Introduction
As surveillance drones become increasingly prevalent in security and patrol operations, the risk of privacy infringement through "indiscriminate shooting" grows. While drones are excellent tools for accessing hazardous or hard-to-reach areas, their high-resolution cameras often capture sensitive personal information without consent.
Traditional privacy-preserving methods, such as blurring or pixelation, present significant technical challenges for autonomous systems. Specifically, these methods can "break" a drone's computer vision systems. If a drone utilizes Simultaneous Localization and Mapping (SLAM) to navigate, it relies on extracting feature points (edges, corners, textures) from its environment. Blurring a large portion of the frame—such as a person's face—removes these critical points, potentially causing the drone to lose orientation or fail to detect obstacles. This paper introduces a real-time drone patrol system that preserves privacy without sacrificing robot perception performance by replacing detected faces with photorealistic, synthetic ones.
System Architecture
The proposed system is built upon the Robot Operating System (ROS) and splits functionality between a ground station and the patrol drone itself.
The hardware configuration includes a customized DJI F550 drone equipped with a ZED stereo camera and an Nvidia Jetson Xavier "companion computer." By performing anonymization on the edge (the drone itself), the system ensures that raw, un-anonymized video is never transmitted over the air, closing a major security loophole.
Proposed Methodology
The core of the system is a two-stage deep learning framework designed for real-time execution on mobile hardware.
1. Face Detection and Segmentation
The system first detects faces using FaceBoxes, a lightweight detector. To handle varying resolutions, the system employs repetitive detection to ensure faces are tightly cropped. A modified CycleGAN then translates the photorealistic face into a semantic segmentation mask, identifying key components like eyes, nose, and mouth.
2. Face Synthesis and Replacement
A second network, based on GauGAN (SPADE), takes the segmentation mask and generates a completely new, photorealistic face. This synthetic face maintains the original pose and expression but belongs to a non-existent person.
Technical Innovations
To ensure high-quality anonymization, the authors introduced several novel modifications:
- VGG Perceptual Loss with Margins: To prevent the network from accidentally regenerating the original identity, a margin ($\epsilon$) was added to the loss function. This forces the generated face to be a specific "distance" away from the original in feature space.
- Component-Level Discrimination: The discriminator investigates individual facial components rather than the whole face. This allows the generator more freedom to create a diverse "new" face while strictly maintaining the structural integrity of the original expression.
Experimental Results
The researchers utilized benchmarks including CelebA-HQ, Helen, FaceScrub, and FaceForensic, alongside custom drone footage, to validate the system.
Privacy and Utility
Quantitative privacy was measured using a Siamese network to calculate the Euclidean distance between original and anonymized faces. In all tests, the distance significantly exceeded the threshold for a "different person," confirming effective anonymization.
Crucially, the system maintained navigation stability. When comparing SLAM performance on original versus anonymized footage, the feature points extracted by ORB-SLAM2 were nearly identical. This proves the drone can calculate distances and map its environment perfectly, even with synthetic faces in the frame.
Conclusion and Future Work
This framework successfully decouples identity from structure, allowing for privacy-compliant drone patrols. While the system achieves high visual quality and maintains SLAM performance, future work is needed to address computational overhead for high-speed flight and to expand anonymization to other identifiers like clothing, tattoos, or gait. Additionally, the robustness of the GAN against adversarial "reversal" attacks remains an open area for research.
Summary updated