DeepVerify

"The DeepVerify dataset is a benchmark multimodal video dataset designed for evaluating integrity verification, continuous candidate authentication, and automated proctoring in remote technical interviews and online assessments. Collected across indoor environments by student research teams at Amrita Vishwa Vidyapeetham with informed participant consent for public academic release, the dataset comprises synchronized facial video recordings from 59 participants (303 video clips) captured under diverse natural and artificial lighting conditions at 30 FPS (640x480 and 1280x720 resolutions).The dataset is partitioned into two core subsets:1. rPPG Physiological Subset: Contains facial video recordings captured under static and varying head orientations, controlled illumination transitions, and non-live display spoofing attacks (screen replay) to benchmark contactless pulse extraction, physiological signal consistency, and pulse-based liveness detection.2. Behavioral Anomaly Subset: Features fine-grained scripted simulations replicating real-world remote assessment misconduct. Scenarios systematically cover baseline normal gaze, off-screen gaze diversions (left, right, downward), secondary mobile device usage, reading from hidden physical notes, simulated tab/window switching, and camera occlusion.Each video clip is accompanied by an individual JSON metadata record capturing camera parameters, timestamps, scenario tags, and ambient lighting conditions. In addition, an aggregated master index (metadata.csv) is provided to enable direct ingestion into PyTorch and TensorFlow deep learning pipelines. Annotations are provided at the scenario level. This dataset provides the experimental benchmarking data for the study entitled \"DeepVerify A Zero Trust Multi Modal Framework for Continuous Deepfake and Behavioral Integrity Verification in Remote Technical Interviews\" under review at Discover Computing (Springer Nature). "

Verbatim fromDataCite· Licence: https://creativecommons.org/licenses/by/4.0/legalcode