Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{206461,
        author = {Simrithaa N V and Royo J Varghese and Mithun Karthik and Shri Kirthikha Gunasekaran and Surya R},
        title = {Robust Deepfake Audio Detection Using Multi-Branch Feature Fusion and Adversarial Hardening},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {13},
        number = {2},
        pages = {1663-1675},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=206461},
        abstract = {Rapid progress in neural text-to-speech and voice conversion has made synthetic speech almost indistinguishable from genuine human speech, creating serious risks for speaker verification, digital forensics, and financial security. Although recent spectro-temporal and convolutional neural network (CNN) detectors achieve high accuracy on benchmark datasets, they remain vulnerable to adversarial perturbations and generalize poorly to real-world audio. This paper presents a robust deepfake speech detector that combines three complementary branches interpretable classical features, an EfficientNet-based CNN operating on log-Mel spectrograms, and self-supervised learning (SSL) embeddings fused at the score level. The CNN branch is hardened using Fast Gradient Sign Method (FGSM) adversarial training, and domain-adaptive augmentation is applied during training to improve robustness under codec, channel, and noise variation. The system is trained and evaluated on the ASVspoof 2021 Logical Access dataset for both binary detection and multi-class spoof attribution. The fused model achieves 99.30% binary accuracy and an Equal Error Rate (EER) of 0.97%, together with 94.20% multi-class attribution accuracy, outperforming the individual branches and the spectro-temporal baseline while remaining stable under adversarial attack.},
        keywords = {Adversarial training, deepfake audio detection, feature fusion, self-supervised learning, spoofing countermeasures.},
        month = {July},
        }