BiLSTM-based Internet of Things Security Posture Prediction Research

In order to improve the sensing accuracy and response efficiency of security posture in IoT environment, a BiLSTM prediction model integrating multi-dimensional features and time-series modeling capability is constructed. The model is based on 11.8 million log data, and 38 standardized features are extracted to construct a 120-step-long time-series input tensor. The bidirectional LSTM architecture is used to mine the contextual dependencies, and the training strategy is optimized by combining cosine annealing, L2 regularity and weighted cross-entropy. The F1 value under six types of attack labels is maintained above 0.91, the accuracy rate is up to 0.968, and the inference time is controlled within 15.7ms, and the analysis concludes that the model has strong robustness and generalization ability while maintaining computational efficiency.