DeepConvIEHO-AdaMaxNet: Secure and Robust Deepfake Detection using Optimized Deep Neural Networks

Advances in artificial intelligence (AI) and deep learning (DL) have enabled sophisticated Deepfake technology, which creates highly realistic fake videos by mapping one person's expressions onto another's face. While these innovations raise concerns about privacy, trust, and misuse in politics and security, they underscore the urgent need for robust detection and prevention measures to protect digital integrity. This paper addresses the challenges of deepfake detection, including high computational demands, difficulty capturing temporal inconsistencies, and susceptibility to adversarial attacks, by proposing a novel framework, DeepConvIEHO-AdaMaxNet. The framework leverages a Gaussian-Adaptive Bilateral Filter (GABF) for noise reduction and edge preservation during preprocessing, followed by a Convolutional neural network (CNN)-based classification optimised using an Improved Elephant Herding Optimization (IEHO) algorithm integrated with the AdaMax optimiser. This bio-inspired approach enhances the hyperparameter tuning process, overcoming limitations such as slow convergence and vulnerability to local minima in traditional methods. Additionally, the framework incorporates adversarial training within federated learning (FL) paradigm to improve robustness against manipulations. The suggested approach offers 96.41% of an area under the receiver operating characteristic curve (AUC), and 97.50% accuracy, as well as for class 0 achieve 100% specificity, 90% recall, 100% precision, and 95% F1-measure and class 1 achieves 93.53% specificity, 90.50% sensitivity, 100% recall, 90% precision, and 95% F1-measure using the Celeb-DF dataset. Evaluated through comprehensive metrics, the proposed model demonstrates superior performance compared to conventional CNNs, showcasing its potential in reliable deepfake detection.