A Coprime Blur Scheme for Data Security in Video Surveillance

Christopher Thorpe, Feng Li, Zijia Li, Zhan Yu, David Saunders, Jingyi YuIEEE TPAMI 2013 · OpenReview

TL;DR

This paper introduces a novel "coprime blur" technique for video surveillance that partially encrypts video streams by blurring them with two distinct, mathematically coprime kernels. This allows for public viewing of general activity while restricting access to sensitive details, which can be fully recovered only when both blurred streams are combined, offering a practical multi-level security solution.

What Problem Does This Paper Address?

Modern video surveillance systems are ubiquitous, offering enhanced security but also raising significant privacy concerns. Existing methods for protecting sensitive visual data often involve complete encryption, rendering the video useless for low-clearance users who might only need to observe general behavior. This paper addresses the challenge of providing a nuanced security approach: enabling the observation of activity in public surveillance feeds without revealing personally identifiable information (like faces), while still allowing authorized personnel with higher clearance to fully reconstruct the original, unblurred video if necessary. This is akin to a "blind vision" approach for surveillance.

Methods

The core of the proposed solution is a "Coprime Blurred Pair" (CBP) model. Instead of full encryption, the original video stream is blurred using two different, mathematically "coprime" kernels. These kernels are designed such that their corresponding bivariate polynomials in the $z$-domain share no common factors (their greatest common divisor, GCD, is 1).

The process involves:

  1. Blur Kernel Selection: Two large, randomly generated blur kernels are chosen. The size of these kernels is crucial for security, making brute-force deblurring difficult.
  2. Spatial Encryption (Blurring): The original video frames are blurred using these two coprime kernels, creating two distinct blurred video streams: a "public stream" and a "private stream."
  3. Deblurring (Coprime Deblurring): To recover the original video, the system leverages the coprime nature of the kernels. The mathematical relationship between the two blurred images ($B_1, B_2$) and the original latent image ($L$) and their respective blur kernels ($K_1, K_2$) can be expressed in the $z$-domain as $b(z_1, z_2) = l(z_1, z_2) \cdot k(z_1, z_2)$. The key insight is that the latent image $L$ can be recovered as the GCD of the $z$-transforms of the two blurred images: $l(z_1, z_2) = \text{GCD}\{b_1(z_1, z_2), b_2(z_1, z_2)\}$.
  4. Algorithm Implementation: The paper details an efficient algorithm to compute this GCD. It involves estimating the degree of the blur kernels, then decomposing the 2D GCD problem into several 1D GCD sub-problems. This is achieved by evaluating the polynomials on the unit circle and using techniques like the Bézout matrix and Fast Fourier Transform (FFT) division.
  5. GPU Acceleration: To make the deblurring process practical for real-time surveillance, the entire pipeline is implemented on a Graphics Processing Unit (GPU) using CUDA.

Datasets

Not discussed in the paper. The paper uses typical surveillance scenarios and video sequences captured by a Point Grey Flea2 firewire camera for evaluation.

Key Insights and Ideas

  • Partial Encryption for Privacy: The most significant insight is the shift from complete encryption to partial encryption (blurring) for privacy-preserving surveillance. This allows for a tiered access system where general activity is visible to all, but sensitive details are hidden.
  • Mathematical Coprimality as a Security Key: The coprime nature of the blur kernels is not just a mathematical curiosity but forms the basis of the decryption key. Access to both blurred streams is essential, and the mathematical relationship between the kernels enables reconstruction without needing to explicitly store or transmit the kernels themselves.
  • GCD as the Deblurring Mechanism: The elegant application of the Greatest Common Divisor (GCD) in the $z$-domain to recover the latent image from two blurred versions is a core theoretical contribution.
  • Efficiency through 1D Decomposition: Decomposing the 2D GCD problem into multiple 1D GCD sub-problems significantly improves computational efficiency and reduces memory requirements.
  • GPU for Real-Time Performance: The successful implementation on a GPU demonstrates that this sophisticated deblurring technique can achieve real-time frame rates, making it viable for practical surveillance applications.

Key Results

  • Effective Privacy Protection: The proposed scheme effectively obscures sensitive identity information (like facial features) in public streams, while still allowing viewers to discern general behavior and activity.
  • Accurate Reconstruction: When both blurred streams are available, the algorithm can faithfully reconstruct the original, unblurred video stream.
  • Real-Time Performance: The GPU implementation achieves frame rates exceeding 20fps for kernel sizes up to 25x25, approaching or exceeding standard video frame rates.
  • Speedups: The GPU implementation provides significant speedups (21x to 77x) compared to CPU-based execution.
  • Scalability: The performance is evaluated across different kernel sizes (9x9 to 25x25), showing how larger kernels (offering more privacy) impact running time.

Limitations and Future Work

  • Bandwidth Requirements: The approach requires double the bandwidth for transmitting two video streams.
  • Computational Cost: While GPU acceleration is effective, the running time is still proportional to the blur kernel size, which can limit the level of privacy achievable for a desired frame rate. This might also restrict some post-processing steps like facial recognition.
  • Future Work:
    • Reducing Running Time: Optimizing expensive operations (like SVD) or exploring hybrid pipelines with hardware accelerators (e.g., FPGAs) to further decrease running time.
  • Integration with Depth Information: Utilizing $z$-cameras to obtain depth information, which could help in computing optimal blur kernel sizes for hiding information more effectively.
  • Multi-level Security Extension: The paper suggests extending the scheme to more than two levels of security by restricting access to different bit precisions of the recovered image.

Summary updated

Abstract

This paper presents a novel coprime blurred pair (CBP) model to improve data security in camera surveillance. While most previous approaches have focused on completely encrypting the video stream, we introduce a spatial encryption scheme by strategically blurring the image/video contents. Specifically, we form a public stream and a private stream by blurring the original video data using two different kernels. Each blurred stream will provide the user who has lower clearance less access to personally identifiable details while still allowing behavior to be monitored. If the behavior is recognized as suspicious, a supervisor can use both streams to deblur the contents. Our approach is based on a new CBP theory where the two kernels are coprime when mapped to bivariate polynomials in the $(z)$ domain. We show that coprimality can be derived in terms of the rank of Bézout matrix formed by sampled polynomials, and we present an efficient algorithm to factor the Bézout matrix for recovering the latent image. To make our solution practical, we implement our decryption scheme on a graphics processing unit (GPU) to achieve real-time performance. Extensive experiments demonstrate that our new scheme can effectively protect sensitive identity information in surveillance videos and faithfully reconstruct the unblurred video stream when both CBP sequences are available.
BibTeX
@article{thorpe2013coprime,
  author = {Christopher Thorpe and Feng Li and Zijia Li and Zhan Yu and David Saunders and Jingyi Yu},
  title = {{A Coprime Blur Scheme for Data Security in Video Surveillance}},
  journal = {IEEE Trans. Pattern Anal. Mach. Intell.},
  year = {2013},
  url = {https://openreview.net/forum?id=sqiApEdPme},
}
Figure