An Identity-Aware Privacy-Preserving Deep Learning Framework for Culturally Sensitive Image Sharing

The blistering development of digital image sharing raises privacy concerns, especially in cultural contexts where image exposure could be ethically and socially provocative. In Islamic societies, sharing images of women without hijab can be deeply sensitive. This study presents SITR, an identity-sensitive privacy-preserving deep learning system aimed at reducing un-intended sharing of sensitive images involving female family members without hijab. SITR integrates three components in a unified deployment-ready pipeline: 1) face recognition with Multi-task Cascaded Convolutional Networks (MTCNN), 2) family-member authentication with FaceNet embeddings stored in a vector database, and 3) hijab detection with an optimized Densely Connected Convolutional Network (DenseNet). The hijab detection model was trained and evaluated on a cleaned dataset of 2,191 images with hijab and non-hijab cases with diverse visual conditions. DenseNet121 was benchmarked against ResNet50, MobileNetV2, and EfficientNet-B0, achieving the best overall performance. To further enhance its effectiveness, DenseNet121 was modified by integrating an Efficient Channel Attention (ECA) mechanism and applying hyperparameter tuning. The optimized selected model achieved 92.16% test accuracy, strong discrimination with precision of 91.63% , and 86.39% F1-score on a held-out test set. The model was deployed as a quantized RESTful API, reduced from 82 MB to 27 MB while maintaining predictive reliability. Results demonstrate the practicability of identity-conditioned, culturally-aware AI systems for privacy protection. This work highlights the role of context-sensitive computer vision beyond generic content moderation toward culturally-aware and ethically accountable applications.