Deepfake-Eval-2024 Benchmark
paperAuthors
Credibility Rating
Good quality. Reputable source with community review or editorial standards, but less rigorous than peer-reviewed venues.
Rating inherited from publication venue: arXiv
Relevant to AI safety discussions around synthetic media, disinformation, and the gap between benchmark performance and real-world robustness of detection systems.
Paper Details
Metadata
Abstract
In the age of increasingly realistic generative AI, robust deepfake detection is essential for mitigating fraud and disinformation. While many deepfake detectors report high accuracy on academic datasets, we show that these academic benchmarks are out of date and not representative of real-world deepfakes. We introduce Deepfake-Eval-2024, a new deepfake detection benchmark consisting of in-the-wild deepfakes collected from social media and deepfake detection platform users in 2024. Deepfake-Eval-2024 consists of 45 hours of videos, 56.5 hours of audio, and 1,975 images, encompassing the latest manipulation technologies. The benchmark contains diverse media content from 88 different websites in 52 different languages. We find that the performance of open-source state-of-the-art deepfake detection models drops precipitously when evaluated on Deepfake-Eval-2024, with AUC decreasing by 50% for video, 48% for audio, and 45% for image models compared to previous benchmarks. We also evaluate commercial deepfake detection models and models finetuned on Deepfake-Eval-2024, and find that they have superior performance to off-the-shelf open-source models, but do not yet reach the accuracy of deepfake forensic analysts. The dataset is available at https://github.com/nuriachandra/Deepfake-Eval-2024.
Summary
Deepfake-Eval-2024 introduces a benchmark of in-the-wild deepfakes collected from social media in 2024, revealing that state-of-the-art open-source detectors suffer 45-50% AUC drops compared to academic benchmarks. The dataset spans 45 hours of video, 56.5 hours of audio, and 1,975 images across 52 languages from 88 websites. Commercial and finetuned models improve but still fall short of human forensic analysts.
Key Points
- •Existing academic deepfake benchmarks are outdated; open-source SOTA models see 45-50% AUC drops when evaluated on real-world 2024 deepfakes.
- •Dataset covers video (45 hrs), audio (56.5 hrs), and images (1,975) from 88 websites in 52 languages, representing latest manipulation technologies.
- •Commercial and finetuned models outperform off-the-shelf open-source models but still lag behind human deepfake forensic analysts.
- •The benchmark highlights a critical deployment gap: high lab accuracy does not translate to real-world detection performance.
- •Dataset is publicly available, enabling community-wide evaluation and improvement of deepfake detection systems.
Cited by 2 pages
| Page | Type | Quality |
|---|---|---|
| AI Content Authentication | Approach | 58.0 |
| AI-Era Epistemic Security | Approach | 63.0 |
Cached Content Preview
Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024
Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024
Nuria Alina Chandra
TrueMedia.org
Ryan Murtfeldt
TrueMedia.org
University of Washington, Seattle
Lin Qiu
TrueMedia.org
University of Washington, Seattle
Arnab Karmakar
TrueMedia.org
University of Washington, Seattle
Hannah Lee
TrueMedia.org
Emmanuel Tanumihardja
TrueMedia.org
University of Washington, Seattle
Kevin Farhat
TrueMedia.org
University of Washington, Seattle
Ben Caffee
TrueMedia.org
University of Washington, Seattle
Sejin Paik
TrueMedia.org
Georgetown University, Washington D.C.
Changyeon Lee
Miraflow AI
Yonsei University, Seoul
Jongwook Choi
TrueMedia.org
Chung-Ang University, Seoul
Aerin Kim
TrueMedia.org
Miraflow AI
Oren Etzioni
TrueMedia.org
University of Washington, Seattle
Abstract
In the age of increasingly realistic generative AI, robust deepfake detection is essential for mitigating fraud and disinformation. While many deepfake detectors report high accuracy on academic datasets, we show that these academic benchmarks are out of date and not representative of real-world deepfakes. We introduce Deepfake-Eval-2024, a new deepfake detection benchmark consisting of in-the-wild deepfakes collected from social media and deepfake detection platform users in 2024. Deepfake-Eval-2024 consists of 44 hours of videos, 56.5 hours of audio, and 1,975 images, encompassing the latest manipulation technologies. The benchmark contains diverse media content from 88 different websites in 52 different languages. We find that the performance of open-source state-of-the-art deepfake detection models drops precipitously when evaluated on Deepfake-Eval-2024, with AUC decreasing by 50% for video, 48% for audio, and 45% for image models compared to previous benchmarks. We also evaluate commercial deepfake detection models and models finetuned on Deepfake-Eval-2024, and find that they have superior performance to off-the-shelf open-source models, but they do not yet reach the accuracy of human deepfake forensic analysts. The dataset is available at https://github.com/nuriachandra/Deepfake-Eval-2024 .
Figure 1: Examples of Deepfake-Eval-2024 video and audio (rows 1–2), and images (rows 3–4), demonstrating a diversity of content styles and generation techniques, including lipsync, faceswap, and diffusion. Images have been resized for presentation.
1 Introduction
Advances in generative AI models have precipitated a surge of highly realistic deepfakes, which have been used to fabricate messages from politicians [ 1 ] , create non-consensual pornographic content [ 2 ] , spread misinformation [ 3 ] , and damage reputations [ 4 ] , harming lives, businesses, and nations [ 5 ] . Between 2023 and 2024
... (truncated, 67 KB total)919c9ed9593285fd | Stable ID: sid_mJnPIH9kBR