We revisit the problem of anonymous communication, in which users wish to send messages to each other without revealing their identities. We propose a novel framework to organize and compare anonymity definitions. In this framework, we present simple and practical definitions for anonymous channels in the context of computational indistinguishability. The notions seem to capture the intuitive properties of several types of anonymous channels (Pfitzmann and Kohntopp 2001) (eg. sender anonymity and unlinkability). We justify these notions by showing they naturally capture practical scenarios where information is unavoidably leaked in the system. Then, we compare the notions and we show they form a natural hierarchy for which we exhibit non-trivial implications. In particular, we show how to implement stronger notions from weaker ones using cryptography and dummy traffic - in a provably optimal way. With these tools, we revisit the security of previous anonymous channels protocols, in particular constructions based on broadcast networks (Blaze et al. 2003), anonymous broadcast (Chaum 1981), and mix networks (Groth 2003, Nguyen et al. 2004). Our results give generic, optimal constructions to transform known protocols into new ones that achieve the strongest notions of anonymity.