Deepfake Detection with Contrastive Learning in Curved Spaces
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: deepfakes, forgery, detection
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Deepfake detectors excel in familiar scenarios but falter when faced with new generation techniques. Improving their generalization can be achieved through synthetic data during training or using one-class anomaly detection methods. However, existing techniques, limited to non-negative-curvature spaces, struggle to effectively identify counterfeit features on the intricate and diverse non-Euclidean human face manifold. Human faces defy simple Euclidean geometry due to their complexity. To overcome this limitation, we introduce a novel and efficient deepfake detector, called CTru, that learns a rich representation of facial geometry across multiple-curvature spaces in a self-supervised manner. During inference, the fakeness score is computed by integrating angle-based similarity in spherical space and model confidence in hyperbolic space with Busemann distance. CTru establishes new SoTA results on various challenging datasets in both cross-dataset and cross-manipulation scenarios, while being trained only on pristine faces, highlighting its impressive generalization performance. Code source will be made available.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2798
Loading