Robust Deepfake Detection and Resilient Adversarial Image Reconstruction with Reduced Features Set

The rapid advancements in deepfake technology pose significant challenges in detecting manipulated media. This research introduces a feature extraction and selection method to address this threat. An optimally integrated pre-trained model is introduced to extract features from face images, composed of three CNN models —DenseNet-121, EfficientNet-B0, and ResNet-18 —and fine-tuned on the Celeb-DF (V2) dataset. These features are stacked for diverse representations, and a novel Assimilation-Elimination (ASEL) selection algorithm is used to minimize redundancies. The selected features are then fed into a KNN classifier to determine if an image is real or manipulated. Experiments on Celeb-DF (V2) achieve an AUC score of 97.77%, confirming the model's robustness. Additionally, reduced feature sets from real images are transmitted over a noisy communication channel, optimizing storage and bandwidth needs. Afterwards, the recovered bits are fed into the proposed error-resilient Feature-Driven Adversarial Image Reconstruction (FDAIR) model, achieving image reconstruction comparable to state-of-the-art methods.