Enhancing Visual Multimedia Authenticity with Lightweight-DNN for Deepfake Detection

Abstract

Detecting visual Deepfakes is essential because of the increasing use of multimedia phishing in online social networks, which are shared to spread false news. This paper proposes a Deepfake detection approach leveraging normal images and frequency domain representations using the Discrete Fourier Transform (DFT). The method, built using Python and Deep Learning (DL), applies a light-weight Deep Neural Network (DNN) for model training. Authentic and forged videos are acquired to build the dataset, followed by pre-processing. Based on MesoNet’s DNN architecture designed with Convolutional Neural Network (CNN), the model detects fraud in individual images and aggregates results to classify videos. Output values are 0 for Deepfakes and “1” for real videos in the time domain, while DFT images yield higher “mean aggregate” scores for original inputs. The model focuses on spatial and frequency domain feature analysis and achieves an accuracy of 83%. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.