Channel-based Physical Layer Authentication (PLA) leverages unique channel characteristics to provide a lightweight and hardware-independent security mechanism for resource-constrained devices. However, this reliance on spatial decorrelation renders it inherently vulnerable to advanced adversaries, particularly co-located and replay attackers who can mimic legitimate channel statistics. To overcome these limitations, this paper proposes a robust multi-domain protection framework that synergizes spatial, temporal, and logical security dimensions. Unlike passive approaches, our scheme utilizes active time-varying Reconfigurable Intelligent Surfaces (RIS) to inject temporal entropy into the channel, thereby invalidating historical fingerprints to thwart replay attacks. Simultaneously, it embeds an RIS-assisted, session-key-based watermark as logical identity features to distinguish spatially correlated co-located adversaries. Theoretically, we analyze the impact of RIS phase quantization on channel randomness and derive closed-form approximations for authentication error probabilities under a Generalized Likelihood Ratio Test (GLRT). Furthermore, to address the limitations of linear statistical metrics in complex environments, we propose InterCRNet, a physics-aware neural network designed to explicitly model the interaction between RIS configurations and Channel State Information (CSI). Experimental results demonstrate that this multi-domain integration significantly outperforms existing baselines, achieving superior robustness and scalability against heterogeneous attacks in multi-user scenarios.
