VisionLabs DeepFake Detector. VisionLabs DeepFake Detector (VL DFD) can detect visual deepfakes on a single image or video. The AI solution detects the probability, that the person in the image or video is a deepfake on a frame, n-frame and video.
The problem is the fraud with personal facial biometrics or the spread of fake content, which is caused by increased access to high-quality deepfake generation algorithms, that are much easier to implement than 2 years ago. Fraud: financial institutions / banks or any organization, that uses authentication with facial biometrics can be attacked by intruders, who wish to gain access to credits or other services remotely. For this purpose, they obtain a photo of a person, whith a real client, generate a deepfake and during the videocall with the bank, substitute initial video stream with generated one. The operator of the call on behalf of the organization may not realize that she is talking to a fake as the video stream may be of low quality due to obvious reasons: poor internet connection, low-resolution web camera, etc. Thus, the operator may be mistaken and approve transaction for the fake person. Another problem is the distribution of fake content generated with deepfake technology. To solve these two problems, VisionLabs provides VL DFD – an AI neural networks solution, that is capable of detecting all the most popular visual deepfakes: • Face swap is a method, used to transfer face from a source image (bona fide) to the target video (attacker) • Face reenactment is a method, used to implement mimics of an attacker onto the image of a bona fide person. This type includes lip-sync, the type of deepfake algorithm used to synchronize video and audio. • Face synthesis is a method used to generate synthetic faces of non-existent persons. VL DFD can operate in real-time during a videocall, using 1 frame or n-frames, which is suitable for the financial sector. Additionally, VL DFD can analyze content post-factum to prevent the spread of fake media content VL DFD can operate both on CPU and GPU in the backend. VL DFD receives inputs either from front-end (frames that can be selected using VL frontend solution or using client’s software) or from backend.
Artificial Intelligence And Digital Services
Artificial intelligence
From 1 hour to 6 hours, depending on which product is selected for installation (frontend or high-load backend system)
The CIS, Russia, The UAE
Competitive advantages are: - Ability to operate both in real-time and post-factum - Support for both CPU and GPU (most NN operate on GPU-only) - Ability to detect deepfakes on a single frame (most competitors work with video or several frames only) - Ability to work with video (full file) in the backend - Ability to detect several faces in the image/video and provide output for each face - Ability to filter files (e.g., if the face is missing, if the resolution is too low, if the quality is low, etc), the filters can be customized - Low competition, as most vendors only provide liveness (presentation attack detection), but not deepfake detection (which attacks the channel) - VL DFD is used together with virtual camera detection (Solution provided by VisionLabs for web, iOS, Android). Deepfake detection combined with virtual camera detection helps to provide a complex security solution - Ability to detect all popular and major deepfake generation algorithms - Continuous monitoring and training of VL DFD against new types of fakes - Strong expertise in anti-fraud solutions (Liveness VisionLabs was successfully tested by iBeta (0 successful fraud attempts at levels 1 and 2) - Strong understanding of the target audience, as VisionLabs is a leader in facial biometrics and has completed numerous projects for financial institutions
The winner of the FINNEXT award for the transformation of Core Banking https://bosfera.ru/bo/obyavleny-imena-pobediteley-premii-finnext First place in the "The Face Deepfake Detection Challenge" https://www.researchgate.net/publication/363928464_The_Face_Deepfake_Detection_Challenge Webinar dedicated to the VisionLabs deepfake detector https://mts.ai/ru/sobytiya/web-deepfake-detection/ Media articles: https://www.vedomosti.ru/technology/articles/2023/06/01/978035-biometricheskie-proverki-nachnut-zaschischat-ot-dipfeikov https://www.biometricupdate.com/202307/visionlabs-introduces-anti-deepfake-tech-clinches-deal-for-govt-biometric-system https://rb.ru/news/visionlabs-deepfake/
Intangible asset VisionLabs LP5 DeepFake detection, which is included in the Register of domestic software (entered to the record LUNA PLATFORM 5, registration record No. 12557 dated 01/14/2022)
