Secure Cryptographic Federated Learning over Air Interface in 6G O-RAN

The integration of federated learning (FL) into 6G O-RAN architectures enhances collaborative model training while preserving data privacy. However, FL still faces challenges such as data corruption by intermediaries and slow convergence rates. To address these issues, this paper employs the Diffie\u2013Hellman (DH) protocol for key management to establish secure shared keys during the initialization phase of FL. These DH keys are subsequently used to derive Advanced Encryption Standard (AES) keys, to encrypt weights. In parallel, FedProx is adopted to promote faster and more stable convergence. To mitigate parameter inversion attacks across the fronthaul interface, Medium Access Control Security (MACsec) is applied to ensure the confidentiality and integrity of transmitted data. Furthermore, to counter Byzantine behaviors, we integrate access control mechanisms to filter malicious participants and utilize Binary Cross-Entropy (BCE) to detect and exclude free-riders during weight aggregation. The overall security challenge is formulated as a Markov Decision Process (MDP) and solved using the Soft Actor-Critic (SAC) algorithm. Experimental results demonstrate that the proposed framework achieves a 20\\% improvement in convergence speed, a 0.03\\% increase in attack prevention effectiveness, and a 0.04\\% reduction in packet drop rate compared with secure aggregation (SA) and homomorphic encryption (HE) baselines.