Researchers at the HSE MIEM analyzed 30 digital semi-fragile watermarking schemes published between 2016 and 2026. These watermarks are designed to remain intact during standard image processing, such as compression or brightness adjustments, while flagging malicious alterations like object insertion or face swapping.
The study reveals that 18 of the examined schemes rely on traditional methods of embedding marks into pixels or frequency characteristics, while 12 utilize neural networks. Among the reviewed methods, 25 can identify the specific area of an image that has been altered, and 10 are capable of restoring the damaged fragment.
The analysis highlights that 18 of the 30 methods operate blindly, meaning they do not require the original watermark to verify authenticity. The researchers categorized these schemes into three levels of resistance based on the number of permissible operations, ranging from low-resistance to high-resistance.
Despite these developments, the researchers identified significant challenges, including a lack of standardized metrics for evaluating sensitivity to tampering. Furthermore, many traditional schemes are tested on limited datasets, and neural network-based methods are often trained exclusively on facial images, limiting their broader application.
