| Author: | Lu, Rui |
| Title: | Privacy-preserving video analytics systems for resource-constraint edge devices |
| Advisors: | Wang, Dan (COMP) |
| Degree: | Ph.D. |
| Year: | 2024 |
| Department: | Department of Computing |
| Pages: | xv, 142 pages : color illustrations |
| Language: | English |
| Abstract: | Currently, video analytics systems (VAS) are extensively used to support such applications as traffic monitoring, security surveillance, autonomous driving, and the emerging domain of the Metaverse. These systems typically equip cameras deployed on resource-constraint edge devices, like smart cameras, AI edges devices and AR/VR headsets, capturing video frames for analytics. Those video frames are inputted into well-trained neural networks to perform model inference and obtain analytics results. Meanwhile, privacy concerns arise with increasing video analytics application development. The complex and vulnerable cyber environment between edge and cloud makes the transmission of sensitive information, such as human faces and vehicle license plates, susceptible to interception. Privacy further leaks if sensitive attributes such as gender and race are extracted. Consequently, protecting the privacy of the video analytics system is necessary. This thesis focuses on protecting the privacy of typical video analytics systems while simultaneously maintaining high analytics accuracy and minimizing latency. It investigates three prevalent scenarios of VAS, categorized according to the location of inference workload offloading: exclusively on the edge (i.e., edge-side), concurrently on both the edge and the cloud (i.e., edge-cloud), or solely within the cloud (i.e., cloud-side). We contribute to the field through the following effective methods and novel system designs. First, we focus on the edge-side VAS, which offloads all inference workloads on the edge and keeps any captured videos local to protect privacy. Only the inference results can be uploaded to the cloud, making the analytics performance challenging. We show that existing VAS comprises two kinds of computing workload: instruction-intensive computing (IIC) handled by the CPU, e.g., video preprocessing, and data-intensive computing (DIC) handled by the GPU, e.g., model inference. The amount of workload is dynamic because of different video contents, resulting in significant fluctuations in accuracy in fields. To address existing edge devices with fixed CPU and GPU resources that are unable to adapt to the above dynamics effectively, we propose a new edge-side VAS, Gemini, enhanced by a dual-image FPGA that can be pre-configured with two FPGA images with negligible image switching time. We pre-configure it into one IIC image and one DIC image and elastically multiplex the dual IIC and DIC resources in the time dimension. We develop a new abstraction of hardware functions to achieve hardware-agnostic, and a bandit learning approach for resource management. Through the implementation and comprehensive analysis of Gemini, we demonstrate its analytics accuracy improvement and latency reduction, thereby validating its efficacy and capability for enhancing video analytics systems. Secondly, we focus on the edge-cloud VAS, where the edge and the cloud cooperatively conduct model inference of the video frames. Specifically, the edge conducts initial analytics on the video frames to a split layer of the NN model and then sends intermediate results to the cloud for follow-up analytics. We demonstrate that a reconstruction attacker attacks the intermediate results, and private information of video frames, e.g., facial features, can be leaked. Therefore, we present a new edge-cloud VAS, Preva, to conduct image transformation on the video frames as preprocessing prior to video analytics. Our goal is to ensure that the intermediate results do not leak private information under attack during inference. We design a policy-based video-frame transformation scheme to ensure high analytics accuracy and minimize privacy leakage in any split layer. We also present a formal privacy analysis to show that Preva can guarantee privacy protection capability under both outsider and insider reconstruction attacks. The evaluation results show that Preva outperforms existing systems in both analytics accuracy and privacy protection. Finally, we focus on the cloud-side VAS, where the edge devices are responsible for securely streaming the original video to the cloud for inference. We investigate the emerging 3D volumetric videos that are widely used in the areas of Metaverse, Esports, etc. Existing privacy-preserving approaches suffer massive computation costs and degrade the quality of the streaming. Therefore, we design Pagoda, a new privacy-preserving volumetric video streaming enhancement incorporating the MPEG V-PCC standard, which protects the privacy of volumetric videos, and maintains high throughput. The core idea is to content-aware transform the privacy attribute information to the geometry domain and content-agnostic protect the geometry information by adding Poisson noise perturbations. These perturbations can be denoised through a Poisson diffusion probabilistic model on the cloud. Edges only need to encrypt a small amount of high-sensitive information to achieve secure streaming. Our designs ensure that volumetric videos can be transmitted in high quality and that attackers are unable to recover them. We evaluate Pagoda and show that it outperforms existing privacy-preserving baselines in protection capability improvement, streaming quality, and latency reduction. In summary, we investigate three kinds of VAS and introduce corresponding privacy protection approaches to protect sensitive information while maintaining high analytics accuracy and low latency. These approaches are implemented and rigorously evaluated using diverse video analytics models and real-world datasets for distinct video analytics tasks, demonstrating their effectiveness in privacy-preserving. Overall, our work has significant implications for video analytics applications, particularly in scenarios where the protection of sensitive information is essential while maintaining high-quality analytics services. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/13516