Saved in:
Bibliographic Details
Main Authors: Weng, Jiasi, Weng, Jian, Li, Ming
Format: Recurso digital
Language:English
Published: Zenodo 2026
Online Access:https://doi.org/10.5281/zenodo.18995804
Tags: Add Tag
No Tags, Be the first to tag this record!
Table of Contents:
  • <p>This paper proposes a built-in framework that embeds a dedicated “Crypto Expert” directly into large language models (LLMs) architecture. As an initial attempt, we design a differentiable proxy tailored to the Advanced Encryption Standard (AES) algorithm, using our customized neuron units, including SoftXOR, SoftLUT and GF-conv neurons. These units provide functional equivalence to the AES within the Boolean domain, while enabling stable gradients for backpropagation. By integrating this differentiable proxy as a specialized expert within a Mixture-of-Expert (MoE) LLM, the LLM learns to autonomously route and encrypt sensitive tokens during the training phase. After training, the differentiable proxy is seamlessly swapped for a real and discrete AES implementation to guarantee provable security at inference. Our empirical evaluations demonstrate that our approach significantly reduces neuron counts and latency compared to prior ReLU-based representation, mitigates continuous differential attacks, and enforces end-to-end data protection without degrading downstream task utility. We expect this attempt to serve as a catalyst for future research into the seamless fusion of formal cryptographic guarantees and deep learning computation graphs.</p>