Collaborative physical layer authentication (CPLA), which leverages spatial diversity, holds promise for enhancing the performance of feature-based physical layer authentication. However, some existing CPLA schemes simply aggregate the local information of collaborators to make final judgments and rarely consider the involvement of malicious collaborators. In this paper, we propose an impression-weighted based local decision aggregation scheme for detecting spoofing attacks in the presence of malicious collaborators. Specifically, the authenticator continually evaluates the authentication capabilities of collaborators by verifying the accuracy of local decisions and then synthesizes their long-term capabilities into impressions using a fuzzy membership function. These impression values will be dynamically updated upon completion of each authentication task. Moreover, a reinforcement learning scheme is employed to find the optimal threshold for authentication in a dynamic environment. Simulation results validate the high robustness and effectiveness of our proposed approach, guaranteeing the CPLA system's reliable operation.