<p>This research work proposes the use of a hybrid deep learning model for identifying cases of fraud in financial and online payment platforms. The proposed method improves the detection of suspicious transactions by combining deep learning technologies with Meta-heuristic optimization approach. The current approaches to fraud detection have some drawbacks like high percentage of false positives, inability to adjust to changes in fraud schemes, and unsuitability for processing massive amounts of real-time data. To overcome these limitations, a new approach of multiple levels feature extraction is introduced to extract low level features and relationships between different levels of the data. Furthermore, a novel hybrid Stallion Clash Optimization (SCO) is used to select the best features that improve the model’s accuracy. The centre of the approach is the new FinShieldNet model, which is a combination of CNN, RBFN, and SA-BiLSTM networks. The final decision is made using an ensemble of voting classifier to make the decision-making process more reliable. Performance evaluation is conducted using metrics such as accuracy, sensitivity, and specificity on three benchmark datasets namely the Financial Fraud Detection Dataset, Online Payments Fraud Detection Dataset, and Credit Card Fraud Detection Dataset. The proposed approach obtained an accuracy of 94% and recall of 92.5% and thus better than previous methods.</p>