An Enhanced MIBKA-CNN-BiLSTM Model for Fake Information Detection

The complexity of fake information and the inefficiency of parameter optimization in detection models present dual challenges for current detection technologies. Therefore, this paper proposes a hybrid detection model named MIBKA-CNN-BiLSTM, which significantly improves detection accuracy and efficiency through a triple-strategy enhancement of the Black Kite Optimization Algorithm (MIBKA) and an optimized dual-channel deep learning architecture. First, three improvements are introduced in the MIBKA. The population initialization process is restructured using circle chaotic mapping to enhance parameter space coverage. The conventional random perturbation is replaced by a random-to-elite differential mutation strategy (DE/rand-to-best/1) to balance global exploration and local exploitation. Moreover, a logarithmic spiral opposition-based learning (LSOBL) mechanism is integrated to dynamically explore the opposition solution space. Second, a CNN-BiLSTM dual-channel feature extraction network is constructed, with hyperparameters such as the number of convolutional kernels and LSTM units optimized by MIBKA to enable adaptive model structure alignment with task requirements. Finally, a high-quality fake information dataset is created based on social media platforms, including CCTV. The experimental results show that our model achieves the highest accuracy on the self-built dataset, which is 3.11% higher than the optimal hybrid model. Additionally, on the Weibo21 dataset, our model's accuracy and F1-score increased by 1.52% and 1.71%, respectively, compared to the average values of all baseline models. These findings offer a practical and effective approach for detecting lightweight and robust false information.