Gait-based gender spoofing detection using depth images

Gender transformation, particularly transgender transitions, has become a significant challenge in biometric security systems, as it complicates the identification of an individual's original gender based on their birth sex. This issue is especially prevalent in gait-based gender identification systems, which can be spoofed by individuals who have undergone gender transformation. To address this challenge, this study proposes a novel gait-based gender spoofing detection method using depth images. Given the absence of publicly available gait-spoofing datasets, a new dataset called spoofing gait dataset (SpooGa) was developed for this research. The SpooGa dataset contains depth images capturing individuals' walking styles, tailored specifically to the study's requirements. The proposed method comprises three main stages: pre-processing, feature extraction, and classification. During the pre-processing stage, the dataset is standardized to ensure uniformity in data dimensions. Feature extraction involves normalizing the depth images using gait energy images (GEI), which are then divided into three parts: the upper body, body, and lower body. This study focuses on the body and lower body parts, which are mapped onto a principal component analysis (PCA) plane to reveal distinctive cyclical patterns indicating changes in viewpoint. Features are subsequently extracted using the leg, toe, hand (LETH) formula. For classification, three independent methods are employed: linear support vector machine (linear SVM), fine decision tree, and weighted k-nearest neighbor (weighted KNN) classifier. The feature dataset is divided into training (70%) and testing (30%) subsets. The performance of the proposed method is evaluated based on its ability to correctly identify the original gender of individual's post-disguise. The experimental results demonstrate the effectiveness of the proposed method, achieving an accuracy of 92.30% with the linear SVM, 96.15% with the weighted KNN, and 92.30% with the fine decision tree classifiers. These findings indicate the potential of the proposed approach to enhance biometric security against gender spoofing attacks.