Non-cryptographic mobile wireless device authentication based on channel signature has aroused extensive attention. This technology uses the mobile device's physical characteristics to mark and verify its identity, and can be used to detect impersonation attacks and information forgery attacks. Channel State Information (CSI) has been used to generate fine-grained channel signatures. However, there are two sticking points in using CSI-based signature for authentication in association phase. The first is that the channel state will change as the device moves, which means that the local authenticator should be updated in real time to adapts to the latest channel state. The second is that the time complexity of authentication should be small enough to do packet-level authentication in association phase and detect attackers in time. In this paper, we propose a CSI-based authentication scheme, which can authenticate mobile devices at the packet level. Further, we provide an packet-level authentication framework based on neural networks. It uses a simple real-time authenticator update method to keep the authenticator valid. What's more, an ensemble of small-scale autoencoders are used to build the authenticator. It has been shown to significantly reduce the authentication's time complexity while maintaining the accuracy, providing the possibility for packet-level authentication. The evaluation shows that the packet-level framework can authenticate legitimate mobile devices with 95.19% accuracy and filter out attackers with even greater accuracy, which has higher time efficiency than traditional large-scale neural networks.