Optimization and Probabilistic Modelling for Privacy-Preserving Scalable Video Analytics

Abstract

Privacy-preserving video analytics (PPVA) aims to extract useful information from video streams while reducing the exposure of sensitive visual content. However, privacy protection, analytics performance, and system scalability are often addressed as separate objectives, making it difficult to provide reliable privacy guarantees under changing scene conditions and limited edge resources. This paper proposes a scene conditioned, chance-constrained probabilistic framework for scalable privacy-preserving video analytics. The proposed approach models privacy leakage as a stochastic risk influenced by the latent sensitivity of the observed scene. A unified optimization formulation jointly determines the privacy transformation strength, video resolution, bandwidth allocation, computational resources, and edge-cloud offloading decisions. The formulation incorporates analytics accuracy, latency, energy consumption, and privacy risk while explicitly constraining the probability of privacy violation. A risk-sensitive objective with Conditional Value-at-Risk (CVaR) is further incorporated to account for high-risk privacy events. The framework is evaluated through privacy-accuracy-resource trade-offs, chance-constraint validation, multi-stream scalability analysis, and optimization convergence analysis. The results demonstrate the expected trade-off between stronger privacy protection and analytics utility while maintaining the specified probabilistic privacy constraints and bounded resource growth as the number of concurrent video streams increases. The proposed formulation provides a unified mathematical basis for adaptive privacy protection and scalable resource management in large-scale video analytics environments.