There has been proposed a data embedding method based on the phase modulation of the two-dimensional sampling lattice to image coding in which the smooth control to the tradeoff between the reliability of embedded information and the quality of decoded image can be attained. In this paper, the effect of the simplest collusion attack with this method, based on the use of the arithmetic mean of pixel values by betrayers is examined. The results show that the method can disclose who they are.
The traditional adding template method has a serious flaw in that the patterns correctly recognized before becomes to error as dictionary learning progresses. To solve this problem, we present a new adding template method that embeds erroneous pattern as a hypersphere with a small radius into Voronoi region generated per category. Also, we present a hierarchical system that outputs the recognition result from Voronoi regions given by reference templates as the first candidate. Next, hypershperical classifiers embedded in the Voronoi region of the first candidate determine the correct candidate. The hierarchical system promotes the recognition rate by dictionary learning and succeeds in overcoming the serious flaw in experiments on handwritten Kanji pattern recognition.