DeepFake-O-Meter open detection platform
Vendor University at Buffalo Media Forensics Lab Official site zinc.cse.buffalo.edu/ubmdfl/deep-o-meter How this card is sourced
Updated · AI Detector Index editors
DeepFake-O-Meter lists 41 detection methods on its detectors endpoint, 19 for images, 10 for video, 11 for audio and 1 for both image and video by the platform's own labels (zinc.cse.buffalo.edu, detectors endpoint, read 10 November 2026). It is not a product. The University at Buffalo Media Forensics Lab runs it as an open research platform that, in its own words, "aggregates results from state-of-the art research independent AI media detection models across image, video, and audio modalities", and its notice reads "All detection method implementations are intended solely for academic and individual use and are not consumer products" (zinc.cse.buffalo.edu, platform, read 10 November 2026). Each hosted detector returns its own score; the platform publishes no accuracy figure of its own. The card joins the detector index as its first academic platform, and how this card is sourced explains why a platform that hosts other people's models carries no claim to rank.
What DeepFake-O-Meter detects
| Field | Value | Source |
|---|---|---|
| Content types detected | Image, video, voice | zinc.cse.buffalo.edu/ubmdfl/deep-o-meter |
| Generators covered | The platform publishes no single generator list; each hosted detector states its own scope (detectors endpoint, read 10 November 2026), Image: CLIP-ViT (2023) for real v.s. Fake images generated by ProGAN, StyleGAN, BigGAN, CycleGAN, StarGAN, GauGAN, CRN, IMLE, SAN, SITD, DeepFakes, Latent diffusion model, Glide and DALL-E; NPR (2024) for ProGAN, StyleGAN, StyleGAN2, BigGAN and others; AIDE (2024) for diffusion model generated image detection; GLFF (2024) for GAN models and DALL-E model generated images; NoDown (2021) for GAN models, Video: AltFreezing (2023) trained on FaceForensics++ with four types of manipulation (Face2Face, FaceSwap, NeuralTexture, and Deepfake); DSP-FWA (2019) for face-swap videos; LIPINC (2024) and WAV2LIP-STA (2022) for lip-syncing videos, Audio: AASIST (2021) trained on ASV spoof dataset; RawNet2 (2021), RawNet3 (2023), LFCC-LCNN (2021), Whisper (2023), RawNet2-Vocoder (2025), SLS (2021), XLSR-Mamba (2025), Nes2Net (2025) for single-speaker clips | zinc.cse.buffalo.edu/ubmdfl/api/deep-o-me... |
| Input limits | Web platform with a login (Sign in to Submit Media; Please log in before testing the file; email verification code at registration). Drop zone: Supported formats: Images (JPG, PNG, BMP, TIFF), Video (MP4, AVI, MOV), Audio (WAV, MP3); a YouTube URL can be pasted instead of a file. The detectors endpoint lists per-detector extensions that add FLAC, OGG, WEBP, TIF, NEF and RAF, and per-detector limits: CFM (2025) Max file size: 50MB for videos, 10MB for images; AVSRDD (2025) Video must be at least 5 seconds long; most face detectors Only for deepfake videos with a single face, prefer frontal face facing without sunglasses or occlusion, with good lighting; audio detectors Input requires a single speaker and a clear voice background. Daily upload limits by tier (model_groups endpoint): regular_user 30 uploads a day, advanced_user 100, premium_user unlimited. The paper reports average running times of about 30 seconds for image and audio modules and about 90 seconds for video modules (Section III.B). | zinc.cse.buffalo.edu/ubmdfl/api/deep-o-me... |
| Free tier | Free with an account: the paper describes nonprofit services and the site calls the methods not consumer products; a regular_user account allows 30 uploads a day with 15 days of record keeping (model_groups endpoint, read 10 November 2026); the paper gives regular users 30 daily tasks and advanced users 300 (the live endpoint gives advanced users 100). No price is published for any tier | zinc.cse.buffalo.edu/ubmdfl/api/deep-o-me... |
| Integrations | Web platform at zinc.cse.buffalo.edu/ubmdfl/deep-o-meter/ (login required; English, French, simplified and traditional Chinese interface), YouTube URL submission (Paste a valid YouTube URL to see available detection methods), Casebook, a second route shown in the interface as not deployed yet (Casebook is not deployed yet. A secure server route is required.), The front end calls a backend API under /ubmdfl/api/deep-o-meter/; no public API documentation is published | zinc.cse.buffalo.edu/ubmdfl/deep-o-meter |
| Data retention | No privacy policy or terms page is published on the platform (read 10 November 2026). The submission history page reads Submissions are kept for {{policy}}, filled from the account's retention_policy, and the model_groups endpoint states record_keeping_days 15 for regular_user, 30 for advanced_user and unlimited for premium_user and admin. The account page lets a user Permanently delete your account and all associated data. This action cannot be undone. The upload form asks Do you want to share data with us?; the paper says users can choose whether they want to share their data for academic research purposes only (Section II.A). | zinc.cse.buffalo.edu/ubmdfl/api/deep-o-me... |
| Launch year | 2021 | arxiv.org/abs/2404.13146 |
| Ownership | Open research platform of the UB Media Forensic Laboratory (UB MDFL), a multi-departmental research lab at the School of Engineering and Applied Sciences, University at Buffalo, State University of New York, 301A Davis Hall, Buffalo, NY 14260, affiliated with the Department of Computer Science and Engineering and the Artificial Intelligence Institute; the lab names the National Science Foundation, DARPA, the National Institute of Justice, the Army Research Office, the US Department of Homeland Security and Google Inc. as funders (lab home page). The platform's about text reads This project is supported by the University at Buffalo and the National Science Foundation under Grant SaTC-2153112 and All detection method implementations are intended solely for academic and individual use and are not consumer products. The rights of individual detection methods belong to their original authors. Version 1.0 was presented in Li, Zhang, Sun, Ke, Ju, Qi and Lyu, Deepfake-o-meter: An open platform for deepfake detection, 2021 IEEE Security and Privacy Workshops; version 2.0 is described in Ju, Sun, Jia, Hou, Si, Datta, Ke, Zhou, Nikolich and Lyu, DeepFake-O-Meter v2.0: An Open Platform for DeepFake Detection (arXiv 2404.13146, submitted 19 April 2024, v2 27 June 2024), the platform's own paper. The site runs at zinc.cse.buffalo.edu under the lab's path with a Google Analytics tag. | ubmdfl.cse.buffalo.edu |
Three media types, through published research code. The detectors endpoint names each method, its year, its originating organisation and its scope: for images, CLIP-ViT (2023) from the University of Wisconsin-Madison "for classifying real v.s. Fake images generated by ProGAN, StyleGAN, BigGAN, CycleGAN, StarGAN, GauGAN, CRN, IMLE, SAN, SITD, DeepFakes, Latent diffusion model, Glide and DALL-E", AIDE (2024) "for diffusion model generated image detection" and NPR (2024); for video, AltFreezing (2023) "trained on lightly compressed FaceForensics++ (2021) dataset, containing four types of manipulation (Face2Face, FaceSwap, NeuralTexture, and Deepfake)", DSP-FWA (2019) for face swaps and LIPINC (2024) for lip-syncing; for audio, AASIST (2021) "trained on ASV spoof dataset", RawNet2, RawNet3, LFCC-LCNN, Whisper and the lab's own RawNet2-Vocoder (zinc.cse.buffalo.edu, detectors endpoint, read 10 November 2026). The platform publishes no single generator list, so the generator count on this card stays empty and the scope lines are quoted instead.
The counts move. The model groups endpoint puts 38 methods in the regular layer and 3 in a retired layer (DMimageDetection 2023, RawNet2-Vocoder 2021 and AVAD 2023), while the detectors endpoint returns 41 in all (zinc.cse.buffalo.edu, model groups endpoint, read 10 November 2026). A notice dated 11/24/2025 announces "three new image detection methods (DIRE, SBI-Image, UIA-ViT), one new video detection method (LipFD), and one new audio detection method (RawNet2-Vocoder-v2)", and the endpoint returns none of the first four by those names (zinc.cse.buffalo.edu, news endpoint, read 10 November 2026). The platform's own paper counted 6 image, 6 video and 5 audio detectors at version 2.0 in April 2024 (arXiv 2404.13146, Table I, read 10 November 2026).
The operator is the UB Media Forensic Laboratory at 301A Davis Hall, Buffalo, NY 14260, "a multi-departmental research lab at the School of Engineering and Applied Sciences, University at Buffalo, State University of New York", which names the National Science Foundation, DARPA, the National Institute of Justice, the Army Research Office, the US Department of Homeland Security and Google Inc. as funders (ubmdfl.cse.buffalo.edu, read 10 November 2026). The platform text adds "This project is supported by the University at Buffalo and the National Science Foundation under Grant SaTC-2153112". Version 1.0 was presented at the 2021 IEEE Security and Privacy Workshops and version 2.0 is described by Ju, Sun, Jia, Hou, Si, Datta, Ke, Zhou, Nikolich and Lyu in the platform's own paper (arXiv 2404.13146, read 10 November 2026). The sibling cards for the same media are Sensity, a commercial forensic platform, and Deepware Scanner, a free video scanner; the paper's Table I compares the platform with both.
Limits and plans
| Plan | Price | Limits | Checked on |
|---|---|---|---|
| Regular user | Free | 30 uploads a day; records kept 15 days; the regular model layer (38 detectors on 10 November 2026). | |
| Advanced user | Free | 100 uploads a day; records kept 30 days; the regular model layer; the tier is assigned by an administrator (the paper's figure for advanced users is 300 daily tasks). | |
| Premium user | Free | Unlimited uploads; unlimited record keeping; sees regular, new and retired models; assigned by an administrator. No price is published for any tier. |
A login is required. The platform's upload area reads "Sign in to Submit Media", a tip dated 07/09/2025 says "Please log in before testing the file", and registration sends a verification code to the user's email (zinc.cse.buffalo.edu, platform, read 10 November 2026). The drop zone states "Supported formats: Images (JPG, PNG, BMP, TIFF), Video (MP4, AVI, MOV), Audio (WAV, MP3)", and a YouTube URL can be pasted instead of a file. Per-detector limits add to that: CFM (2025) sets "Max file size: 50MB for videos, 10MB for images", AVSRDD (2025) requires that "Video must be at least 5 seconds long", and most face detectors are "Only for deepfake videos with a single face, prefer frontal face facing without sunglasses or occlusion, with good lighting" (zinc.cse.buffalo.edu, detectors endpoint, read 10 November 2026). The paper reports average running times of about 30 seconds for image and audio modules and about 90 seconds for video (arXiv 2404.13146, Section III.B, read 10 November 2026).
Nothing is sold, and allowances are set by tier. The model groups endpoint defines a regular_user with "upload_limit_per_day" 30 and "record_keeping_days" 15, an advanced_user with 100 uploads a day and 30 days, and a premium_user with unlimited uploads and unlimited records who also sees the new and retired model layers (zinc.cse.buffalo.edu, model groups endpoint, read 10 November 2026). The paper described the same ladder with different figures: "Advanced users can submit up to 300 daily tasks, while regular users are limited to 30" (arXiv 2404.13146, Section II.A, read 10 November 2026). The card records the live endpoint's 100 and quotes the paper's 300. The front end calls a backend API under the lab's path, and no public API documentation is published, so the API field is marked unavailable.
Retention is stated per tier and nowhere else. The platform publishes no privacy policy or terms page; the submission history page reads "Submissions are kept for" the account's retention policy, filled from the same tier table, and the account page offers to "Permanently delete your account and all associated data. This action cannot be undone" (zinc.cse.buffalo.edu, platform, read 10 November 2026). The retention rank field stays empty: the 15 days is an API tier value and the platform publishes no privacy policy or terms page, so no policy period exists to rank; the three tiers are quoted here and in the second table of the data retention ranking (ruling of 10 November 2026). The upload form asks "Do you want to share data with us?", and the paper says users "can choose whether they want to share their data for academic research purposes only" (arXiv 2404.13146, Section II.A, read 10 November 2026).
What the vendor claims
No accuracy figure is published for the platform. About text: DeepFake-o-Meter aggregates results from state-of-the art research independent AI media detection models across image, video, and audio modalities. The system provides structured analytical outputs to support media authenticity assessment. Paper abstract (arXiv 2404.13146): an open-source and user-friendly online platform [...] that integrates state-of-the-art methods for detecting Deepfake images, videos, and audio; it serves as an evaluation and benchmarking platform for researchers in digital media forensics to compare the performance of multiple algorithms on the same input. The platform notice reads DeepFake-O-Meter reaches 10,000 active monthly users; the paper reports 632 registered users and 4,091 tasks between February 8, 2024 and June 1, 2024 (Section III.A). Each detector returns its own likelihood score and the results page shows an Overall System Assessment.
No accuracy figure is claimed for the platform. The paper's abstract describes "an open-source and user-friendly online platform, DeepFake-O-Meter v2.0, that integrates state-of-the-art methods for detecting Deepfake images, videos, and audio" and says it "serves as an evaluation and benchmarking platform for researchers in digital media forensics to compare the performance of multiple algorithms on the same input" (arXiv 2404.13146, read 10 November 2026). The results page shows each detector's likelihood score, labels results by risk level and an "Overall System Assessment", and warns that the assessment waits until "all results are ready" (zinc.cse.buffalo.edu, platform, read 10 November 2026). The claim gap field on this card is empty because there is no percentage to gap.
The usage figures are the vendor's and they differ by date. The paper reports "632 registered users, among which 445 users submitted 4,091 tasks" between February 8, 2024 and June 1, 2024, with "on average, 34 new submissions each day" and 41% of access from the United States (arXiv 2404.13146, Section III.A, read 10 November 2026). The platform's notice board states "DeepFake-O-Meter reaches 10,000 active monthly users", undated, and its news feed dated 01/12/2026 and 12/09/2025 records the platform's use by AFP and Vishvas News fact checkers (zinc.cse.buffalo.edu, news endpoint, read 10 November 2026). The card keeps the paper's counted figures and the notice's round one apart.
The caveats are built into the scope lines. The platform states that "The rights of individual detection methods belong to their original authors" and that "The views, opinions, and findings expressed are those of the authors and do not necessarily reflect the official policies or views of the University at Buffalo, the National Science Foundation, or the U.S. Government" (zinc.cse.buffalo.edu, platform, read 10 November 2026). Each detector carries its own limit, such as LFCC-LCNN's "Input requires a single speaker and a clear voice background" (zinc.cse.buffalo.edu, detectors endpoint, read 10 November 2026). A score from a model trained on ProGAN images in 2023 says what that model says; the platform does not average them into a verdict it vouches for.
What independent studies found
No record in the index tested DeepFake-O-Meter as a platform, read 10 November 2026 in the data file, and the study count on this card is zero. The speech records ASVspoof 5 and SAFE 2025 and the video record DFDC 2020 score anonymised entrants and Kaggle teams, not this platform. Several hosted methods descend from that literature, AASIST being "trained on ASV spoof dataset" by its own scope line, but a detector's training set is not a test of the platform, and the index counts none.
The platform is itself a comparison tool. Its paper says it lets researchers "compare the performance of multiple algorithms on the same input", and its own usage analysis found that "the use of image and video detectors is significantly more widespread than that of audio detectors" (arXiv 2404.13146, Section III.B, read 10 November 2026). A reader who needs an outside test of a hosted method should look for that method's own paper, which the endpoint links beside each name.
Where DeepFake-O-Meter appears in the index
Rankings
Fits these use cases, questions and comparisons
The card appears on AI detectors by content type with image, voice and video as its columns and text blank, and in studies by modality with zero records against three non-text types, read 10 November 2026 in the data file. It is absent from the study-based rankings, which need an indexed record, and from the claim gap ranking, because no percentage is claimed; it carries no retention period (the tier values are quoted, not ranked) and its 30 uploads a day is a media unit the free checks per day ranking tables apart. The voice and video question lists the cards that claim both media, and this platform claims both. Every detector card on the register shows the same fourteen fields, so a grant-funded research platform sits beside the commercial detectors on the same terms.
Questions about DeepFake-O-Meter open detection platform
Is DeepFake-O-Meter free?
Yes, with an account: the regular tier allows 30 uploads a day with records kept 15 days, advanced 100 a day for 30 days, and premium unlimited, with no price on any tier, per the platform's model groups endpoint read 10 November 2026.
Which detectors does DeepFake-O-Meter run?
41 methods on 10 November 2026, 19 image, 10 video, 11 audio and 1 for both image and video, among them CLIP-ViT, AIDE, NPR, AltFreezing, DSP-FWA, LIPINC, AASIST, RawNet2 and Whisper, each with its paper and code linked.
How accurate is DeepFake-O-Meter?
The platform claims no accuracy figure: it returns each hosted detector's own score and an overall assessment, and no indexed study tested it, read 10 November 2026. Its paper describes it as a benchmarking platform for comparing algorithms.
What files can you upload to DeepFake-O-Meter?
JPG, PNG, BMP and TIFF images, MP4, AVI and MOV video, WAV and MP3 audio, or a YouTube URL, with per-detector limits such as 50 MB for video and 10 MB for images on CFM, read 10 November 2026.
Who runs DeepFake-O-Meter?
The University at Buffalo Media Forensics Lab, supported by NSF grant SaTC-2153112; version 1.0 dates from 2021 and version 2.0 is described in the lab's own paper (arXiv 2404.13146), read 10 November 2026.
Sources and revisions
Sources
- zinc.cse.buffalo.edu/ubmdfl/deep-o-meterRead 10 November 2026
- zinc.cse.buffalo.edu/ubmdfl/api/deep-o-me...Read 10 November 2026
- zinc.cse.buffalo.edu/ubmdfl/api/deep-o-me...Read 10 November 2026
- zinc.cse.buffalo.edu/ubmdfl/api/deep-o-me...Read 10 November 2026
- ubmdfl.cse.buffalo.eduRead 10 November 2026
- arxiv.org/abs/2404.13146Read 10 November 2026
- arxiv.org/pdf/2404.13146Read 10 November 2026
Revision log
- Card created2026-11-10: retention period set to not ranked at the merge (an api tier value with no policy page)
