Unique substrings in genomes may indicate high level of specificity which is crucial and fundamental to many genetics studies, such as PCR, microarray hybridization, Southern and Northern blotting, RNA interference (RNAi), and genome (re)sequencing. However, being unique sequence in the genome alone is not adequate to guaranty high specificity. For example, nucleotides mismatches within a certain tolerance may impair specificity even if an interested substring occur only once in the genome. In this study we propose the concept of unique-m substrings of genomes for controlling specificity in genome-wide assays. A unique-m substring is defined if it only has a single perfect match on one strand of the entire genome while all other approximate matches must have more than m mismatches. We developed a pattern growth approach to systematically mine such unique-m substrings from a given genome. Our algorithm does not need a pre-processing step to extract sequential information which is required by most of other rival methods. The search for unique-m substrings from genomes is performed as a single task of regular data mining so that the similarities among queries are utilized to achieve tremendous speedup. The runtime of our algorithm is linear to the sizes of input genomes and the length of unique-m substrings. In addition, the unique-m mining algorithm has been parallelized to facilitate genome-wide computation on a cluster or a single machine of multiple CPUs with shared memory.