Publication Date

Spring 2026

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Nada Attar

Second Advisor

Amith Kamath Belman

Third Advisor

Noha Elfiky

Keywords

GenderStego, Gait-based Gender Recognition, ST-GCN, Video Steganog- raphy, DCT, AES-256-CTR, HMAC, Biometric Privacy, Ethical AI, Computer Vision, YOLOv8

Abstract

The convenience of surveillance technology has raised several serious ethical questions regarding an individual’s right to privacy and the security of their biometric information. To address the conflict between the use of surveillance technology to protect society and the need to preserve individual anonymity, we propose a new method called GenderStego. GenderStego is a novel framework for gait-based gender recognition in conjunction with privacy-preserving video steganography. Traditional gender recognition methods rely on identifying facial or other visual characteristics. In contrast, this research demonstrates how our system uses a privacy-preserving pipeline to identify a person’s gender at long distances using only skeletal motion. Our proposed system employs a Spatial-Temporal Graph Convolutional Network

(ST-GCN) to analyse poses derived from skeletal keypoints tracked using the state- of-the-art keypoint tracker YOLOv8x-Pose, thereby removing the need for visual

identification. Our experiments demonstrate a validation accuracy of 72.65% on the large-scale, in-the-wild DenseGait dataset, without signs of overfitting. We also observe strong evidence that our model continues to improve even after reaching the maximum training budget of 20 epochs.

We additionally incorporate a privacy module that removes the subject’s visible appearance using YOLOv8-based object segmentation and HSV Gaussian inpainting, overlaying the subject with a blank background and colour-coded skeletons (blue: Male; red: Female), thereby providing comprehensive protection of biometric information contained in the original frames. The original video is then embedded into this skeletal output using a modified Discrete Cosine Transform (DCT) steganographic

scheme applied to 4 × 4 frequency-domain blocks. For added security, we encrypt

the hidden payload with AES-256-CTR for confidentiality and append an HMAC- SHA256 signature for tamper detection, enabling authorised entities to recover the

data for forensic analysis while keeping sensitive biometric information private from unauthorised parties. In addition to qualitative evaluations across multiple parameter settings, our results indicate that the recommended configuration (QP = 27, β = 30, α = 0.05) yields a stego PSNR of 34.48 dB—below the threshold of human perceptual detection—and a reconstruction PSNR of 36.51 dB with 99.34% cosine similarity relative to the original video. Alongside quantitative testing on six real-world surveillance videos, which demonstrates stable behaviour under typical conditions and graceful degradation in challenging environments, GenderStego emerges as one of the first systems to combine robust security with accurate video analytics for ethical AI and responsible surveillance.

Download

Available for download on Saturday, May 22, 2027

Share

COinS