Journal of Al-Qadisiyah for Computer Science and Mathematics

The rapid rise in the use of artificially generated faces has significantly increased the risk of identity theft in biometric authentication systems. Modern facial recognition technologies are now vulnerable to sophisticated attacks using printed images, replayed videos, and highly realistic 3D masks. This creates an urgent need for advanced, reliable, and mobile-compatible fake face detection systems. Research indicates that while deep learning models have demonstrated strong performance in detecting artificially generated faces, deploying these models on consumer mobile devices remains challenging due to limitations in computing power, memory, privacy, and processing speed. This paper highlights several key challenges: (1) optimizing deep learning models to operate efficiently on mobile devices, (2) ensuring real-time inference without compromising accuracy, (3) maintaining user privacy when processing sensitive facial data, and (4) addressing the variability in mobile phone cameras, input resolution, and platform limitations across Android and iOS. Furthermore, the increasing sophistication of identity spoofing attacks—such as 3D masks and AI-generated faces—demands more sophisticated, robust, and multimodal detection technologies. The research findings provide a clear roadmap toward practical solutions. By evaluating the latest deep learning architectures, datasets, and anti-spoofing metrics, the study proposes a comprehensive React Native deployment path using TensorFlow Lite and TensorFlow.js to ensure cross-platform compatibility. The proposed system offers a unified classification of identity spoofing attacks and defense mechanisms, along with a structured evaluation framework that compares on-device processing with server-side detection. The results demonstrate that optimized models can achieve high accuracy, low false accept/rejection rates, and sub-second processing speeds on mobile devices. Ultimately, the study provides practical design guidelines for building robust, privacy-preserving, efficient, and real-world consumer-grade fake face detection systems.

Chest X-ray is the most popular examination type for thoracic diseases, but its interpretation exhibits error rates which are still subject to inter-observer variability and economic workload constraints. This work can be found in this paper: "A reproducible deep learning pipeline for pneumonia identification vs normal based on DenseNet-121 from chest X-ray". The dataset originated from the NIH ChestX-ray14 corpus and was downsampled to 8,500 frontal radiographs (1,050 pneumonia-positive, 7,450 normal) and split at the patient level into training, validation and testing sets. Preprocessing: Grayscale normalization, Resize, Targeted Augmentation and Training (with) Early Stopping, Learning Rate Scheduling, Class Weighting and Post-Hoc Probability Calibration. In the held-out test set, the model achieved ROC-AUC: 0.87, PR-AUC: 0.72, as well as a general accuracy of 93.2%, sensitivity: 82.8% and specificity: 94.6%. Calibration analysis contributed to improving the Brier score from 0.042 to 0.019 and led to good-fitting reliability curves. Interpretability was built into the inference using Grad-CAM and Integrated Gradients, with explanation faithfulness quantitatively checked (deletion AUC = 0.84, insertion AUC = 0.87, sanity check pass rate = 98%, pointing-game hit rate = 76%). Based on the above results, it can be seen that CNN-based diagnosis is promising to achieve a good accuracy as well as interpretability and reproducibility simultaneously. Hence, the proposed framework provides a white-box baseline for clinical examination and future multi-label thoracic disease detection extensions.