Circumventing Shortcuts in Audio-visual Deepfake Detection Datasets with Unsupervised Learning | Read Paper on Bytez