APATE Deepfake Frauds

This dataset was funded by the ANR project APATE: ANR-22-CE39-0016 This dataset of deepfakes was created for the ANR APATE project for evaluation purposes. A data-acquisition campaign was conducted at IDEMIA in which 11 people provided a total of 43 videos portraying various simulated frauds. IDEMIA's videos are used as driving videos only. The 11 identities have been modified visually to ressemble subjects from the Microsoft VCD dataset, and the audio tracks have been modified to sound like subjects from the Mozille Common Voices dataset. Driving video statistics: 43 videos 10 men / 1 woman 12 English (2 IDs) / 31 French (10 IDs) 21 with glasses / 22 without 4 outdoors (2 IDs) Role Num. videos Banker 12 Romantic partner 12 President 9 Influencer 7 Son 2 Colleague 1 An Excel file containing detailed meta data for the driving videos can be found with the dataset (driving_video_meta_data.xlsx). Deepfake methods Face-swap methods The FaceFusion package was used to run the following four face-swap methods: INSwapper HiFiFace SimSwap Uniface These methods were deemed to be amongst the better-performing available models. Each of the videos was then further processed using the GFPGAN (v1.4) face-image-restoration model, again using the FaceFusion package. Avatar methods Method URL LivePortrait https://github.com/KlingTeam/LivePortrait X-NeMo https://github.com/bytedance/x-nemo-inference Voice conversion Audio deepfakes were generated by running the 24KHz FreeVC voice conversion model via the Coqui TTS package. LICENCE Although faces and voices in driving videos have been replaced, there is a risk that the deepfake videos and audio may still contain personal information of employees of IDEMIA Public Security. The dataset is therefore released under a custom licence to APATE partners only, as per the licence agreements signed by IPS employees during collection of the driving videos. Please see LICENCE.txt for details.