Comparative Study of Convolutional Neural Networks for Deepfake Video Detection on Social Media

Lamia Omar; Idris Skloul Ibrahim; Mehran Sharghi

Comparative Study of Convolutional Neural Networks for Deepfake Video Detection on Social Media

Lamia Omar^*
, Idris Skloul Ibrahim
, Mehran Sharghi

^*Corresponding author for this work

Research output: Contribution to conference › Paper › peer-review

Abstract

This study evaluates four CNN architectures—XceptionNet, EfficientNetB4, EfficientNetB7, MesoNet, and ResNet152—for deepfake detection across controlled and real-world datasets. We assess performance on DeepfakeTIMIT (controlled) and WildDeepfake (internet-sourced with compression artifacts). Results reveal a substantial performance gap between datasets. On DeepfakeTIMIT, models achieved 79.17%-100% accuracy, with MesoNet demonstrating perfect performance. However, accuracy dropped significantly on WildDeepfake to 40.46%-69.21%, with ResNet152 achieving the highest real-world performance. The findings emphasize that evaluation on challenging, diverse datasets is critical, as controlled benchmarks may overestimate model readiness for deployment.

Original language	English
Number of pages	7
Publication status	Published - 14 Nov 2025
Event	International Conference on Advances in Artificial Intelligence for Society (ICA2S 2025) - Online (Hosted by Gulf Medical University, Ajman, UAE), Ajman, United Arab Emirates Duration: 13 Nov 2025 → 14 Nov 2025 Conference number: 1 https://www.ica2s.com/

Conference

Conference	International Conference on Advances in Artificial Intelligence for Society (ICA2S 2025)
Country/Territory	United Arab Emirates
City	Ajman
Period	13/11/25 → 14/11/25
Internet address	https://www.ica2s.com/

Keywords

Deepfake Detection Convolutional Neural Networks (CNN) Social Media Security Artificial Intelligence Deep Learning Video Forensics Computer Vision

Cite this

@conference{69690740cfec4a59afb2a286f468b7c1,

title = "Comparative Study of Convolutional Neural Networks for Deepfake Video Detection on Social Media",

abstract = "This study evaluates four CNN architectures—XceptionNet, EfficientNetB4, EfficientNetB7, MesoNet, and ResNet152—for deepfake detection across controlled and real-world datasets. We assess performance on DeepfakeTIMIT (controlled) and WildDeepfake (internet-sourced with compression artifacts). Results reveal a substantial performance gap between datasets. On DeepfakeTIMIT, models achieved 79.17\%-100\% accuracy, with MesoNet demonstrating perfect performance. However, accuracy dropped significantly on WildDeepfake to 40.46\%-69.21\%, with ResNet152 achieving the highest real-world performance. The findings emphasize that evaluation on challenging, diverse datasets is critical, as controlled benchmarks may overestimate model readiness for deployment.",

keywords = "Deepfake Detection Convolutional Neural Networks (CNN) Social Media Security Artificial Intelligence Deep Learning Video Forensics Computer Vision",

author = "Lamia Omar and Ibrahim, \{Idris Skloul\} and Mehran Sharghi",

year = "2025",

month = nov,

day = "14",

language = "English",

note = "International Conference on Advances in Artificial Intelligence for Society (ICA2S 2025) ; Conference date: 13-11-2025 Through 14-11-2025",

url = "https://www.ica2s.com/",

}

TY - CONF

T1 - Comparative Study of Convolutional Neural Networks for Deepfake Video Detection on Social Media

AU - Omar, Lamia

AU - Ibrahim, Idris Skloul

AU - Sharghi, Mehran

N1 - Conference code: 1

PY - 2025/11/14

Y1 - 2025/11/14

N2 - This study evaluates four CNN architectures—XceptionNet, EfficientNetB4, EfficientNetB7, MesoNet, and ResNet152—for deepfake detection across controlled and real-world datasets. We assess performance on DeepfakeTIMIT (controlled) and WildDeepfake (internet-sourced with compression artifacts). Results reveal a substantial performance gap between datasets. On DeepfakeTIMIT, models achieved 79.17%-100% accuracy, with MesoNet demonstrating perfect performance. However, accuracy dropped significantly on WildDeepfake to 40.46%-69.21%, with ResNet152 achieving the highest real-world performance. The findings emphasize that evaluation on challenging, diverse datasets is critical, as controlled benchmarks may overestimate model readiness for deployment.

AB - This study evaluates four CNN architectures—XceptionNet, EfficientNetB4, EfficientNetB7, MesoNet, and ResNet152—for deepfake detection across controlled and real-world datasets. We assess performance on DeepfakeTIMIT (controlled) and WildDeepfake (internet-sourced with compression artifacts). Results reveal a substantial performance gap between datasets. On DeepfakeTIMIT, models achieved 79.17%-100% accuracy, with MesoNet demonstrating perfect performance. However, accuracy dropped significantly on WildDeepfake to 40.46%-69.21%, with ResNet152 achieving the highest real-world performance. The findings emphasize that evaluation on challenging, diverse datasets is critical, as controlled benchmarks may overestimate model readiness for deployment.

KW - Deepfake Detection Convolutional Neural Networks (CNN) Social Media Security Artificial Intelligence Deep Learning Video Forensics Computer Vision

M3 - Paper

T2 - International Conference on Advances in Artificial Intelligence for Society (ICA2S 2025)

Y2 - 13 November 2025 through 14 November 2025

ER -

Comparative Study of Convolutional Neural Networks for Deepfake Video Detection on Social Media

Abstract

Conference

Keywords

Fingerprint

Cite this