Improved Deepfake Detection with Optimized Preprocessing for Low-Quality Images

Deepfake detection has been one of the fastest-growing areas of research due to the increasing threat manipulated media is posing to the world. One of the major challenges still underlying this problem, it remains poorly researched how to detect deepfakes in low-quality images[1]. In this paper, a framework is proposed to enhance the Mesonet model in the detection of deepfakes using the implementation of a preprocessing optimization layer. This layer refines the quality of input images through an enhanced preprocessing layer. In our experiments, deepfake images were generated from a DCGAN model trained on the CelebA dataset, simulating the real-world scenarios of low-quality deepfakes. Our framework was tested on a dataset comprising 1,000 low-quality real images and 1,000 low-quality deepfake images, also derived from the CelebA dataset. Accuracy was found to increase from 69.40% to 92.50% in the optimized model. Moreover, the area under the ROC curve increased from 0.75—75% to 0.96—96%, also improving the model’s discriminatory power very much. These results shed light on the proportion of preprocessing-related optimizations for performance improvement in deepfake detection. This work has contributed to the deepfake detection methodology with robust ways to improve model performance under very challenging conditions. While the research has significantly enhanced the detection which can be seen when dealing with low quality-wire etc., there are still some issues associated with the complexity of the approach. This extra requirement makes the demand for additional computing more critical. Intuitively, the researchers could direct future works towards seeking more flexible and efficient optimization without sacrificing the desirable accuracy and the ability to use limited resources.