Quantitative steganalysis of spatial LSB based stego images using reduced instances and features

Universal quantitative steganalysis suffers from the curse of dimensionality and requires large number of instances (training samples) to produce qualitative results. In this paper, a universal quantitative steganalyser for spatial LSB based algorithms using reduced number of instances and features is proposed. A rich combination of global and local features models are employed as core features. The AdaBoost ensemble regressor with regression trees as its base learners is used to estimate the change rate caused in the stego images. To obtain low dimensional features and training instances, a three level optimisation approach is proposed. The trilevel optimisation comprises of bilevel optimisation of feature selection and an intermediate level of novel instance selection method. The first level of optimisation is done for choosing the optimal concatenation of discriminant feature models from the core feature set by Greedy Randomised Adaptive Search Procedure (GRASP). The instances of the optimal concatenated feature model are selected by the proposed Discretized-All Condensed Nearest Neighbour (D-AllCNN) method. The optimally concatenated features of the reduced instances are then reduced dimensionally by Recursive Feature Elimination (RFE) feature selection process. This process yields an improved quantitative steganalyser working on reduced instances and features. Experimental results confirm that the proposed steganalyser is better than state-of-the-art quantitative steganalysers for both traditional non-adaptive (LSBR, LSBM, LSBMR, LSBR2, LSBRmod5) and content adaptive (HUGO, WOW and SW) spatial LSB based steganographic schemes.