ManTra-Net: Manipulation Tracing Network for Detection and Localization of Image Forgeries with Anomalous Features

Posted on 21/10/2019, in Paper.
  • Overview: This paper designed new FCN layer that captures local anomaly and achieve better performance by modeling forgery detection as an image-segmentation problem.
  • ZPool2D: Given a feature map, the authors suggest that human identifies potential forged regions by comparing each region with a dominant features of its neighbours. The local anomaly can be captured by a local Z-score. More specifically, \begin{equation} Z_F[i, j] = (F[i,j] - \mu_F) / \sigma_F \end{equation}, where $\mu_F, \sigma_F$ are mean and standard deviation within an n by n window respectively. Window size need to be tuned in case the average are dominant by the forged region. In practice, $\sigma_F$ are clipped with a learnable parameter $w_\sigma$: \begin{equation} \sigma_F^* = \text{max}(\sigma_F, \epsilon + w_\sigma) \end{equation}
  • Experiments: Dresden Image Database provides the original image. During training, a random structured binary mask $M$ is generated. Within the masked region, four manipulation ($y$) are applied to the original image $P$, namely: a)splicing; b)copy-move; c)removal; d)enhancement. The final forged image is a blend of both: \begin{equation} Z = P \cdot (1-M) + y(P)\cdot M \end{equation} For evaluation, there are four datasets, NIST2016, CASIA, COVERAGE and Columbia datset, contains 564, 6044, 100 and 180 samples respectively.
  • Limitation: The authors identify several failed mode as: a) Highly correlated noise everywhere; b) Multple edits; c) Fully regenerated.


  • Forgeries detection seems to be a problem with unlimited training data, given enough augmenting approaches. It also the dual problem of adversial attack, where we try to leverage the sensitivity of NN.