PrivHAR: Supplementary Material - Privacy-preserving Human Action Recognition With Light Propagation Modeling.pdf

PrivHAR - Privacy-Preserving Human Action Recognition from Camera Sensor Data

This document provides detailed supplementary information to support the findings presented in the main paper, "PrivHAR: Recognizing Human Actions from Privacy-preserving Lens."

1. Light Propagation and Optics Modeling:

The paper utilizes a differentiable Fourier optics model [7] to describe light transport within the camera system. This model is crucial for understanding how the proposed phase mask and lens configuration modify the incident light wavefront, ultimately forming the point-spread function (PSF).

2. Point-Spread Function (PSF) Frequency Analysis:

The Modulation Transfer Function (MTF) [1] is employed to validate the PSF. Figure 2 illustrates the MTF of various PSFs, highlighting the proposed PSF's low invertibility characteristics, particularly in the high-frequency range. This property contributes to PrivHAR's robustness against face detection attacks based on high-frequency textures (e.g., skin color).

3. Face Recognition Results:

The paper employs a Tensorflow implementation of ArcFace [3], a deep learning-based face recognition network, for evaluating the privacy preservation effectiveness. Experiments are conducted on three datasets: LFW [11], AgeDB-30 [14], and CFP-FP [17].

Pretrained Model: Utilizes a pretrained ArcFace model on "non-private" images to test performance on "private" (PrivHAR-optimized) images.
Trained Model: Trains the ArcFace model from scratch using the private MS-Celeb-1M dataset.
Finetuned Model: Loads pretrained weights from the original ArcFace model and fine-tunes it on the private MS-Celeb-1M dataset.

Figure 3 presents face recognition performance on AgeDB-30 and CFP-FP datasets, demonstrating significantly degraded performance for all three testing approaches when using PrivHAR-optimized images.

4. Precision-Recall (PR) and Receiver Operating Characteristic (ROC) Curves:

The paper analyzes the performance of adversarial networks used in PrivHAR through PR and ROC curves.

Rubiksnet Backbone: The best-performing adversarial network achieves AUC values close to a random classifier (p<0.01), indicating limited discriminative power against PrivHAR-optimized images.

C3D Backbone: Similar results are observed with C3D as the backbone, further emphasizing the difficulty of detecting human actions from PrivHAR-processed data.

5. Additional Experiments and Discussion:

The supplementary material includes:

A video showcasing more qualitative results and failure cases of the proposed PrivHAR network (with Rubiksnet backbone).
Details on hardware experiments used to validate the system's functionality.
Creators and license information for assets used in the paper.
Discussion on potential negative impacts of the work, emphasizing responsible AI development and deployment.
A section addressing personal data/human subjects considerations related to using camera-based privacy-preserving techniques.