The theory of noninterference supports the analysis of information leakage and the execution of secure computations in multi-level security systems. Classical equivalence-based approaches to noninterference mainly rely on weak bisimulation semantics. We show that this approach is not sufficient to identify potential covert channels in the presence of reversible computations. As illustrated via a database management system example, the activation of backward computations may trigger information flows that are not observable when proceeding in the standard forward direction. To capture the effects of back-and-forth computations, it is necessary to switch to a more expressive semantics, which has been proven to be branching bisimilarity in a previous work by De Nicola, Montanari, and Vaandrager. In this paper we investigate a taxonomy of noninterference properties based on branching bisimilarity along with their preservation and compositionality features, then we compare it with the taxonomy of Focardi and Gorrieri based on weak bisimilarity.