On digital image trustworthiness

• Three trustworthiness evaluation models based on information fusion are proposed.
• Dempster-Shafer evidence theory is used at three information fusion levels.
• An improved least square method is presented to reduce evidence conflicts.
• The method can be adapted for digital image forensic with high detection accuracy.

Digital images are facing a crisis of trustworthiness with the emergence of various digital image processing and steganography tools. This paper proposes a novel approach that can evaluate the trustworthiness of a digital image. In this approach, an information fusion method is used to combine base digital image forensic models at the feature level and the decision level. When using different kinds of base forensic models to get supporting evidence for different kinds of digital image manipulations, there exist uncertainties introduced by base forensic models and conflicts among evidence provided by different forensic models. We use the Dempster-Shafer (D-S) evidence theory and an improved least square method to tolerate the uncertainties of forensic models and reduce the evidence conflicts. The lower and upper limits of digital image trustworthiness can then be reliably evaluated by the D-S theory. Three information fusion models based on the D-S theory are proposed. The first model uses the D-S theory at the feature fusion level. The second uses the D-S theory at the decision fusion level, where an improved least square method is designed to reduce the evidence conflicts. The last model is a combination of the first and the second one, where the D-S theory is applied at both the feature fusion and decision fusion levels. Experiments are carried out on four kinds of digital image manipulations. The experimental results show that the three proposed models are very stable in evaluating different kinds of natural images and tampering images. While the first model can only give the upper limit of the trustworthiness of a digital image, the second and the third one can give both lower and upper limits of the trustworthiness of a digital image, as well as the uncertainties of the evidence produced by base forensics models. Compared with the second model, the third one can further reduce the uncertainties. The experimental results also show that when a digital image undergoes many kinds of manipulations, our models can validly compute a soft degree to measure the trustworthiness of the image, while current ordinary digital image forensic models may fail to predict it correctly. Experimental results also demonstrate that the proposed digital image trustworthiness evaluation models can be adapted as digital image forensic classification models with very high detection accuracy.

Figure optionsDownload as PowerPoint slide