allows one to compute on the encrypted information and does not require it to be decrypted at any point. Hypothetically, it allows privacy-preserving inference of the LLM, where the user can make encrypted inquiries and get encrypted responses, ignoring that the model does not know anything about the particular content of the information. Nevertheless, despite its conceptual strength, homomorphic encryption faces serious difficulties in the case of large scale deep learning models like transformers. HE schemes are computationally heavy and have significant overhead in terms of time and memory. Also, multiple computations of encrypted data acquire noise in the encrypted data, which may ultimately distort computations unless handled carefully. These bottlenecks make the straightforward use of HE to full-sized LLMs, involving billions of parameters and intricate attention mechanisms impracticable. Therefore, there has to be a balance between privacy and calculations in practice. The paper is limited to the potentially realistic discussion of the prospects of homomorphic encryption to the inference of the LLM, especially in the context of situations in which the absolute confidentiality of data is of paramount importance, like in asking clinical questions. Instead of trying to encrypt very advanced models, we explore the performance of reduced and shallow transformer-based models in combination with algorithmic HE designs. Our experiments measure important trade-offs between model accuracy, inference latency, and cryptographic security, and gives a systematic study of current capabilities. These findings suggest that all homomorphic implementation of deep transformer nets is still not a viable solution to realtime prediction as they have very high computational requirements. Times of inference with HE may go up hundreds of milliseconds to several minutes or even minutes, and make interactive application unrealistic. Besides, noise in cipher text restricts sequential operations, which restricts the depth and expressiveness of the model. Nonetheless, our research reports that, actually, with a tedious selection of parameters and with reduced model architectures, relatively basic reasoning tasks, which may include binary classification or even simple clinical question answering, can be performed with sufficiently low latency with only small losses in precision. We also determine major research directions that allow us to facilitate future improvement. They consist of creation of HE friendly model architectures, the minimization of the number of polynomials in attention systems and the means of binding cryptographic libraries and machine learning systems more closely to one another. As well, hybrid schemes involving the application of both homomorphic encryption with communication schemes like secure multiparty computation (MPC) or trusted execution environments (TEEs) are promising to be more efficient and at the same time offer strong privacy assurance. To conclude, this paper is a methodical analysis of the potential and restriction of homomorphic encryption in the inference of privacy-preserving language models. Today, even partial deployment with full-scale encryption of the LLM is beyond the capability of current technology, but great improvements can be achieved with customized architectures and optimization of cryptography. We would like to inform the researchers and practitioners who may desire to apply secure and privacy-conscious AI systems across settings that place confidentiality above negotiable using this paper to inform and present the trade-offs and practical limitations.

Zitieren

APA

Chakraborty, A., Keshri, R. K., Chakraborty, R., Mallick, C., & Barik, K. (2026). Reasoning Under Encryption: Assessing the Use of Large Language Models in Privacy-Preserving Drug Research. International Journal of Drug Delivery Technology, 16(47s). https://doi.org/10.25258/ijddt.16.47s.91

Vancouver

Chakraborty A, Keshri RK, Chakraborty R, Mallick C, Barik K. Reasoning Under Encryption: Assessing the Use of Large Language Models in Privacy-Preserving Drug Research. International Journal of Drug Delivery Technology. 2026;16(47s). doi:10.25258/ijddt.16.47s.91

BibTeX
@article{Chakraborty2026,
  author = {Chakraborty, Amit and Keshri, Raj Kumar and Chakraborty, Rajdeep and Mallick, Chirantana and Barik, Kasturi},
  title = {Reasoning Under Encryption: Assessing the Use of Large Language Models in Privacy-Preserving Drug Research},
  journal = {International Journal of Drug Delivery Technology},
  year = {2026},
  volume = {16},
  number = {47s},
  doi = {10.25258/ijddt.16.47s.91}
}
Kostenlos starten →