<![CDATA[Deepfakes are synthetic audio-visual content that is artificially engineered that is increasingly difficult to distinguish from real content and is mostly circulated through social media platforms. The problem studied in this study is the decline in the accuracy of the deepfake detection model when the test content has gone through a series of social media compression processes, such as H.264 re-encoding and resolution degradation, which removes some of the forensic traces of the pixels that are the basis for detection. The objectives of this study were to measure the magnitude of the decline in detection accuracy in the lightweight architecture of MobileNetV3, to find the compression threshold where the decline began to be significant, and to test the augmentation of compressed data as a mitigation strategy. The method used was a quantitative experiment on the Celeb-DF v2 dataset with three levels of compression (raw, c23, and c40) simulated using FFmpeg; The MobileNetV3-Small model was trained on raw data and then tested at all three levels of compression, then compared to training scenarios using combined data of all three levels of compression. Test results on 5,180 face images of standard test data showed a decrease in accuracy from 70.23% in raw data to 69.54% in c23 and 58.11% in c40, with the sharpest decrease occurring between medium compression and high compression and mainly due to a drop in recall from 89.79% to 61.24%. Training scenarios with compressed data augmentation were shown to increase AUC at all levels of compression, from 59.51% to 64.29% at c40, although accompanied by a slight decrease in accuracy on raw data. The contribution of this study is empirical evidence of the compression resistance of lightweight architectures that are rarely studied as well as practical mitigation recommendations for the development of deepfake detection systems on mobile devices.]]>
