Paper 2026/1209 Original: eprint.iacr.org/2026/1209 ↗

Latency-Aware, High-Throughput Homomorphic AES Evaluation with CKKS

Abstract

Homomorphic Advanced Encryption Standard (AES) evaluation refers to evaluating the AES circuit with a fully homomorphic encryption (FHE)-encrypted secret key. Applications include in particular Transciphering, which converts AES-encrypted data into FHE ciphertexts without exposing the secret key. Existing homomorphic AES evaluations show a clear separation between latency-oriented solutions and throughput-oriented solutions. CKKS-based methods exploit massive SIMD parallelism and focus on throughput by processing many AES blocks in parallel. They are hardly suitable for latency-critical settings. In contrast, TFHE-based methods process a small number of blocks efficiently. They are preferable for low-latency settings, but provide very limited throughput. In this work, we show that AES-CKKS evaluation can achieve both interactive latency and high throughput. Our first variant is optimized for latency and decrypts a single AES block in only 26ms on an NVIDIA RTX-5090. This is more than 6× faster than recent TFHE-based state-of-the-art approaches; further, an extension of it processes 4 AES blocks at once in 29ms. Our second variant is based on a new embedding of $\textrm{GF}(16)$, the finite field with 16 elements, into CKKS message space. It is optimized for throughput and processes up to 2048 AES blocks at once, achieving 238KB/s throughput (a more than 3.41× improvement over the state-of-the-art CKKS-based approaches), while maintaining latency comparable to TFHE-based methods. To the best of our knowledge, this is the first AES-FHE evaluation algorithm combining good latency and throughput properties, bringing homomorphic outsourcing with AES within reach of real-time applications on constrained devices. Our main ingredients are redundant structures that maximize SIMD utilization, improved algorithms for the SubBytes step (one of them being based on inversion in $\textrm{GF}(256)$ using CKKS), fusion of linear layers into bootstrapping, and carefully crafted FHE parameters.