PrivHAR-Bench: A Graduated Privacy Benchmark Dataset for Video-Based Action Recognition (v1.0.0)

PrivHAR-Bench is a multi-tier benchmark dataset for evaluating the privacy-utility trade-off in video-based human activity recognition (HAR). The dataset applies a graduated spectrum of visual privacy transformations to 1,932 source videos drawn from 15 classes of UCF101 (Soomro et al., 2012), producing 9 parallel tiers of increasing privacy strength: - Original (clear video)- Tier 1: Gaussian Blur (sigma=15)- Tier 2: Canny Edge Detection- Tier 3: AES Block Scrambling at B=4, B=8, B=16- Background-removed (NoBG) variants for all three Tier 3 block sizes All frames are stored as lossless PNG at 224x224 resolution. The dataset includes per-frame bounding boxes, estimated pose keypoints (17 COCO-format joints via YOLOv8n-Pose), and group-based train/test splits. An evaluation toolkit (eval.py) and the complete generation pipeline are available at: https://github.com/SamarAnsariUK/PrivHAR-Bench. ------------- Associated paper: [arXiv URL] ------------- Reference: Soomro, K., Zamir, A. R., & Shah, M. (2012). UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402.