Detector results on deepfakes found in the wild

How Deepfake-Eval-2024 was assembled

Most deepfake benchmarks are built by generating manipulated media on purpose, with known tools, under known conditions. Deepfake-Eval-2024 was assembled the other way round, from in-the-wild deepfakes collected from social media and from users of a deepfake detection platform during 2024 [1]. It contains 45 hours of video, 56.5 hours of audio and 1,975 images, drawn from 88 different websites in 52 different languages [1].

The distinction is the entire point of the exercise. A laboratory benchmark measures whether a detector can recognise the output of the specific tools used to build it. A collection of material that was actually circulating measures something closer to the question a journalist or an investigator is really asking.

What the reported detector numbers do on Deepfake-Eval-2024

The authors state directly that academic benchmarks are out of date and not representative of real-world deepfakes [1]. Their measurement is that the performance of open-source state-of-the-art detection models drops precipitously when evaluated on Deepfake-Eval-2024, with AUC decreasing by 50 percent for video, 48 percent for audio and 45 percent for image models compared with previous benchmarks [1].

Two cautions belong with those figures. The first is that AUC is a ranking measure rather than a share of files correctly classified, so a fall in AUC is not the same quantity as a fall in accuracy, and the two should not be quoted interchangeably. The second is that the percentages above are relative changes against each model's own earlier benchmark result, which is how the authors express them, rather than absolute differences in the score.

Commercial and finetuned models, and the ceiling above them

The same work evaluates commercial detection models and models finetuned on Deepfake-Eval-2024, and finds they have superior performance to off-the-shelf open-source models, but do not yet reach the accuracy of deepfake forensic analysts [1]. That last clause is the one worth carrying around. The comparison the authors draw is not between a good detector and a bad one; it is between every model they tested and a trained human examiner, and the human is still ahead.

Tuning changes the detector picture more than architecture does

A second paper revisits the same benchmark and complicates the simple reading that open-source detection is hopeless. Initial reporting on Deepfake-Eval-2024 showed three finetuned open-source models achieving accuracies between 61 and 69 percent, against a leading commercial detector at 82 percent accuracy [2]. Revisiting one of those baseline approaches, which adapts standard pretrained vision backbones, the authors show that with better-tuned hyperparameters the same simple method reaches 81 percent accuracy on Deepfake-Eval-2024, surpassing the previously reported accuracy of that baseline by 18 percent and competing with commercial detectors [2].

Read together, the two papers say something more specific than "detectors do not work". A large part of the reported gap between open and commercial systems, on this benchmark, was a gap in tuning rather than in method. That is encouraging about the ceiling and discouraging about the published figures, because it means a headline number can move substantially without the underlying approach changing at all.

The second paper also frames the choice as a set of tradeoffs between accuracy, computational cost and interpretability [2]. Interpretability is the one most likely to matter to anyone who has to explain a result to a court, an editor or a client, and it does not appear in a single accuracy figure.

Reading a detector claim after these results

Three questions follow from the measurements above, and none of them require agreeing with any particular conclusion.

• Which benchmark produced the quoted figure. A number from a laboratory dataset and a number from in-the-wild material are not comparable, and the reported drop between them is large [1].
• Which metric the figure is. AUC and accuracy are different quantities, and both appear across this literature [1] [2].
• Whether the comparison being offered is against another tool or against a human examiner. On this benchmark, the models tested did not reach the accuracy of forensic analysts [1].

None of this makes detection useless. It makes a detector output a piece of evidence with known limits rather than an answer, which is a more modest role and a more defensible one.

---

• Why deepfake detectors fail on new generators , why performance collapses on an unfamiliar generator
• What the US AI Safety Institute says still needs research , what a standards body says is still unsolved
• What a deepfake detector score does not tell you , how to read a detector claim after these results

---

Sources

1. [1]
   Chandra, N. A., Lee, H., Murtfeldt, R., Qiu, L., Karmakar, A., et al. Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024. arXiv:2503.02857.arxiv.org
2. [2]
   Castaneda, O., So-Tang, K., Gurung, K. Revisiting Simple Baselines for In-The-Wild Deepfake Detection. arXiv:2509.04150.arxiv.org