White-Box AI Model: Next Frontier of Wireless Communications

TL;DR

The paper addresses the opacity of conventional black-box AI in 6G wireless systems by proposing White-box AI (WAI), an explainable framework built on principled theory. It integrates Bayesian inference, information bottleneck, coding-rate reduction, and finite-horizon optimization within model-driven architectures such as deep unfolding and CRATE transformers to yield transparent, verifiable optimization paths. Core contributions include a modular large-scale WAI architecture with explicit pre-processing, task adaptation, and verifiable training, plus two case studies—an IB-based edge graph neural network for cell-free MIMO and a deep-unfolding PGD approach—that demonstrate improved spectrum efficiency and robustness under challenging conditions. The work highlights practical impact for reliable wireless optimization and outlines future directions in privacy, security, and edge intelligence for real-time, interpretable network operation.

Abstract

White-box AI (WAI), or explainable AI (XAI) model, a novel tool to achieve the reasoning behind decisions and predictions made by the AI algorithms, makes it more understandable and transparent. It offers a new approach to address key challenges of interpretability and mathematical validation in traditional black-box models. In this paper, WAI-aided wireless communication systems are proposed and investigated thoroughly to utilize the promising capabilities. First, we introduce the fundamental principles of WAI. Then, a detailed comparison between WAI and traditional black-box model is conducted in terms of optimization objectives and architecture design, with a focus on deep neural networks (DNNs) and transformer networks. Furthermore, in contrast to the traditional black-box methods, WAI leverages theory-driven causal modeling and verifiable optimization paths, thereby demonstrating potential advantages in areas such as signal processing and resource allocation. Finally, we outline future research directions for the integration of WAI in wireless communication systems.

Figures (6)

• Figure 1: Overview of WAI model: definition and fundamentals, with a comparative analysis of white-box and black-box model features.
• Figure 2: Fundamental theories supporting WAI model: (A) Bayesian Inference, (B) Information Bottleneck, (C) Coding Rate Reduction, and (D) Finite horizon strategy.
• Figure 3: ReduNet: A WAI model as an alternative to traditional black-box DNN model.
• Figure 4: Architecture of the white-box wireless large AI model. The model comprises four core modules: pre-processing, multi-task adaptation, white-box training, and multi-task output. With the Transformer serving as a representative example, illustration of its white-box counterpart, CRATE.
• Figure 5: Applications of WAI in wireless communication: architecture and optimization techniques. Key modules include signal processing and wireless resource management. Furthermore, we provide an overview of the WAI optimization design for core technologies in wireless communication.