Abstract

The proliferation of counterfeit coins poses a substantial threat to the integrity of monetary systems and the stability of financial markets. Advanced counterfeiting techniques allow these fraudulent coins to closely mimic genuine ones, complicating the detection process and necessitating robust methods capable of discerning minute differences between genuine and fake coins. This thesis addresses the problem of counterfeit coin detection by introducing a diverse dataset comprising high-resolution images of both Danish and Chinese coins, categorized into genuine and counterfeit sets across multiple years.
To tackle the detection task, we employ two advanced approaches: a Vision Transformer (ViT) model and a multimodal GPT-4 model. The ViT model leverages its self-attention mechanisms to capture intricate patterns and details within the coin images, while the GPT-4 model integrates both visual and textual data, utilizing various prompting techniques to enhance its performance. Our results show that the ViT model outperforms previous methods and the state-of-the-art in terms of accuracy and robustness, achieving a remarkable 99.31% accuracy. The GPT-4 model, although primarily designed for natural language processing, demonstrates promising capabilities in counterfeit detection, particularly with advanced prompting strategies like Chain-of-Thought and Generated Knowledge.
This research advances the current state-of-the-art in counterfeit coin detection and highlights the potential of few-shot learning and transfer learning in achieving high accuracy with limited training data.