The advent of sixth-generation (6G) networks is expected to start a new era in mobile networks, characterized by unprecedented high demands on dense connectivity, ultra-reliability, low latency, and high throughput. Artificial intelligence (AI) is at the forefront of this progress, optimizing and enabling intelligence for essential 6G functions such as radio resource allocation, slicing, service offloading, and mobility management. However, AI is subject to a wide range of security risks, most notably adversarial attacks. Recent studies, inspired by computer vision and natural language processing, show that adversarial attacks have significantly reduced performance and caused incorrect decisions in wireless communications, jeopardizing the perspective of transforming AI-based 6G core networks. This survey presents a thorough investigation into the landscape of adversarial attacks and defenses in the realm of AI-powered functions within classic wireless networks, open radio access networks (O-RAN), and 6G networks. Two key findings are as follows. First, by leveraging shared wireless networks, attackers can provide noise perturbation or signal sampling for interference, resulting in misclassification in AI-based channel estimation and signal classification. From these basic weaknesses, 6G introduces new threat vectors from AI-based core functionalities, such as malicious agents in federated learning-based service offloading and adversarial attacks on O-RAN near-real-time RIC (xApp). Second, adversarial training, trustworthy mmWave/Terahertz datasets, adversarial anomaly detection, and quantum technologies for adversarial defenses are the most promising strategies for mitigating the negative effects of the attacks. This survey also identifies possible future research topics for adversarial attacks and countermeasures in 6G AI-enabled technologies.

中文翻译：

---

6G 网络中 AI 驱动型应用对抗性攻击的安全风险及对策：一项调查


第六代 （6G） 网络的出现有望开启移动网络的新时代，其特点是对密集连接、超可靠性、低延迟和高吞吐量的空前高要求。人工智能 （AI） 处于这一进展的最前沿，它优化和实现 6G 基本功能的智能，例如无线电资源分配、切片、服务卸载和移动性管理。但是，AI 面临广泛的安全风险，最明显的是对抗性攻击。受计算机视觉和自然语言处理启发的最新研究表明，对抗性攻击显著降低了性能，并导致无线通信做出错误决策，从而危及基于 AI 的 6G 核心网络转型前景。本调查对经典无线网络、开放式无线接入网络 （O-RAN） 和 6G 网络中 AI 驱动的功能领域的对抗性攻击和防御形势进行了全面调查。两个主要发现如下。首先，通过利用共享无线网络，攻击者可以提供噪声扰动或信号采样来干扰，从而导致基于 AI 的信道估计和信号分类错误分类。从这些基本弱点开始，6G 从基于 AI 的核心功能中引入了新的威胁向量，例如基于联合学习的服务卸载中的恶意代理和对 O-RAN 近实时 RIC （xApp） 的对抗性攻击。其次，对抗性训练、值得信赖的毫米波/太赫兹数据集、对抗性异常检测和用于对抗性防御的量子技术是减轻攻击负面影响的最有前途的策略。 该调查还确定了 6G AI 技术中对抗性攻击和对策的未来可能研究主题。