What question did this study set out to answer?

The aim is to develop a comprehensive framework for unimodal deepfake detection across different modalities.

June 7, 2026Open Access

VerbaFake, EchoFake, and PixelFake: lightweight architectures with advanced augmentation for unimodal deepfake detection

UAUsman AliUniversiti Brunei Darussalam MSMuhammad Rizwan ShoukatUniversity of Education MRMahjbeen RajputUniversity of Education

Key Points

The aim is to develop a comprehensive framework for unimodal deepfake detection across different modalities.
Developed lightweight detection models for images, audio, and text including CNNs and RNNs.
Utilized publicly available datasets with over 26,000 images, 20,000 audio clips, and 31,429 text entries.
Integrated specialized pipelines for each modality ensuring unimodal consistency.
Image detection models achieved 97% accuracy using MobileNetV2 and EfficientNetV2-S.
Text-based detection models reached 99% accuracy with VerbaFake and Attention-based BiLSTM.
Audio detection models attained accuracies of 96% for Xception and 98% for EchoFake.

Abstract

The rapid evolution of deepfake technology, driven by deep learning and generative models such as Generative Adversarial Networks, has revolutionized digital content creation while simultaneously introducing serious threats to privacy, security, and digital trust. This study addresses the increasing risks posed by deepfakes, synthetically generated images, audio, and text, by developing a comprehensive unimodal deepfake detection framework. The system leverages a range of deep learning models including CNNs, attention mechanism (NLP), GRU and BiLSTM (RNNs), and capsule networks. For image detection, models such as MobileNetV2, EfficientNetV2-S, and the PixelFake (scratch-built) model achieved high accuracy, with MobileNetV2 and EfficientNetV2-S reaching 97%. Text-based detection models, including a VerbaFake (scratch-built), Efficient-CapsNet, and Attention-based BiLSTM, consistently achieved 99% accuracy. Audio deepfake detection models, such as Xception, DenseNet201, and the EchoFake (scratch-built) model, attained accuracies of 96%, 97%, and 98%, respectively.The framework integrates specialized pipelines for each modality, ensuring unimodal consistency and enhancing detection reliability.Publicly available datasets from Kaggle and additional curated sources were used, comprising over 26,000 image samples, 20,000 audio clips, and 31,429 text entries.Experimental results confirm the system’s effectiveness, especially when employing CNN-based models.This research contributes a lightweight, real-time-capable solution with practical applications in journalism, cybersecurity, and law enforcement. While also raising awareness about the ethical implications of deepfakes and informing policy development. The proposed model codes and associated data are publicly available at https://github.com/UsmanAlExp/DeepFake-Detection GitHub repository,with DOI: https://doi.org/10.5281/zenodo.19232320 .

Ask AI

Helpful

Bookmark

View Full Paper