Training TFHE-Based Neural Networks with Approximated Floating-Point Arithmetic

This dataset repository contains the specific preprocessed variants of standard machine learning benchmarks (including MNIST, Fashion-MNIST, Ternary-MNIST, Derma-MNIST, Blood-MNIST, and CIFAR-10) used to evaluate the framework presented in the paper "Training TFHE-Based Neural Networks with Approximated Floating-Point Arithmetic" (Accepted at Privacy Enhancing Technologies Symposium, 2026).

Because Fully Homomorphic Encryption (FHE) frameworks, specifically those utilizing TFHE-assisted floating-point operations, impose strict constraints on data ingestion pipelines, these assets have undergone highly specific structural preprocessing. These modifications ensure optimal bit-width alignment, normalization, and tensor formatting required by the system's underlying cryptographic layers to achieve the exact convergence rates and cryptographic performance benchmarks reported in our study.

Structure & Usage To optimize storage and bandwidth, the benchmarks are partitioned into individual, modular archive files (e.g., ternary_mnist.zip, fashion_mnist.zip). Users do not need to download this 2GB repository manually. The official implementation companion repository automatically interfaces with these Zenodo records under the hood. Upon executing a specific experiment via our interactive CLI tool (cargo run --release), the code validates local assets, fetches the respective target zip file from this archive, and extracts it on-demand to guarantee seamless, deterministic reproducibility.

Keywords Fully Homomorphic Encryption (FHE), TFHE, Privacy-Preserving Machine Learning (PPML), Encrypted Neural Networks, Floating-Point Arithmetic, Reproducible Research.