AgentCam: Privacy-Aware and Security-Oriented Video Sensing for Heterogeneous Cloud-Edge-Terminal Surveillance Networks

Heterogeneous cloud-edge-terminal surveillance networks integrate resource-constrained cameras, edge devices, base stations, and cloud servers for security monitoring in smart campuses, transportation systems, and public spaces. Continuous full-color video uploading increases bandwidth pressure, enlarges the sensing pipeline’s attack surface, exposes privacy-sensitive background content, and may delay security analysis under congested edge links. To address these challenges, this paper presents AgentCam, a privacy-aware and security-oriented video sensing framework for heterogeneous surveillance networks. AgentCam combines a dual-mode grayscale-color camera with a cloud-hosted reinforcement learning (RL) agent. The sensing unit first transmits a low-cost grayscale observation; the cloud-side agent then selects security-relevant regions of interest (ROIs), and the terminal/edge unit applies the returned ROI action to transmit only selected color content. The cloud side reconstructs high-fidelity color video from grayscale frames and sparse color ROIs through restoration, temporal inference, and super-resolution. Experiments on surveillance-oriented video datasets show that AgentCam improves reconstruction quality over existing dual-mode reconstruction baselines while substantially reducing transmission overhead. These results indicate that selective ROI transmission can support scalable and privacy-aware security monitoring without continuously exposing full-color video streams. The framework is positioned as sensing-layer visual-data minimization rather than a complete cryptographic or adversarial-defense mechanism.