Deepfake detection IEEE and CVPR papers: where the research is published

This page belongs to the deepfake detection hub and uses one table, from the DeepfakeBench benchmark, as a window onto where the field publishes. It counts the detectors by venue and year, lists where the main datasets appeared and checks whether a venue or a year predicts a score.

01Which venues do the benchmarked detectors cite?

Table 1 of the DeepfakeBench paper lists each compared detector with a backbone, a repository and a reference. The reference column gives a venue and year. Counting it gives the table below.

CVPR alone accounts for six of the 15, and CVPR, ICCV and ECCV together for nine of the 13 detectors dated 2019 or later. The two oldest rows come from a forensics workshop: MesoNet's authors report it as "accepted to WIFS 2018". The SPSL authors name their venue in full as the "IEEE Conference on Computer Vision and Pattern Recognition (CVPR)". No row in the table is a journal article; all 15 references are conference or workshop papers.

02How do the counts change by year?

The detector years run 2018 (two), 2019 (four), 2020 (four), 2021 (two), 2022 (two) and 2023 (one). The shape reflects a selection, not a census, and the next section says why. Within it, the move from a forensics-specific workshop in 2018 to the general computer vision conferences from 2019 is clear: after the two WIFS rows, every detector cites CVPR, a CVPR workshop, ICCV, ECCV, ICML or ICASSP.

The year in the table is the year of the reference, which is not always the year of the idea. Xception is cited to ICCV 2019 because the benchmark takes the row from the FaceForensics++ repository and paper, which used XceptionNet as the detector it benchmarked. EfficientNet-B4 is cited to ICML 2019 because that is where the EfficientNet architecture was published, so the row records a backbone's venue rather than a forensics method's.

03What does the table not count?

It counts what the benchmark authors chose to compare. The original DeepfakeBench release has 15 detectors and nine datasets, and the current repository supports 36 detectors, 28 for images and 8 for video. A list that grows from 15 to 36 shows how much of the field one table leaves out.

Five rows carry a note in the repository column: Face X-ray, UCF, F3Net, SPSL and SRM read "Unpublished code, reproduced by us". The caption explains: "For detectors without publicly available repositories, we undertake careful re-implementation." A venue label on those rows says where the method was published, not who wrote the code that produced the benchmark score. The read a detection paper page lists the checks that follow from this, starting with whether the code was released.

04Where were the main datasets published?

Datasets have their own venues. The table lists the dataset papers cited elsewhere on this site, with the venue each paper's authors or repository report.

Four of the nine appeared at CVPR or ICCV, two in the NeurIPS Datasets and Benchmarks track, one at ACM Multimedia and two as preprints. The two NeurIPS rows are the two most recent benchmark frameworks, which suggests where benchmark papers have gone since 2023 but is a pattern in two cases. The datasets these papers introduce are compared on size and rights in their own page.

05Does the venue or the year predict the score?

No. DeepfakeBench Table 3 scores every detector trained on FaceForensics++ at c23, using frame-level AUC, on its own test split and on eight other datasets. The table below picks seven detectors across the venues.

The best cross-dataset average belongs to SPSL, a CVPR 2021 paper, at 0.7875, and the best score on the training distribution to UCF, an ICCV 2023 paper, at 0.9705. The newest detector in the table is therefore not the best on unseen sets: SPSL leads UCF by 0.0074 and leads the 2019 Xception baseline by 0.0157. The benchmark's authors note that a plain Xception classifier registers "high AUC values that are comparable to more sophisticated algorithms". The SPSL detector and UCF detector pages cover the two methods, and the cross-dataset generalisation page explains why the second column falls so far below the first.

The comparison is one benchmark, one training set and one metric, so it shows only that these scores do not order by year. The scores each detector reached are on the results page with their conditions.

06How to read a venue label in a results table

1. Match the year to the reference, not to the repository: the paper, the code and the benchmark run can be years apart.
2. Separate main-conference rows from workshop rows. CVPRW and CVPR are different proceedings.
3. Check whether the row is a forensics method or a general backbone, such as EfficientNet-B4.
4. Check whether the benchmark used the authors' code or its own re-implementation.
5. Never rank detectors by venue. Rank them by a score measured under the same training data and test sets.