Abstract
This paper presents and experimentally validates an electronic voting model that stores results in a decentralized blockchain while using biometric voter verification and automatic detection of synthetically generated (deepfake) facial images. For identity verification, Vision Transformer and ResNet-50 architectures are fine-tuned with LowRank Adaptation (LoRA). Frequency-domain information from the Fast Fourier Transform (FFT) is also included as a descriptor for detecting synthetic images. Since blockchain records cannot be altered, the model introduces a controlled revoting process based on self-destructing ballots and timestamps recorded on the blockchain; the votes themselves are submitted through an Ethereum smart contract. In the experiments, the LoRA-adapted ViT configuration distinguishes real from synthetic facial images with an accuracy of 97.48%, exceeding the performance of the LoRA-adapted ResNet-50 baseline. However, because no separate FFT ablation study has been conducted, this work does not claim that the frequency descriptor makes an independent contribution. The proposed system is presented as part of Smart City digital infrastructure and lays the groundwork for future research on privacypreserving biometric protocols and blockchain platforms capable of scaling to electoral processes.