The proliferation of deepfake videos poses a significant threat to the integrity of digital content and challenges the reliability and authenticity of the social media content. With the rise of sophisticated deepfake technologies, the detection of manipulated videos has become very challenging. Also, this has sparked worries about the security dangers posed by deepfakes. The identification and classification of image forgeries have emerged as critical issues in the field of multimedia forensics. When tested across different datasets, current deepfake detectors frequently struggle to retain their performance, even when faced with high-quality faces in particular datasets. To overcome such, we proposed an ensemble approach for deepfake detection using the power of enhanced convolution neural network i.e. CSP-Dencenet and LSTM network, a hybrid multi-task strategy to enhance feature maps for tasks related to deepfake detection and classification. CSP-DenceNet is introduced as a foundational element, combining DenceNet’s efficiency with the Cross Stage Partial (CSP) attention mechanism. This architecture is adept at extracting discriminative features crucial for identifying subtle manipulations inherent in deepfake videos. The fusion of CSP-DenceNet with LSTM further enhances the model’s capacity to discern temporal dependencies within video sequences, effectively capturing dynamic patterns indicative of deepfake content. Through comprehensive experimentation on benchmark dataset ie. Faceforensic++, the efficacy of the proposed hybrid model is validated, showcasing superior performance compared to conventional detection methods in detecting diverse deepfake contents i.e. facial swaps, lip-sync alterations, and facial reenactments etc. The proposed method is robust under different compression factors and provides state-of-the-art performance with 96.0% classification accuracy on the FaceForensics++ datasets, based on experimental results.