Unmasking the Unknown: Facial Deepfake Detection in the Open-Set Paradigm

TL;DR

This work tackles the limitation of closed-set deepfake detectors by framing detection as an open-set problem, where unseen forgery methods are flagged as unknown. It introduces a three-stage pipeline built on a weighted supervised contrastive learning loss, an SWA-augmented encoder, and a threshold-based open-set detector, demonstrating strong performance on FaceForensics++c23 and competitive cross-dataset results on CelebDFv2. Key contributions include a novel loss formulation that emphasizes real-data clustering, empirical evidence of improved unknown detection (AUROC) and state-of-the-art True Open-Set Classification (TOSC) scores, and explainability analyses via UMAP and Grad-CAM. The approach offers a robust, scalable tool for detecting both known and emerging deepfake techniques, with practical implications for forensic analysis and security. Future work may address domain shift, video-level modeling, and localized priors to further enhance robustness.

Abstract

Facial forgery methods such as deepfakes can be misused for identity manipulation and spreading misinformation. They have evolved alongside advancements in generative AI, leading to new and more sophisticated forgery techniques that diverge from existing 'known' methods. Conventional deepfake detection methods use the closedset paradigm, thus limiting their applicability to detecting forgeries created using methods that are not part of the training dataset. In this paper, we propose a shift from the closed-set paradigm for deepfake detection. In the open-set paradigm, models are designed not only to identify images created by known facial forgery methods but also to identify and flag those produced by previously unknown methods as 'unknown' and not as unforged/real/unmanipulated. In this paper, we propose an open-set deepfake classification algorithm based on supervised contrastive learning. The open-set paradigm used in our model allows it to function as a more robust tool capable of handling emerging and unseen deepfake techniques, enhancing reliability and confidence, and complementing forensic analysis. In open-set paradigm, we identify three groups including the "unknown group that is neither considered known deepfake nor real. We investigate deepfake open-set classification across three scenarios, classifying deepfakes from unknown methods not as real, distinguishing real images from deepfakes, and classifying deepfakes from known methods, using the FaceForensics++ dataset as a benchmark. Our method achieves state of the art results in the first two tasks and competitive results in the third task.

Figures (5)

• Figure 1: This figure illustrates the latent space and decision boundaries of both the closed-set and open-set paradigms. (A) In the case of a closed-set classifier, the model learns a feature space and decision boundaries based solely on the known deepfake techniques and real data, distinguishing between Real and Known Deepfakes. (B) However, this approach cannot identify samples from previously unseen techniques. In contrast, open-set recognition establishes more flexible decision boundaries around the known classes and learns more general features.
• Figure 2: This figure shows the proposed methodology. (a) Supervised contrastive learning is used on the dataset to learn representations of various deepfake/real images via an encoder. (b) During the second stage, the encoder is frozen and a classifier is learnt on top of it. (c) In this phase, the logits of the training samples are employed to establish thresholds, which are then utilized for detecting unknown classes.
• Figure 3: Sample face images from different facial forgery methods utilized in this study. The first column showcases pristine (non-manipulated) frames, while the subsequent four columns depict images generated by DeepFakes (DF), Face2Face (F2F), FaceSwap (FS), and Neural Textures (NT). DF and FS are identity swapping methods whereas NT and F2F are Facial Re-enactment methods
• Figure 4: Class activation maps produced for images created by known and unknown facial forgeries for two different cases of unknown test classes. Notably, the images with unknown facial forgeries show multiple clusters of activations.
• Figure 5: This figure shows the TSNE visualizations of the representations learned by various methods