High Complexity Time Series Prediction Generation Based on CNN-LSTM-MHA and its Application in Image Encryption Algorithm

To address the shortcomings of classical chaotic time series in image encryption algorithms in terms of low complexity, fewer control parameters, and limited range of value domains, this paper proposes a chaotic time series prediction and generation method based on CNN-LSTM and Multi-Head Attention Mechanism (CNN-LSTM-MHA) models. First, Coupling the Sine mapping with the Tent mapping, and then cascading the coupled mapping with the Sine mapping to increase its complexity. Second, the processed chaotic time series are predicted using the CNN-LSTM-MHA model and the prediction results are compared with those of CNN, LSTM, and CNN-LSTM models. Subsequently, the information entropy and NIST test results of the CNN-LSTM-MHA model generated sequences are comprehensively analyzed. The test results show that the generated sequences have good distributional and stochastic properties. Finally, the time series generated by CNN-LSTM-MHA is used for image encryption experiments and analyzed for histogram, correlation, differential attack, and information entropy. The experimental results prove that the time series generated by the CNN-LSTM-MHA model exhibits stronger randomness and good plaintext sensitivity in image encryption.