Differential-Privacy-Based Collaborative Protection for Visual and Location Data in UAV Semantic Communications

Area of Science:

• Computer Science
• Electrical Engineering
• Information Security

Background:

• Unmanned aerial vehicle (UAV) semantic communications face challenges with limited resources and vulnerable air-to-ground links.
• Transmitting visual semantic features and location metadata creates a dual-privacy risk for sensitive content reconstruction by eavesdroppers.
• Existing methods struggle to balance data utility with robust privacy protection in dynamic UAV environments.

Purpose of the Study:

• To propose a differential privacy-based collaborative protection framework for UAV semantic communications.
• To address the dual-privacy vulnerability arising from simultaneous transmission of visual and location data.
• To enhance the privacy-utility trade-off for UAV sensing, inspection, and emergency response applications.

Main Methods:

• Developed a region-aware differential privacy mechanism for visual data, applying stronger noise to sensitive semantic regions.
• Implemented a scenario-adaptive strategy for location data, using randomized and Laplace-based differential privacy for different coordinate types.
• Formulated a joint optimization problem to balance privacy budgets and transmit power for maximizing semantic task performance.
• Utilized a Block Coordinate Descent (BCD)-based algorithm to solve the non-convex optimization problem.

Main Results:

• Empirically verified reduced attacker-side recoverability at the optimized operating point.
• Demonstrated stable convergence of the BCD-based algorithm within a few iterations.
• Achieved a superior task-level privacy-utility trade-off compared to uniform differential privacy.
• Showcased selective sensitive-region protection for visual data with comparable whole-image attack suppression.

Conclusions:

• The proposed differential privacy framework effectively protects sensitive visual and location data in UAV semantic communications.
• The region-aware and scenario-adaptive mechanisms offer a superior privacy-utility balance, crucial for resource-constrained UAVs.
• This approach enhances security for critical applications like intelligent inspection and emergency response without significant data utility loss.