Analysing Statistical methods for Automatic Detection of Image Forgery

  • 2021-11-24 17:48:26
  • Umar Masud, Anupam Agarwal
  • 0


Image manipulation and forgery detection have been a topic of research formore than a decade now. New-age tools and large-scale social platforms havegiven space for manipulated media to thrive. These media can be potentiallydangerous and thus innumerable methods have been designed and tested to provetheir robustness in detecting forgery. However, the results reported bystate-of-the-art systems indicate that supervised approaches achieve almostperfect performance but only with particular datasets. In this work, we analyzethe issue of out-of-distribution generalisability of the currentstate-of-the-art image forgery detection techniques through severalexperiments. Our study focuses on models that utilise handcrafted features forimage forgery detection. We show that the developed methods fail to performwell on cross-dataset evaluations and in-the-wild manipulated media. As aconsequence, a question is raised about the current evaluation andoverestimated performance of the systems under consideration. Note: This workwas done during a summer research internship at ITMR Lab, IIIT-Allahabad underthe supervision of Prof. Anupam Agarwal.


Quick Read (beta)

loading the full paper ...