<![CDATA[The development of Generative Artificial Intelligence has produced realistic synthetic images that are difficult to distinguish from authentic images through visual inspection. This study aims to implement an Android-based mobile digital forensics application for detecting AI-generate d images using MobileNetV2 converted to TensorFlow Lite for on-device inference. A quantitative-experimental approach used 2,000 images, consisting of 1,000 authentic and 1,000 AI-generate d images. The authentic class comprised 500 smartphone photographs and 500 GenImage samples, while the AI-generate d class included Stable Diffusion v1.4, Stable Diffusion v1.5, Wukong, and Midjourney images. The dataset was divided into 1,000 training, 500 validation, and 500 testing images. Training data were used for model development, validation data for model selection and threshold determination, and testing data only for final evaluation. Robustness testing used copies of 500 testing images without retraining. Under normal conditions with a threshold of 0.680, the model achieved 85.40% accuracy, 85.53% precision, 85.40% recall, and an 85.39% F1-score. JPEG q=65 compression produced 86.00% accuracy. The largest degradation occurred with 112 × 112 resizing combined with JPEG q=65, resulting in 62.40% accuracy, a decrease of 23.00 percentage points, and a 56.66% F1-score. The application performed local inference, displayed prediction labels and confidence scores, and stored detection history. This study contributes an on-device detection application and robustness evaluation using a consistent test subset, positioning the system as an initial detection aid rather than a final forensic verification tool.]]>
