Development of Ensemble Strategies for Generalization in Deepfake Image Detection

Abstract

The growing accessibility of generative models has enabled the rapid proliferation of deepfake content, posing significant challenges in image-based biometric security and media authenticity. In this thesis, six diverse facial deepfake image datasets are assembled, and four modern detection models are evaluated in a cross-domain scenario. We observe that individual models fail to generalize to images generated by techniques outside the scope of their training data. This often hinders the applicability of a single model in real-world deepfake detection. This thesis proposes ensemble strategies as a means of addressing this lack of generalization. We find that the ensemble models outperform individual models in classifying deepfake images, particularly in terms of accuracy and recall. An exhaustive evaluation of combinations of models shows that ensembles of similar models provide limited benefit, whereas ensembles of complementary models lead to significant improvements in classification performance. Ensembling models based specifically on accuracy and recall metrics also produces models that lower the rate of more harmful false negative predictions. This work highlights the value of ensemble models in improving generalization across diverse image families and provides a framework for building robustness in real-world deepfake detection systems.