Vishal Asnani

Vishal Asnani

Research Scientist at Adobe

Trustworthy AI · Watermarking · Media provenance

Research / 2022

Proactive Image Manipulation Detection

Vishal Asnani, Xi Yin, Tal Hassner, Sijia Liu, Xiaoming Liu

Published in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022

Read paper Code repository All publications
Proactive detection learns a template set, adds a selected template to an image, and recovers its signal to distinguish manipulation.
Template protection happens before manipulation; the recovery network then looks for evidence of changes to that signal. Figure 2 from the CVPR 2022 paper. Select the figure to enlarge it.

Overview

Can images be prepared to make future manipulation easier to detect? This work moves part of the forensic process before the edit: a learned template is added to an original image, giving the detector a signal to examine later.

How it works

  • Learn a set of low-magnitude, high-frequency templates with constraints on diversity and recoverability.
  • Protect an image with a selected template before it reaches a manipulation model.
  • Recover the signal from protected or manipulated images and use it for classification.

A reported result

92.43% mean average precision across 12 unseen generative models, versus 82.97% for the Wang et al. baseline in Table 4.

The proactive model is trained using STGAN in this comparison. The number is an average of per-model average precision, not classification accuracy or a guarantee at a particular false-alarm rate.

Scope

Protection must be applied before manipulation. This is not a passive detector for arbitrary unprotected images, and performance varies by manipulation model.

Sources and resources

The numerical comparison is from Section 4 and Table 4.

Published paper Supplementary GitHub Project Webpage

Bibliographic reference

Asnani, V., Yin, X., Hassner, T., Liu, S., & Liu, X. (2022). Proactive Image Manipulation Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.