Skip to content
#graph neural networks Book Open access

A Framework for Automatic Conversion of Neural Networks to Homomorphic Equivalents

Nov 2026 · Proceedings of the 14th Workshop on Encrypted Computing & Applied Homomorphic Cryptography · 0 citations · 11 references

TL;DR

This paper presents a framework that automatically converts a pre-trained Keras model into a CKKS equivalent, handling parameter inference, bootstrap placement and encrypted execution without retraining, with the user only providing a configuration of the encryption parameters.

Abstract

Machine learning inference on sensitive data exposes user inputs to the server, raising privacy and legal concerns. Fully homomorphic encryption (FHE) resolves this by allowing inference to run directly on encrypted data. However, deploying an FHE neural network currently requires deep cryptographic expertise and significant manual effort for each model. This paper presents a framework that automatically converts a pre-trained Keras model into a CKKS equivalent, handling parameter inference, bootstrap placement and encrypted execution without retraining, with the user only providing a configuration of the encryption parameters. The framework supports dense, convolutional, residual and multi-head attention architectures through a graph-based intermediate representation. The framework is evaluated against Orion on MNIST and CIFAR-10, on the same machine and inputs. On every evaluated input, the encrypted models reproduce the predictions of their plaintext counterparts, with a higher numerical precision than Orion. Orion is faster on a single thread, but with multi-threading and batching the framework reaches up to a 2.6× lower amortised latency. Multi-head attention is demonstrated on a small custom model, representing an architectural capability not present in Orion.

Read PDF

Similar papers

#artificial intelligence Preprint Nov 2026

Perseus: A Full-Stack System for Homomorphic Model Inference

Large Language Models (LLMs) have made it exceptionally convenient for users to send their private data to model providers, which process them in plaintext. Fully Homomorphic Encryption (FHE) promises to secure remote inference by computing directly on encrypted data, yet it has remained prohibitively slow. The effort...

Alessandro Zirilli, Davide Marincione, E. Kornaropoulos et al. · 0 citations
#small language model Book Open access Nov 2026

Classification in the Dark: Performance Insights and Novel Methods for Homomorphic Binary Networks

A performance evaluation of three binary classification neural networks on consumer hardware, and a novel method for deriving a network's final classification using the binary step function are proposed.

Lukas Horn, Lennart Svoboda, Simon Engel et al. · 0 citations
Sep 2026

An Open-Source End-to-End FHE Implementation for Privacy-Preserving Llama 3 8B Inference

Odin is the first open-source end-to-end GPU CKKS implementation of Llama-3, an FHE inference system that co-designs ciphertext packing and model execution for Llama and uses a feature-major cross-layer layout to unify residual connections and layer interfaces.

Yu-Hang Fan, Yu-Si Chen, Kan-Yu Ye et al. · 1 citation
Preprint Oct 2026

Pacifo: An Efficient Fully Homomorphic Encryption-based Lightweight ResNet for Trustworthy Machine Learning as a Service

This paper proposes Pacifo, a framework that adapts customized fine-tuning strategies to efficiently evaluate the ResNet20 model in the context of FHE, and introduces a public-private layer partitioning scheme and an efficient ReLU approximation, significantly reducing the number of private layers and costly FHE operat...

Yang-Zhuo Li, Gang He, Rui-Chen Zhang et al. · 0 citations
Preprint Oct 2026

Fully Homomorphic Encryption for Statistical Modeling

Fully homomorphic encryption allows arithmetic to be carried out on encrypted values, so that a party performing a computation need not see the data it operates on. This is useful wherever collaborating parties or sites cannot share data records or computed summaries yet need to do a joint analysis. We demonstrate the...

B. Narasimhan · 0 citations
Sep 2026

SpliTEE: Fast and Private LLM Inference by Coupling GPU-Assisted Trusted Execution Environments with Differential Privacy

This paper demonstrates that prompt reconstruction, even with knowledge of the DP mechanism, cannot recover more information than is contained in an unrelated prompt, and derives an upper bound on the floating-point error from masking and subsequent noise cancellation as a function of the privacy parameter epsilon.

Shashie Dilhara Batan Arachchige, Robin Carpentier, H. Asghar et al. · 0 citations

Related blog posts

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.