Mutual Effects of Face-Swap Deepfakes and Digital Watermarking-A Region-Aware Study

Area of Science:

• Digital Image Forensics
• Computer Vision
• Generative Adversarial Networks (GANs)

Background:

• Face-swapping technology is advancing rapidly, raising concerns about digital media authenticity.
• Current watermark placement strategies often assume localized edits, typically focusing on the face region.
• This assumption may not hold true for sophisticated face-swapping algorithms.

Purpose of the Study:

• To investigate the non-local effects of face-swapping on watermarks placed in image backgrounds.
• To quantify the impact of face-swapping on both visible and invisible watermark integrity.
• To evaluate the relationship between watermark strength, identity transfer, and retention across different face-swapping architectures.

Main Methods:

• Utilized a region-aware protocol with tunable-strength visible and invisible watermarks.
• Tested six distinct face-swap families on the VGGFace2 dataset.
• Measured watermark degradation using background-only Peak Signal-to-Noise Ratio (PSNR) and Pearson correlation, comparing against a locality-preserving baseline.

Main Results:

• Face-swapping generators demonstrably alter background statistics and degrade watermarks even in non-face regions, violating the locality assumption.
• The dependencies between watermark strength, identity transfer, and retention are complex, non-monotonic, and influenced by the specific GAN architecture.
• Segmentation-weighted methods that confine edits to the face preserve background signal integrity better than globally trained GANs; invisible watermarks in the background offer higher correlation than visible ones at similar perceptual distortion levels.

Conclusions:

• Classical robustness tests for watermarks are insufficient as they do not account for non-local edit propagation.
• Watermark evaluation protocols must incorporate region-wise metrics, considering watermark strength and the specific face-swapping architecture.
• Future research should focus on developing robust watermarking techniques and evaluation methods that address the non-local nature of generative image manipulation.