Abstract: 

The rectangular layout of typical camera sensors allows photographers to choose between landscape or portrait format. The choice between landscape and portrait format appears to be purely artistic, but can in fact have significant impact for security applications that operate on image statistics. One reason for this is that many state-of-the-art methods assume that image statistics are similar in the horizontal and vertical directions, allowing them to reduce the number of features (or trainable weights) by merging coefficients. On the one hand, this artificial symmetrization tends to suppress directional properties of natural images, causing a loss of performance. On the other hand, unaddressed directionality causes learning-based methods to overfit to a single orientation. This makes them vulnerable to manipulation if an adversary chooses inputs with the less common orientation.
This talk takes a comprehensive approach, identifies and systematizes causes of directionality at several stages of the image acquisition, and demonstrates its effects in three selected security applications (steganalysis, forensic source identification, and the detection of synthetic images).