"Deepfake Detection using Gait Analysis"

"This dataset supports the development and evaluation of deepfake detection systems that rely on skeletal gait analysis rather than facial artifact inspection. It comprises walking videos from 13 unique subjects recorded in controlled indoor conditions, with each subject performing multiple walking passes across two camera perspectives: frontal (F) and side-profile (S), yielding approximately 65 raw videos. Through a 16\u00d7 data augmentation pipeline \u2014 including temporal jitter, Gaussian noise injection on keypoints, speed perturbation, horizontal flipping, and occlusion simulation \u2014 the dataset expands to 1,056 augmented video samples.For each video, 78-dimensional skeletal gait feature vectors are extracted per frame using MediaPipe Pose, derived from 12 biomechanically relevant keypoints (shoulders, hips, knees, ankles, heels, and foot indices). The feature vector encodes hip-centered 3D coordinates (36 dimensions), six joint flexion angles at the knee, hip, and ankle (6 dimensions), and frame-to-frame velocity signals (36 dimensions). All sequences are normalized to exactly 60 frames (~2 complete gait cycles at 30 fps).The dataset is organized at the subject level to support Leave-One-Out Cross-Validation (LOOCV), ensuring strict subject-level separation to prevent data leakage during model evaluation. Subjects include male and female participants ranging in age from approximately 19\u201325 years. Videos are stored in MP4 format and follow the naming convention SubjectName_ViewN.mp4. Pre-extracted gait features are provided as a serialized .pkl file containing (sequence, label, subject_id) tuples, ready for direct model training without requiring MediaPipe installation.This dataset was used to train and evaluate a CNN+BiLSTM+Transformer hybrid deepfake detection model, achieving an AUC-ROC of 94.95% \u00b1 2.81% and an F1 score of 86.56% under LOOCV, filling a confirmed research gap where no prior published dataset exists specifically for gait-based deepfake verification."