Securing Federated Diffusion Model with Dynamic Quantization for Generative AI Services in Multiple-Access Artificial Intelligence of Things

Abstract

Generative diffusion models (GDMs) have emerged as potent tools for generating high-quality, creative content across various media, including audio, images, videos, and 3-D models. Their application in artificial intelligence-generated content (AIGC) marks a pivotal advancement in the evolution from the Internet of Things (IoT) to the Artificial Intelligence of Things (AIoT). Considering the inherent multiple-access nature of AIoT, training GDMs via federated learning and deploying them collaboratively is paramount. However, such approaches introduce considerable security risks and energy consumption challenges. To address these issues, we propose a comprehensive architecture for GDMs, encompassing both training and sampling stages. This architecture, termed secure and sustainable diffusion (SS-Diff), aims to thwart trigger-based security threats, such as backdoor attacks and trojan attacks, while simultaneously reducing energy consumption in multiple-access AIoT. The SS-Diff architecture incorporates a dynamic quantization mechanism within the training phase, significantly reducing communication overhead and thereby improving both spectrum and energy efficiency. During the sampling stage, a detection-based defense strategy is employed to identify and negate trigger inputs associated with malicious attacks. Through extensive simulations, we evaluate the performance of the SS-Diff architecture. The results demonstrate that the SS-Diff can effectively train GDMs and eliminate the impact of the attacks, compared with existing schemes.