The generalization ability of deepfake detectors is vital for their applications in real-world scenarios. One effective solution to enhance this ability is to train the models with manually-blended data, which we termed ''blendfake'', encouraging models to learn generic forgery artifacts like blending boundary. Interestingly, current SoTA methods utilize blendfake $\textit{without}$ incorporating any deepfake data in their training process. This is likely because previous empirical observations suggest that vanilla hybrid training (VHT), which combines deepfake and blendfake data, results in inferior performance to methods using only blendfake data (so-called "1+1<2"). Therefore, a critical question arises: Can we leave deepfake behind and rely solely on blendfake data to train an effective deepfake detector? Intuitively, as deepfakes also contain additional informative forgery clues ($\textit{e.g.,}$ deep generative artifacts), excluding all deepfake data in training deepfake detectors seems counter-intuitive.

The generalization ability of deepfake detectors is vital for their applications in real-world scenarios. One effective solution to enhance this ability is to train the models with manually-blended data, which we termed "blendfake", encouraging models to

With the prevalence of artificial intelligence (AI)-generated content, such as audio deepfakes, a large body of recent work has focused on developing deepfake detection techniques. However, most models are evaluated on a narrow set of datasets, leaving their generalization to real-world conditions uncertain. In this paper, we systematically review 28 existing audio deepfake datasets and present an open-source benchmarking toolkit called AUDDT (https://github.com/MuSAELab/AUDDT). The goal of this toolkit is to automate the evaluation of pretrained detectors across these 28 datasets, giving users direct feedback on the advantages and shortcomings of their deepfake detectors. We start by showcasing the usage of the developed toolkit, the composition of our benchmark, and the breakdown of different deepfake subgroups. Next, using a widely adopted pretrained deepfake detector, we present in- and out-of-domain detection results, revealing notable differences across conditions and audio manipulation types. Lastly, we also analyze the limitations of these existing datasets and their gap relative to practical deployment scenarios.

The growing presence of AI-generated videos on social networks poses new challenges for deepfake detection, as detectors trained under controlled conditions often fail to generalize to real-world scenarios. A key factor behind this gap is the aggressive, proprietary compression applied by platforms like YouTube and Facebook, which launder low-level forensic cues. However, replicating these transformations at scale is difficult due to API limitations and data-sharing constraints. For these reasons, we propose a first framework that emulates the video sharing pipelines of social networks by estimating compression and resizing parameters from a small set of uploaded videos. These parameters enable a local emulator capable of reproducing platform-specific artifacts on large datasets without direct API access. Experiments on FaceForensics++ videos shared via social networks demonstrate that our emulated data closely matches the degradation patterns of real uploads. Furthermore, detectors fine-tuned on emulated videos achieve comparable performance to those trained on actual shared media. Our approach offers a scalable and practical solution for bridging the gap between lab-based training and real-world deployment of deepfake detectors, particularly in the underexplored domain of compressed video content.

Deepfakes powered by advanced machine learning models present a significant and evolving threat to identity verification and the authenticity of digital media. Although numerous detectors have been developed to address this problem, their effectiveness has yet to be tested when applied to real-world data. In this work we evaluate modern deepfake detectors, introducing a novel testing procedure designed to mimic real-world scenarios for deepfake detection. Using state-of-the-art deepfake generation methods, we create a comprehensive dataset containing more than 500,000 high-quality deepfake images. Our analysis shows that detecting deepfakes still remains a challenging task. The evaluation shows that in fewer than half of the deepfake detectors tested achieved an AUC score greater than 60%, with the lowest being 50%. We demonstrate that basic image manipulations, such as JPEG compression or image enhancement, can significantly reduce model performance. All code and data are publicly available at https://github.com/SumSubstance/Deepfake-Detectors-in-the-Wild.

We present the first large-scale open-set benchmark for multilingual audio-video deepfake detection. Our dataset comprises over 250 hours of real and fake videos across eight languages, with 60% of data being generated. For each language, the fake videos are generated with seven distinct deepfake generation models, selected based on the quality of the generated content. We organize the training, validation and test splits such that only a subset of the chosen generative models and languages are available during training, thus creating several challenging open-set evaluation setups. We perform experiments with various pre-trained and fine-tuned deepfake detectors proposed in recent literature. Our results show that state-of-the-art detectors are not currently able to maintain their performance levels when tested in our open-set scenarios. We publicly release our data and code at: https://huggingface.co/datasets/unibuc-cs/MAVOS-DD.

AI-generated deepfake videos depicting humans are getting more advanced, and more common, by the day. The most sophisticated tools can now produce manipulated content that is indistinguishable to the average human observer. Deepfake detectors, which use their own AI models to analyze video clips, attempt to bypass this deception by searching for hidden tells. One of those is the presence of a human pulse. In the past, AI models that detected a noticeable pulse or heart rate could confidently classify those clips as genuine.

In this paper, we introduce the idea of using adversarially-generated samples of the input images that were classified as deepfakes by a detector, to form perturbation masks for inferring the importance of different input features and produce visual explanations. We generate these samples based on Natural Evolution Strategies, aiming to flip the original deepfake detector's decision and classify these samples as real. We apply this idea to four perturbation-based explanation methods (LIME, SHAP, SOBOL and RISE) and evaluate the performance of the resulting modified methods using a SOTA deepfake detection model, a benchmarking dataset (FaceForensics++) and a corresponding explanation evaluation framework. Our quantitative assessments document the mostly positive contribution of the proposed perturbation approach in the performance of explanation methods. Our qualitative analysis shows the capacity of the modified explanation methods to demarcate the manipulated image regions more accurately, and thus to provide more useful explanations.

Deepfakes are on the rise, with increased sophistication and prevalence allowing for high-profile social engineering attacks. Detecting them in the wild is therefore important as ever, giving rise to new approaches breaking benchmark records in this task. In line with previous work, we show that recently developed state-of-the-art detectors are susceptible to classical adversarial attacks, even in a highly-realistic black-box setting, putting their usability in question. We argue that crucial 'robust features' of deepfakes are in their higher semantics, and follow that with evidence that a detector based on a semantic embedding model is less susceptible to black-box perturbation attacks. We show that large visuo-lingual models like GPT-4o can perform zero-shot deepfake detection better than current state-of-the-art methods, and introduce a novel attack based on high-level semantic manipulation. Finally, we argue that hybridising low- and high-level detectors can improve adversarial robustness, based on their complementary strengths and weaknesses.