Research Article
DCAN-FASSP: A Dual-Path Cross-Modal Adapter Network with Frequency-Aware Self-Supervised Pretraining for Robust and Explainable Deepfake Detection
@INPROCEEDINGS{10.4108/eai.19-12-2025.2366190, author={Nitish Kumar and Chirag Sharma and Anshu Sharma}, title={DCAN-FASSP: A Dual-Path Cross-Modal Adapter Network with Frequency-Aware Self-Supervised Pretraining for Robust and Explainable Deepfake Detection}, proceedings={Proceedings of the 1st International Conference on Computational Techniques in Data Science, IICTDS 2025, 19-20 December 2025, Chandigarh, India}, publisher={EAI}, proceedings_a={IICTDS}, year={2026}, month={11}, keywords={Deepfake Detection Vision Transformers Frequency Analysis Self-Supervised Learning Domain Generalization Explainable AI}, doi={10.4108/eai.19-12-2025.2366190} }- Nitish Kumar
Chirag Sharma
Anshu Sharma
Year: 2026
DCAN-FASSP: A Dual-Path Cross-Modal Adapter Network with Frequency-Aware Self-Supervised Pretraining for Robust and Explainable Deepfake Detection
IICTDS
EAI
DOI: 10.4108/eai.19-12-2025.2366190
Abstract
Recent developments in generative models have made the generation of hyper-realistic synthetic media, also known as deepfakes, possible and undetectable to the naked eye. Although the identified approaches to deep learning-based detection have proven successful, generalization of unseen manipulation techniques remains weak, while model interpretation and efficiency tend to be compromised. The presented paper presents DCAN-FASSP, a Dual-path Cross-modal Adapter Network with Frequency-Aware Self-supervised Pretraining, which combines semantic, frequency, and geometric information and guarantees effective deep fake detection. The framework combines lightweight adapter-based Vision Transformers (ViTs) with frequency- and geometry-aware CNNs, trained through a forgery-aware self-supervised pretraining scheme and a domain-adversarial head for manipulation-agnostic learning. Evaluations on FaceForensics++, Celeb-DF, and DFDC datasets demonstrate superior cross-dataset generalization (AUC=0.991, EER=2.1%), outperforming current state-of-the-art methods while remaining computationally efficient and explainable.