Introduction
In the rapidly evolving landscape of edge-cloud video analytics—encompassing smart city surveillance and autonomous driving—sensitive data is frequently offloaded to cloud servers for processing. This transmission exposes data to "inference attacks," where malicious actors attempt to extract private attributes (e.g., identity, gender) from the video stream. While various protection schemes, such as noise perturbation or object removal, have been proposed, practitioners currently face a "blind choice" problem. There is no established analytical method to determine if a specific video stream can be effectively protected for a given task without exhaustive, resource-intensive empirical testing.
This paper introduces Privacy Protectability ($\\mathcal{P}$-score), a novel information-theoretic metric designed to quantify the intrinsic protectability of a video stream before any protection scheme is applied. By evaluating the overlap between features required for utility and those that leak private information, the $\\mathcal{P}$-score offers a predictive framework that significantly enhances evaluation efficiency, offering over a 1000x speedup in assessment.
The $\\mathcal{P}$-score Framework
The core of the proposed approach is to treat video features as participants in a cooperative game, utilizing Shapley Values to quantify their contributions.
- Feature Contribution Analysis: The framework measures how much each individual feature $Z$ contributes to the target analytics task ($Y_a$) versus the private attribute ($Y_{pri}$).
- Protectable Features ($Z_P$): A feature is classified as "protectable" if its contribution to the private attribute remains below a user-defined threshold ($\\epsilon$).
- $\\mathcal{P}$-score Calculation: The $\\mathcal{P}$-score is defined as the ratio of the contribution of protectable features to the total contribution of all features for the analytics task. A high score indicates that the task primarily relies on features that do not compromise privacy.
- Level-of-Protection ($\\mathcal{LP}$-score): Complementing the $\\mathcal{P}$-score, the $\\mathcal{LP}$-score evaluates the actual performance of an already-implemented protection scheme (e.g., Differential Privacy or DISCO).
To ensure computational feasibility, the authors developed the Privacy Protectability Estimation (PPE) algorithm. PPE utilizes Mutual Information as a proxy for predictive power and employs Monte Carlo sampling to approximate Shapley values, bypassing the exponential complexity typically associated with such calculations.
Experimental Evaluation
The researchers validated the $\\mathcal{P}$-score across several diverse datasets, including CelebA (200,000 face images), LISA (traffic light detection), and real-world traffic streams from NYC. The evaluation demonstrated a remarkable 0.99 Pearson correlation between the predicted $\\mathcal{P}$-score and empirical measurements of privacy protection performance, confirming the metric's reliability as a predictive tool.
Results and Discussion
The application of the $\\mathcal{P}$-score reveals that privacy is an intrinsic property of the relationship between data, the specific task, and the sensitive attribute. If the features required for a task overlap perfectly with those identifying a private attribute, perturbation-based schemes are fundamentally limited.
Key findings include:
- Task Sensitivity: Tasks like "Mouth Open" classification exhibit a high $\\mathcal{P}$-score (0.90), indicating they are easy to protect. In contrast, "Eyeglasses" classification has a low $\\mathcal{P}$-score (0.23), as the features identifying eyewear are inextricably linked to gender identification.
- Computational Efficiency: The PPE algorithm achieved a 1143x reduction in computational cost compared to traditional model-based evaluations (35 TFLOPs vs. 3859 TFLOPs).
- Decision Support: The $\\mathcal{P}$-score provides a theoretical upper bound for perturbation schemes. If the score is low, engineers can immediately pivot to high-overhead solutions like encryption or Trusted Execution Environments (TEEs) rather than wasting resources on ineffective noise-based methods.
Conclusion
The $\\mathcal{P}$-score framework transforms privacy evaluation from an empirical "trial-and-error" process into a formal, predictive science. By identifying the intrinsic protectability of video tasks, it enables massive speedups in system design and provides a rigorous basis for selecting privacy-preserving technologies. Future work will focus on refining threshold selection and exploring alternative feature contribution measures to further automate the protection pipeline.
Summary updated