Document is current

Article History

Received: 14 October 2025

Accepted: 10 February 2026

First Online: 2 March 2026

Declarations

:

: The authors declare no competing interests.

: This research was conducted with strict adherence to ethical guidelines for digital media and machine learning research. The primary dataset used for training the attention model, UCF-101, is a publicly available and widely accepted benchmark for academic research. The 50-video test set curated for the deepfake experiments was sourced exclusively from public stock video repositories, Pexels and Pixabay. All videos from these platforms are provided under licenses that grant the right to use and modify the content for both commercial and non-commercial purposes, which covers the scope of our research. There were no human subjects directly involved in this study so that the Institutional Review Board (IRB) approval is required. The people in the stock videos had already signed a document that allowed their pictures to be used publicly and edited according to the conditions of the Pexels/Pixabay license. No specific person identifiable data were gathered and the study was carried out on existing and publicly available data. We admit that technologies of deepfakes and watermarking fall into the category of dual usage. Nevertheless, our model is specifically created to be defensive: to safeguard content creators and confirm authenticity of digital media, which would prevent the misinformation propagation. The structure and location of the watermark depends on our proprietary model of spatio-temporal attention. Attackers that do not have access to this secret model are not able to extract, copy, or forge a valid watermark to a new content. This secrecy of the model architecture is a big obstacle to illegitimate use of the watermarking technology.