EUVIMEDEUVIMED entdecken
Crossref

AV-CACNet: A Cross-Attention Consistency Network for Audio-Visual Deepfake Detection in Cybersecurity

Assem Turarbek, Sharipa Temirgaziyeva, Ahmet Sakir Dokuz, Alina Azanova, Nazgul Mametzhanova

International Journal of Advanced Computer Science and Applications · 2026 · 17 (10)

DOI: 10.14569/ijacsa.2026.0171095 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal of Advanced Computer Science and Applications, 2026. Verfasst von Assem Turarbek und 4 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The rapid development of audiovisual deepfake technologies has made the automatic verification of digital content authenticity a pressing issue, as methods relying solely on video or audio may fail to fully characterize complex forms of multimodal manipulation. This study proposes the AV-CACNet (Audio-Visual Cross-Attention Consistency Network) architecture, designed to jointly model the interdependencies between audio and visual features. The proposed method processes facial imagery, the mouth region, and audio information as separate data streams, integrating CBAM, Temporal Transformer, Face–Mouth Cross-Attention, Audio–Visual Cross-Attention, a Consistency Module, and Adaptive Gated Fusion mechanisms into a single architecture. Experiments were conducted using the FakeAVCeleb dataset. Based on a source-level splitting strategy, a balanced set of 1,996 videos derived from 499 unique sources was created and divided into training, validation, and test sets for four-class classification. Assessment on previously unseen test data yielded an accuracy of 88.82% for AV-CACNet. The associated macro-averaged precision, recall, and F1-score were 89.50%, 88.82%, and 88.76%, respectively. In the binary deepfake detection scenario, the model achieved a ROC-AUC of 97.76%, with an Equal Error Rate of 7.89%. The highest F1-score by class reached 91.93% for the FakeVideo-FakeAudio category. The results obtained for the four-class FakeAVCeleb classification outperformed previously published results from Multi-view CNN (85.96%) and AMID (87.65%). These results indicate that integrating the connections and coherence between audio and visual data can enhance the examination of multimodal deepfake materials.

Zitieren

APA

Turarbek, A., Temirgaziyeva, S., Dokuz, A. S., Azanova, A., & Mametzhanova, N. (2026). AV-CACNet: A Cross-Attention Consistency Network for Audio-Visual Deepfake Detection in Cybersecurity. International Journal of Advanced Computer Science and Applications, 17(10). https://doi.org/10.14569/ijacsa.2026.0171095

Vancouver

Turarbek A, Temirgaziyeva S, Dokuz AS, Azanova A, Mametzhanova N. AV-CACNet: A Cross-Attention Consistency Network for Audio-Visual Deepfake Detection in Cybersecurity. International Journal of Advanced Computer Science and Applications. 2026;17(10). doi:10.14569/ijacsa.2026.0171095

BibTeX
@article{Turarbek2026,
  author = {Turarbek, Assem and Temirgaziyeva, Sharipa and Dokuz, Ahmet Sakir and Azanova, Alina and Mametzhanova, Nazgul},
  title = {AV-CACNet: A Cross-Attention Consistency Network for Audio-Visual Deepfake Detection in Cybersecurity},
  journal = {International Journal of Advanced Computer Science and Applications},
  year = {2026},
  volume = {17},
  number = {10},
  doi = {10.14569/ijacsa.2026.0171095}
}
Kostenlos starten →