The open radio access network (O-RAN) architecture's native virtualization and embedded intelligence facilitate RAN slicing and enable comprehensive end-to-end services in post-5G networks. However, any vulnerabilities could harm security. Therefore, artificial intelligence (AI) and machine learning (ML) security threats can even threaten O-RAN benefits. This paper proposes a novel approach to estimating the optimal number of predefined VNFs for each slice while addressing secure AI/ML methods for dynamic service admission control and power minimization in the O-RAN architecture. We solve this problem on two-time scales using mathematical methods for determining the predefined number of VNFs on a large time scale and the proximal policy optimization (PPO), a Deep Reinforcement Learning algorithm, for solving dynamic service admission control and power minimization for different slices on a small-time scale. To secure the ML system for O-RAN, we implement a moving target defense (MTD) strategy to prevent poisoning attacks by adding uncertainty to the system. Our experimental results show that the proposed PPO-based service admission control approach achieves an admission rate above 80\% and that the MTD strategy effectively strengthens the robustness of the PPO method against adversarial attacks.

6 pages