Evaluating Xception with Multitask Self-Supervised Learning for AI-Generated Image and Deepfake Video Detection
Abstract
This study evaluates an Xception-based Multitask Self-Supervised Learning (MSSL) framework for detecting real and synthetic visual content across image and video modalities. Separate models were developed for CIFAKE images and FaceForensics++ videos. Beyond the primary binary-classification objective, MSSL integrates three auxiliary representation-learning tasks: rotation prediction, patch reconstruction, and artifact-map prediction. These tasks are intended to encourage the model to learn robust semantic and manipulation-related visual features. Experiments used class-balanced subsets of 5,000 CIFAKE images and 840 FaceForensics++ videos. For video classification, five frames were selected through uniform temporal sampling, and frame-level probabilities were averaged to obtain video-level predictions; a three-frame setting was also examined. A fixed decision threshold of 0.50 was defined a priori and applied consistently across both modalities. On CIFAKE, Xception+MSSL achieved 92.60% accuracy and 0.9797 ROC-AUC, outperforming the baseline Xception model, which obtained 79.60% accuracy and 0.8937 ROC-AUC. On FaceForensics++, five-frame MSSL achieved 68.25% video-level accuracy and 0.7365 ROC-AUC, compared with 62.70% and 0.6652 for Xception. The results indicate consistent within-dataset improvements, although cross-dataset and cross-generator generalization remain unverified.
Related
- OUT-OF-DISTRIBUTION AI-GENERATED IMAGE DETECTION WITH ZERO-PHASE DECONVOLUTION FOR RESIDUAL WHITENING
- Steer-to-Detect: Probing Hidden Representations for Detection of LLM-Generated Texts
- Dual-domain self-supervised feature alignment via spectral–spatial representation learning for deepfake anomaly detection
- Robust Fault Detection in Mechanical Multimodal Time Series via Self-Supervised Cross-Modal Reconstruction
- Robust Detection of LLM-Generated Text under Contamination
BibTeX
@article{rochim2026evaluating,
title = {{Evaluating Xception with Multitask Self-Supervised Learning for AI-Generated Image and Deepfake Video Detection}},
author = {Iksan Nur Rochim and R. Rizal Isnanto and Aris Sugiharto},
journal = {Journal of Information Systems and Informatics},
year = {2026},
month = oct,
doi = {10.63158/journalisi.v8i5.1864},
url = {https://doi.org/10.63158/journalisi.v8i5.1864},
note = {CC BY},
}
RIS
TY - JOUR
TI - Evaluating Xception with Multitask Self-Supervised Learning for AI-Generated Image and Deepfake Video Detection
AU - Iksan Nur Rochim
AU - R. Rizal Isnanto
AU - Aris Sugiharto
PY - 2026
DA - 2026/10/24
JO - Journal of Information Systems and Informatics
DO - 10.63158/journalisi.v8i5.1864
UR - https://doi.org/10.63158/journalisi.v8i5.1864
AB - This study evaluates an Xception-based Multitask Self-Supervised Learning (MSSL) framework for detecting real and synthetic visual content across image and video modalities. Separate models were developed for CIFAKE images and FaceForensics++ videos. Beyond the primary binary-classification objective, MSSL integrates three auxiliary representation-learning tasks: rotation prediction, patch reconstruction, and artifact-map prediction. These tasks are intended to encourage the model to learn robust semantic and manipulation-related visual features. Experiments used class-balanced subsets of 5,000 CIFAKE images and 840 FaceForensics++ videos. For video classification, five frames were selected through uniform temporal sampling, and frame-level probabilities were averaged to obtain video-level predictions; a three-frame setting was also examined. A fixed decision threshold of 0.50 was defined a priori and applied consistently across both modalities. On CIFAKE, Xception+MSSL achieved 92.60% accuracy and 0.9797 ROC-AUC, outperforming the baseline Xception model, which obtained 79.60% accuracy and 0.8937 ROC-AUC. On FaceForensics++, five-frame MSSL achieved 68.25% video-level accuracy and 0.7365 ROC-AUC, compared with 62.70% and 0.6652 for Xception. The results indicate consistent within-dataset improvements, although cross-dataset and cross-generator generalization remain unverified.
SN - 2656-5935
SN - 2656-4882
ER -