Security-oriented error-detecting codes are used to detect fault injection attacks on cryptographic devices. These codes are usually designed for uniformly distributed codewords, i.e., for codes that have maximal entropy. In practice, the codewords are not uniformly distributed; thus, their entropy is smaller and their efficiency in detecting attacks degrades. This paper analyzes the relation between the entropy of a code and its worst error masking probability. Based on this relation, a method for determining the rate and structure of a code that provides the required error masking probability is presented.