Securing the Integrity of Postpaid Electricity Billing Processes via MobileNetV2-Based Image Validation

The postpaid electricity meter reading process has a critical vulnerability: field officers can bypass actual readings by uploading fake photos due to penalty avoidance incentives. The current system only verifies numeric consumption trends without visual validation, leading to financial risks. This paper proposes a back-office photo validation system using MobileNetV2 to detect invalid submissions before billing. Trained and evaluated on a dataset of 6,633 field images (with a test set of 994 images) from PLN ULP Cianjur Kota, the model averaged across three random seeds achieves 97.42 percent accuracy with a standard deviation of 0.28 percent, a false negative rate of 1.30 percent with a standard deviation of 0.13 percent, a macro F1-score of 0.975 with a standard deviation of 0.003, and a specificity of 97.89 percent with a standard deviation of 0.42 percent. On an Intel i7-1255U CPU, the FP16 model achieves 6.51 milliseconds latency with a 5.76 megabyte footprint, while the INT8 variant reduces latency to 4.14 milliseconds with a 3.35 megabyte footprint. Grad-CAM confirms the model focuses on structural meter features. The system demonstrates excellent reproducibility with minimal variance across random initializations.