2026 |
Pahor, Jure; Peer, Peter; Štruc, Vitomir; Batagelj, Borut Less Learning, More Detection: Identity Document Presentation Attack Detection with Frozen Features Proceedings Article V: Proceedings of the IEEE International Joint Conference on Biometrics (IJCB 2026), str. 1–9, 2026. Povzetek | Povezava | BibTeX | Oznake: forgery detection, identity cards, information forensics, media forensics @inproceedings{LukaIJCB2026,Remote identity verification systems are increasingly exposed to physical presentation attacks (PAs), including print, screen, and composite attacks on identity documents. Detecting such attacks, commonly referred to as identity document presentation attack detection (PAD), is challenging because the relevant forensic cues are often local, subtle, and sensitive to acquisition conditions. Recent PAD methods therefore rely on increasingly sophisticated deep learning pipelines, including task-specific heads, learned patch fusion modules, and backbone adaptation. In this work, we revisit identity document PAD from the perspective of feature reuse and ask whether frozen self-supervised foundation model representations already encode the local forensic cues needed for reliable detection. We propose PADCore, a lightweight patch-based framework that decomposes each document into local patches, embeds them using a frozen DINOv2 (or DINOv3) backbone, constructs compact per-class coresets of representative embeddings, and fits a Linear Discriminant Analysis classifier on top. PADCore requires no backbone fine-tuning and can be efficiently re-fitted as new attack data becomes available. On FakeIDet2-db, PADCore matches or surpasses the state-of-the-art FakeIDet2 approach in in-distribution evaluation and achieves stronger generalization on KID34K, while reducing training time from hours to under four minutes. Its patch-based design is also compatible with privacy-preserving workflows based on pre-extracted anonymized document patches. |
Brodarič, Marko; Scheirer, Walter; Jain, Deepak Kumar; Peer, Peter; Štruc, Vitomir Learning Forgery Signals from Synthetic Depth Targets for Generalizable Deepfake Detection Proceedings Article V: Proceedings of the IEEE International Joint Conference on Biometrics (IJCB 2026), str. 1–10, 2026. Povzetek | Povezava | BibTeX | Oznake: deepfake, deepfake DAD, deepfake detection, deepfakes, information forensics, media forensics @inproceedings{BrodaricIJCB2026,Deepfake detection remains challenging under distribution shifts caused by unseen generation pipelines, post-processing, and unconstrained capture conditions. While most existing detectors primarily reason in RGB appearance space, facial manipulations also disturb geometric structure and consistency, making depth an attractive complementary cue. Existing depth-aware methods, however, typically use depth as an auxiliary input or guidance signal and may therefore underutilize the representational richness of modern monocular depth foundation models. To address this limitation, we propose a mechanism that adapts a pretrained monocular depth foundation model into a forgery-aware representation extractor for deepfake detection. Based on the adapted model, we then propose a novel depth-driven detector, termed {FADepth}, that uses the resulting depth representation as the primary source of evidence and complements it with RGB appearance cues. Extensive experiments on six benchmark datasets and with 24 state-of-the-art detectors show that FADepth yields highly competitive performance, achieving an overall mAUC of 88.70. The source code of the model is available at https://github.com/markobrodaric/FADepth |
Hrvatič, Anja; Peer, Peter; Štruc, Vitomir; Batagelj, Borut I-SPAI: A Human Perception Dataset for Deepfake Detection and Perception-Aware Auxiliary Supervision Proceedings Article V: Proceedings of the IEEE International Joint Conference on Biometrics (IJCB 2026), str. 1–10, 2026. Povzetek | Povezava | BibTeX | Oznake: deepfake, deepfake DAD, deepfake detection, deepfakes, media forensics @inproceedings{AnjaIJCB2026,Deepfake detectors typically overfit to low-level artifacts of the generative models seen during training, leading to poor generalization on unseen manipulation techniques. This work explores human visual perception as an alternative and more stable source of supervision. We introduce I-SPAI (textbf{I}nconsistency, textbf{S}aliency, and textbf{P}erception textbf{A}nnotations for Deepfake textbf{I}dentification), the first publicly available deepfake dataset with human perception annotations, consisting of 2,880 SBI-generated image pairs labeled by 435 participants with inconsistency types, spatial saliency maps, and confidence scores. Using I-SPAI, we propose a multitask detection framework and conduct a systematic evaluation of three types of human perceptual supervision: categorical inconsistency labels, spatial saliency map prediction, and indirect saliency-based loss supervision via CYBORG. Results on four benchmark datasets show that not all human perception signals are equally effective. Spatial saliency maps incorporated via a consecutive decoder provide the most consistent cross-dataset improvement, while categorical and indirect supervision offer limited gains. This suggests that spatial localization is the most transferable form of human perceptual knowledge for deepfake detection, and that the form of the supervision signal matters as much as its source. |
Larue, Nicolas; Štruc, Vitomir; Peer, Peter; Vu, Ngoc-Son Learning the Manifold of Authenticity: Hybrid-Curvature Representation Learning for Generalizable Deepfake Detection Članek v strokovni reviji V: IEEE Access, str. 1–14, 2026, ISBN: 2169-3536. Povzetek | Povezava | BibTeX | Oznake: deep learning, deepfake, deepfake DAD, deepfake detection, hyperbolic learning, media forensics @article{AccessHyperbolic,The practical utility of deepfake detectors is crippled by a crisis of generalization: models that perform well on known manipulation techniques consistently fail when faced with unseen forgeries.We argue this failure stems from a fundamental geometric mismatch. Existing methods implicitly assume that the manifold of authentic faces can be modeled in a space of uniform curvature, typically Euclidean, which inade-quately captures the complex, multi-scale structure of facial features. This paper validates the hypothesis that authentic faces lie on a manifold whose geometry is inherently hybrid, requiring both angular compactness (a spherical property) and hierarchical organization (a hyperbolic property). To resolve this geometric mismatch, we introduce a novel detector, CTrue, that learns a unified, hybrid-curvature representation of facial authenticity. Trained exclusively on real faces via self-supervised learning, our method simultaneously projects facial embeddings onto two complementary manifolds: a hypersphere to enforce compactness and a hyperbolic space to model the natural feature hierarchy. A single set of mathematically-optimal prototypes acts as a ‘‘geometric bridge’’, unifying the learning objectives in both spaces. At inference, a composite score measures an embedding’s deviation from this learned manifold. On challenging cross-dataset and cross-manipulation benchmarks, our method achieves competitive generalization under a strictly pristine-only training setting, showing that hybrid-curvature representations provide an effective and data-efficient alternative for deepfake detection. |
Marić, Nikola; Ivanovska, Marija; Štruc, Vitomir Exploring Multimodal Large Language Models for Morphing Attack Detection Proceedings Article V: Proceedings of the 29th Computer Vision Winter Workshop, str. 1-10, 2026. Povzetek | Povezava | BibTeX | Oznake: biometrics, face analysis, face morphing, face morphing attack, face morphing detection, media forensics @inproceedings{CVWW2026,Existing single-image morphing attack detection (S-MAD) systems often suffer from poor cross-dataset generalization and operate as opaque “black boxes,” which is particularly problematic in high-stakes border control scenarios. This paper investigates the adoption of opensource Multimodal Large Language Models (MLLMs) for S-MAD under strict cross-dataset evaluation through two different approaches. First, we assess selected MLLMs in zero-shot settings using a structured forensic prompting framework, which elicits multi-step semantic analysis with human-readable regional attributions. Second, leveraging the lightweight and parameter-efficient LoRA approach and a synthetic training dataset of morphs, we adapt the best-performing MLLM to the morphing attack detection (MAD) task in an efficient, generalizable, and privacy-preserving manner, enhancing the model’s sensitivity to diverse morphing artifacts. Our experimental results show that the proposed prompting strategy significantly improves overall attack detection accuracy compared to naive prompting. Moreover, our LoRA-adapted MLLM, Gemma-3 12B, achieves an average equal error rate (EER) of 14.81% across various morphing attack benchmarks, outperforming widely used MAD models. |
2025 |
Soltandoost, Elahe; Plesh, Richard; Schuckers, Stephanie; Peer, Peter; Struc, Vitomir Extracting Local Information from Global Representations for Interpretable Deepfake Detection Proceedings Article V: Proceedings of IEEE/CFV Winter Conference on Applications in Computer Vision - Workshops (WACV-W) 2025, str. 1-11, Tucson, USA, 2025. Povzetek | Povezava | BibTeX | Oznake: CNN, deepfake DAD, deepfakes, faceforensics++, media forensics, xai @inproceedings{Elahe_WACV2025,The detection of deepfakes has become increasingly challenging due to the sophistication of manipulation techniques that produce highly convincing fake videos. Traditional detection methods often lack transparency and provide limited insight into their decision-making processes. To address these challenges, we propose in this paper a Locally-Explainable Self-Blended (LESB) DeepFake detector that in addition to the final fake-vs-real classification decision also provides information, on which local facial region (i.e., eyes, mouth or nose) contributed the most to the decision process.~At the heart of the detector is a novel Local Feature Discovery (LFD) technique that can be applied to the embedding space of pretrained DeepFake detectors and allows identifying embedding space directions that encode variations in the appearance of local facial features. We demonstrate the merits of the proposed LFD technique and LESB detector in comprehensive experiments on four popular datasets, i.e., Celeb-DF, DeepFake Detection Challenge, Face Forensics in the Wild and FaceForensics++, and show that the proposed detector is not only competitive in comparison to strong baselines, but also exhibits enhanced transparency in the decision-making process by providing insights on the contribution of local face parts in the final detection decision. |
2024 |
Brodarič, Marko; Peer, Peter; Struc, Vitomir Towards Improving Backbones for Deepfake Detection Proceedings Article V: Proceedings of ERK 2024, str. 1-4, 2024. BibTeX | Oznake: CNN, deep learning, deepfake detection, deepfakes, media forensics, transformer @inproceedings{ERK_2024_Deepfakes, |
Brodarič, Marko; Peer, Peter; Štruc, Vitomir Cross-Dataset Deepfake Detection: Evaluating the Generalization Capabilities of Modern DeepFake Detectors Proceedings Article V: Proceedings of the 27th Computer Vision Winter Workshop (CVWW), str. 1-10, 2024. Povzetek | Povezava | BibTeX | Oznake: data integrity, deepfake, deepfake detection, deepfakes, difussion, face, faceforensics++, media forensics @inproceedings{MarkoCVWW,Due to the recent advances in generative deep learning, numerous techniques have been proposed in the literature that allow for the creation of so-called deepfakes, i.e., forged facial images commonly used for malicious purposes. These developments have triggered a need for effective deepfake detectors, capable of identifying forged and manipulated imagery as robustly as possible. While a considerable number of detection techniques has been proposed over the years, generalization across a wide spectrum of deepfake-generation techniques still remains an open problem. In this paper, we study a representative set of deepfake generation methods and analyze their performance in a cross-dataset setting with the goal of better understanding the reasons behind the observed generalization performance. To this end, we conduct a comprehensive analysis on the FaceForensics++ dataset and adopt Gradient-weighted Class Activation Mappings (Grad-CAM) to provide insights into the behavior of the evaluated detectors. Since a new class of deepfake generation techniques based on diffusion models recently appeared in the literature, we introduce a new subset of the FaceForensics++ dataset with diffusion-based deepfake and include it in our analysis. The results of our experiments show that most detectors overfit to the specific image artifacts induced by a given deepfake-generation model and mostly focus on local image areas where such artifacts can be expected. Conversely, good generalization appears to be correlated with class activations that cover a broad spatial area and hence capture different image artifacts that appear in various part of the facial region. |
Ivanovska, Marija; Štruc, Vitomir On the Vulnerability of Deepfake Detectors to Attacks Generated by Denoising Diffusion Models Proceedings Article V: Proceedings of WACV Workshops, str. 1051-1060, 2024. Povzetek | Povezava | BibTeX | Oznake: deep learning, deepfake, deepfake detection, diffusion models, face, media forensics @inproceedings{MarijaWACV24,The detection of malicious deepfakes is a constantly evolving problem that requires continuous monitoring of detectors to ensure they can detect image manipulations generated by the latest emerging models. In this paper, we investigate the vulnerability of single–image deepfake detectors to black–box attacks created by the newest generation of generative methods, namely Denoising Diffusion Models (DDMs). Our experiments are run on FaceForensics++, a widely used deepfake benchmark consisting of manipulated images generated with various techniques for face identity swapping and face reenactment. Attacks are crafted through guided reconstruction of existing deepfakes with a proposed DDM approach for face restoration. Our findings indicate that employing just a single denoising diffusion step in the reconstruction process of a deepfake can significantly reduce the likelihood of detection, all without introducing any perceptible image modifications. While training detectors using attack examples demonstrated some effectiveness, it was observed that discriminators trained on fully diffusion–based deepfakes exhibited limited generalizability when presented with our attacks. |
2023 |
Larue, Nicolas; Vu, Ngoc-Son; Štruc, Vitomir; Peer, Peter; Christophides, Vassilis SeeABLE: Soft Discrepancies and Bounded Contrastive Learning for Exposing Deepfakes Proceedings Article V: Proceedings of the International Conference on Computer Vision (ICCV), str. 21011 - 21021, IEEE 2023. Povzetek | Povezava | BibTeX | Oznake: CNN, deepfake detection, deepfakes, face, media forensics, one-class learning, representation learning @inproceedings{NicolasCCV,Modern deepfake detectors have achieved encouraging results, when training and test images are drawn from the same data collection. However, when these detectors are applied to images produced with unknown deepfake-generation techniques, considerable performance degradations are commonly observed. In this paper, we propose a novel deepfake detector, called SeeABLE, that formalizes the detection problem as a (one-class) out-of-distribution detection task and generalizes better to unseen deepfakes. Specifically, SeeABLE first generates local image perturbations (referred to as soft-discrepancies) and then pushes the perturbed faces towards predefined prototypes using a novel regression-based bounded contrastive loss. To strengthen the generalization performance of SeeABLE to unknown deepfake types, we generate a rich set of soft discrepancies and train the detector: (i) to localize, which part of the face was modified, and (ii) to identify the alteration type. To demonstrate the capabilities of SeeABLE, we perform rigorous experiments on several widely-used deepfake datasets and show that our model convincingly outperforms competing state-of-the-art detectors, while exhibiting highly encouraging generalization capabilities. The source code for SeeABLE is available from: https://github.com/anonymous-author-sub/seeable. |
Objave
2026 |
Less Learning, More Detection: Identity Document Presentation Attack Detection with Frozen Features Proceedings Article V: Proceedings of the IEEE International Joint Conference on Biometrics (IJCB 2026), str. 1–9, 2026. |
Learning Forgery Signals from Synthetic Depth Targets for Generalizable Deepfake Detection Proceedings Article V: Proceedings of the IEEE International Joint Conference on Biometrics (IJCB 2026), str. 1–10, 2026. |
I-SPAI: A Human Perception Dataset for Deepfake Detection and Perception-Aware Auxiliary Supervision Proceedings Article V: Proceedings of the IEEE International Joint Conference on Biometrics (IJCB 2026), str. 1–10, 2026. |
Learning the Manifold of Authenticity: Hybrid-Curvature Representation Learning for Generalizable Deepfake Detection Članek v strokovni reviji V: IEEE Access, str. 1–14, 2026, ISBN: 2169-3536. |
Exploring Multimodal Large Language Models for Morphing Attack Detection Proceedings Article V: Proceedings of the 29th Computer Vision Winter Workshop, str. 1-10, 2026. |
2025 |
Extracting Local Information from Global Representations for Interpretable Deepfake Detection Proceedings Article V: Proceedings of IEEE/CFV Winter Conference on Applications in Computer Vision - Workshops (WACV-W) 2025, str. 1-11, Tucson, USA, 2025. |
2024 |
Towards Improving Backbones for Deepfake Detection Proceedings Article V: Proceedings of ERK 2024, str. 1-4, 2024. |
Cross-Dataset Deepfake Detection: Evaluating the Generalization Capabilities of Modern DeepFake Detectors Proceedings Article V: Proceedings of the 27th Computer Vision Winter Workshop (CVWW), str. 1-10, 2024. |
On the Vulnerability of Deepfake Detectors to Attacks Generated by Denoising Diffusion Models Proceedings Article V: Proceedings of WACV Workshops, str. 1051-1060, 2024. |
2023 |
SeeABLE: Soft Discrepancies and Bounded Contrastive Learning for Exposing Deepfakes Proceedings Article V: Proceedings of the International Conference on Computer Vision (ICCV), str. 21011 - 21021, IEEE 2023. |