Information similarity metrics in information security and forensics.

摘要

We study two information similarity measures, relative entropy and the similarity metric, and methods for estimating them. Relative entropy can be readily estimated with existing algorithms based on compression. The similarity metric, based on algorithmic complexity, proves to be more difficult to estimate due to the fact that algorithmic complexity itself is not computable. We again turn to compression for estimating the similarity metric. Previous studies rely on the compression ratio as an indicator for choosing compressors to estimate the similarity metric. This assumption, however, is fundamentally flawed. We propose a new method to benchmark compressors for estimating the similarity metric. To demonstrate its use, we propose to quantify the security of a stegosystem using the similarity metric. Unlike other measures of steganographic security, the similarity metric is not only a true distance metric, but it is also universal in the sense that it is asymptotically minimal among all computable metrics between two objects. Therefore, it accounts for all similarities between two objects. In contrast, relative entropy, a widely accepted steganographic security definition, only takes into consideration the statistical similarity between two random variables. As an application, we present a general method for benchmarking stegosystems. The method is general in the sense that it is not restricted to any covertext medium and therefore, can be applied to a wide range of stegosystems. For demonstration, we analyze several image stegosystems using the newly proposed similarity metric as the security metric. The results show the true security limits of stegosystems regardless of the chosen security metric or the existence of steganalysis detectors. In other words, this makes it possible to show that a stegosystem with a large similarity metric is inherently insecure, even if it has not yet been broken.