"Accuracy prediction, prevention strategies, and latency testing for xAPP, alongside cryptographic FL-DDQN."

"The integration of federated learning (FL) in 6G O-RAN architectures enhances collaborative learning between distributed users while enabling network monitoring capabilities and preserving data privacy. However, despite its significant advantages, FL has critical security vulnerabilities in the transmission of packets\/weights to\/from the Open Radio Unit (ORU) to the user. These challenges require the use of robust prevention mechanisms to mitigate malicious actors disrupting weight transmission. This paper addresses the security challenges in FL within 6G O-RAN by employing cryptographic methods to secure communications between users and the ORU. Initially, Diffie-Hellman (DH) cryptography is utilized during the initialization phase of FL to establish shared keys securely. These DH shared keys are subsequently used to derive Advanced Encryption Standard keys, enabling the encryption and decryption of weight exchanged between ORUs and users during FL. To further enhance security, Binary Cross Entropy is employed to train the extended application for predicting free rider attacks and issuing mitigation policies. The security problem is formulated as a Markov Decision Process and addressed using the Double Deep Q Network algorithm. Our empirical results indicate a 4.23% improvement in attack detection accuracy, a 5% increase in attack prevention accuracy, and a 5% reduction in latency during FL."

在6G开放式无线接入网（O-RAN）架构中集成联邦学习（FL），可增强分布式用户间的协同学习能力，同时实现网络监控功能并保障数据隐私。然而，尽管具备显著优势，FL在数据包/权重往返于开放式无线电单元（ORU）与用户之间的传输过程中存在严重的安全漏洞。这些挑战需要采用强健的预防机制，以缓解恶意攻击者对权重传输的干扰。本文针对6G O-RAN中FL的安全挑战，采用密码学方法保障用户与ORU之间的通信安全。首先，在FL初始化阶段采用Diffie-Hellman（DH）密码学技术，以安全地建立共享密钥。这些DH共享密钥随后被用于生成高级加密标准（Advanced Encryption Standard）密钥，从而实现FL过程中ORU与用户之间交换权重的加密和解密。为进一步提升安全性，本文采用二元交叉熵（Binary Cross Entropy）训练扩展应用，以预测搭便车攻击并制定缓解策略。该安全问题被建模为马尔可夫决策过程（Markov Decision Process），并采用双深度Q网络（Double Deep Q Network）算法加以解决。我们的实证结果表明，FL过程中的攻击检测准确率提升了4.23%，攻击预防准确率提高了5%，延迟降低了5%。