A Multi-Layered Privacy-Preserving Approach to Obfuscated Human Action
Recognition for Robust Anomaly Detection
Abstract
Privacy-preserving human activity anomaly detection has become
critical in privacy-sensitive fields like video surveillance,
healthcare, and assisted living. While human action recognition offers
significant advantages in automated analysis, it raises confidentiality
concerns. This paper introduces Obfuscated Action Detection, a novel
framework that uses Generative Adversarial Networks (GANs) for temporal
obfuscation, ensuring privacy while maintaining high accuracy. By
integrating Deep Neural Networks, the framework delivers robust anomaly
detection with real-time feasibility. Tested on the UCF101 dataset, the
model achieves high accuracy (98.59% to 100%) and strong
generalization to unseen data (88.44% test accuracy). With impressive
precision (98.14%), recall (99.56%), and F1 score (98.84%),
Obfuscated Action Detection effectively balances privacy with
performance. The framework shows promise for real-world applications in
privacy-critical domains, offering robust privacy protection without
compromising detection accuracy. Extensive experiments demonstrate the
capability of Obfuscated Action Detection to achieve robust privacy
protection without compromising detection precision, making it a viable
solution for applications that prioritize both privacy and reliability.
Additionally, this paper presents an overview of related works,
summarizing recent advancements and methodologies in privacy-preserving
anomaly detection.