The watchful forensic analyst: Multi-clue information fusion with background knowledge

Image forgery localization through the fusion of camera-based, feature-based and pixel-based techniques

“…As it is possible to observe from Table 2 the score obtained by the three approaches used in this work are comparatively good. In the future research we will focus especially on the fusion of the available information by a more effective strategy [22,23,24] than the empirical rules used in this contest. Demosaicking [21] 0.1013 Table 2: Comparisons of the three approaches described in this work with some state-of-the-art methods.…”

Unsupervised fusion for forgery localization exploiting background information

“…In order to define the mapping from the localization maps to BBAs (Eq. ( 4)), we adopt the method proposed in [13], choosing the following set of properties to locally characterize the reliability of each tool τ : the value of the last compression factor (if any) q 2 , the mean value µ and the std. deviation σ of pixels intensity and the the first compression factor q 1 , if present.…”

“…The generic analysis block is thus described by the vector v = (o τ , q 2 , µ, σ, q 1 ), where o τ denotes the value of the block in the map produced by tool τ (in our case, τ ∈ T = {AJPEG; NAJPEG; CFA}). By applying the approach proposed in [13], each vector is associated to scalar values L τ , N τ and D τ (see Eq. 4); as to the parameters required in [13], we used α = 0.85 and k = 12 for all tools, whereas γ = 0.5 for CFA tool, γ = 512 for AJPEG tool and γ = 2048 for NAJPEG tool.…”

“…Other approaches work at the score level, where the scalar output of the tools is considered during fusion [11,12]. The overall performance of the above methods can be further improved by taking into account background information during fusion [13]. As to forgery localization, simple pixel-level fusion of different forensic tool outputs has been investigated in [14,15]: however, these works do not consider tools reliability and compatibility, and they are based on very simple rules for fusion (sum, product, logical disjunction/conjunction).…”

“…It is here that tool-based background information enters the picture: besides considering the value of the localization map in the block i, a set of local properties of the image is evaluated (e.g., mean value or variance of pixels in the analysis block i) and used to determine the mentioned values for equation (4). To perform this mapping from tool outputs and background information to BBAs, we rely on the method recently proposed in [13]: such method exploits a set of training images to learn how local properties affect the output of the tool. Thus, given image and forgery localization map, using this approach we obtain values for (4) for each block of pixels.…”