Academic Fraud Detection Based on Handwritten Documents Using Siamese ResNet Approach

Abstract

Academic cheating, particularly involving the forgery of handwriting documents, has become a significant challenge in the field of education. One of the most difficult modes to detect is identity misuse, where an individual writes or completes tasks on behalf of someone else. This study aimed to develop an academic cheating detection model based on handwriting using the Siamese Network approach and cosine similarity. The experiment was conducted using an HP Z8 G5 device equipped with two NVIDIA RTX 6000 GPUs, and the dataset used came from Universitas Sjakhyakirti. The dataset consisted of 101,475 pairs of handwriting images, each labeled as 1 (similar) for pairs from the same writer and 0 (dissimilar) for pairs from different writers. The data was divided into 70% for training, 15% for validation, and 15% for testing. This research dataset was sourced from handwriting documents of 450 different students, consisting of 450 positive pairs (label = 1) and 101,025 negative pairs (label = 0). The model was evaluated using a cosine similarity threshold of 0.5, with training accuracy reaching 95.34%, validation accuracy at 84.12%, and testing accuracy at 83.87%. This study contributes to the development of a handwriting-based academic cheating detection system that can be implemented in higher education institutions.