Enhancing Deepfake Detection using SE Block Attention with CNN

Authors: Subhram Dasgupta, Janelle Mason, Xiaohong Yuan, Olusola Odeyomi, Kaushik Roy

Published: 2025-06-12 13:29:26+00:00

AI Summary

This research proposes a lightweight Convolutional Neural Network (CNN) enhanced with a Squeeze and Excitation (SE) block for deepfake detection. The SE block improves efficiency and accuracy by focusing on informative features, resulting in a model that achieves high accuracy with minimal computational resources.

Abstract

In the digital age, Deepfake present a formidable challenge by using advanced artificial intelligence to create highly convincing manipulated content, undermining information authenticity and security. These sophisticated fabrications surpass traditional detection methods in complexity and realism. To address this issue, we aim to harness cutting-edge deep learning methodologies to engineer an innovative deepfake detection model. However, most of the models designed for deepfake detection are large, causing heavy storage and memory consumption. In this research, we propose a lightweight convolution neural network (CNN) with squeeze and excitation block attention (SE) for Deepfake detection. The SE block module is designed to perform dynamic channel-wise feature recalibration. The SE block allows the network to emphasize informative features and suppress less useful ones, which leads to a more efficient and effective learning module. This module is integrated with a simple sequential model to perform Deepfake detection. The model is smaller in size and it achieves competing accuracy with the existing models for deepfake detection tasks. The model achieved an overall classification accuracy of 94.14% and AUC-ROC score of 0.985 on the Style GAN dataset from the Diverse Fake Face Dataset. Our proposed approach presents a promising avenue for combating the Deepfake challenge with minimal computational resources, developing efficient and scalable solutions for digital content verification.