CANOPIC: Pre-Digital Privacy-Enhancing Encodings for Computer Vision

Tan, Jasper; Khan, Salman Siddique; Boominathan, Vivek; Byrne, Jeffrey; Baraniuk, Richard G.; Mitra, Kaushik; Veeraraghavan, Ashok

doi:10.1109/icme46284.2020.9102956

Cited by 6 publications

(9 citation statements)

References 15 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In recent years, there is a growing interest in developing privacy-preserving vision systems for various computer vision tasks such as action recognition [37,30], face detection [33], pose estimation [14,32], fall detection [3], and posture classification [12]. Here, we provide a brief overview of privacy-preserving frameworks for various computer vision tasks, especially human action recognition.…”

Section: Related Workmentioning

confidence: 99%

“…Such frameworks learn to preserve privacy via adversarial training [28] using Deep Neural Networks (DNNs) that train the parameters of an encoder to actively inhibit the sensitive attributes in an visual data against an adversarial DNN whose task is to learn privacy attributes while allowing attributes that are essential for the computer vision task [5,24,33,37,16,21]. Note that the encoder can be a hardware module [14,33] or a software module [37,28]. Besides these works, there are other imaging systems that use optical operations to hide sensitive attributes.…”

Section: Related Workmentioning

confidence: 99%

“…Unfortunately, such a formulation is not differentiable and therefore not suitable for back-propagation. Thus, following [38,33], we approximate the above quantization function with a differentiable version by replacing the Heaviside function with the sigmoid σ() function, which results in the following formulation:…”

Section: Bdq: Blur Difference Quantizationmentioning

confidence: 99%

“…any sensitive information, but allow it to capture useful information to successfully perform the CV task [37,30,33,14,26]. For a satisfactory user experience and strong privacy protection, a CV system must satisfy the following properties:…”

Section: Introductionmentioning

confidence: 99%

“…This is necessary for maintaining a good user experience. For example, a privacy-preserving face detection model must detect faces with high precision without revealing facial identity [33], a privacypreserving pose estimation model must detect body key-points without revealing the person identity [14], and an action recognition model must recognize human actions without revealing their identity information [37,30]. -Strong privacy protection.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Privacy-Preserving Action Recognition via Motion Difference Quantization

Kumawat¹,

Nagahara²

2022

Preprint

View full text Add to dashboard Cite

The widespread use of smart computer vision systems in our personal spaces has led to an increased consciousness about the privacy and security risks that these systems pose. On the one hand, we want these systems to assist in our daily lives by understanding our surroundings, but on the other hand, we want them to do so without capturing any sensitive information. Towards this direction, this paper proposes a simple, yet robust privacy-preserving encoder called BDQ for the task of privacy-preserving human action recognition that is composed of three modules: Blur, Difference, and Quantization. First, the input scene is passed to the Blur module to smoothen the edges. This is followed by the Difference module to apply a pixel-wise intensity subtraction between consecutive frames to highlight motion features and suppress high-level privacy attributes. Finally, the Quantization module is applied to the motion difference frames to remove the low-level privacy attributes. The BDQ parameters are optimized in an end-to-end fashion via adversarial training such that it learns to allow action recognition attributes while inhibiting privacy attributes. Our experiments on three benchmark datasets show that the proposed encoder design can achieve state-of-the-art trade-off when compared with previous works. Furthermore, we show that the trade-off achieved is at par with the DVS sensor-based event cameras. Code available at: https://github.com/suakaw/BDQ_PrivacyAR

show abstract