With the fast-paced development in technology and the broad effect brought about by the Internet on business and communication, cybersecurity has become one of the hottest topics in the world today. In this research work, a novel FinSafeNet based deep learning model is introduced for secured cash transaction in digital banking environment. This FinSafeNet is developed by blending the Bi-LSTM model with Convolutional Neural Network (CNN) and dual-attention mechanism. On the other hand, the transaction time has been the major criterion, which has been satisfied by selecting the optimal features with Improved Snow-lion optimization model (I-SLOA). This I-SLOA mimics the Hierarchical Particle Swarm Optimization (HPSO) in attacking phase and Adaptive Differential Evolution (ADE) in the reproduction process. Thus, making the approach a highly convergent once. In addition to this, the dimensions of the selected features have been reduced using the new Multi-kernel PCA(MKPCA) with Nyström Approximation approach. The improved correlation as well as Joint Mutual Information Maximisation based features have been newly considered along with the existing data features, to find the correlation between the variables. Together, the entire model filled with advanced approaches has successively surpassed the traditional approaches, when validated with standard database. The accuracy recorded by the proposed model is 97.8% for paysim database.