Abstract. The notion of key privacy for asymmetric encryption schemes was formally de ned by Bellare, Boldyreva, Desai and Pointcheval in 2001: it states that an eavesdropper in possession of a ciphertext is not able to tell which speci c key, out of a set of known public keys, is the one under which the ciphertext was created. Since anonymity can be misused by dishonest users, some situations could require a tracing authority capable of revoking key privacy when illegal behavior is detected. Prior works on traceable anonymous encryption miss a critical point: an encryption scheme may produce a covert channel which malicious users can use to communicate illegally using ciphertexts that trace back to nobody or, even worse, to some honest user. In this paper, we examine subliminal channels in the context of traceable anonymous encryption and we introduce a new primitive termed mediated traceable anonymous encryption that provides con dentiality and anonymity while preventing malicious users to embed subliminal messages in ciphertexts. In our model, all ciphertexts pass through a mediator (or possibly several successive mediators) and our goal is to design protocols where the absence of covert channels is guaranteed as long as the mediator is honest, while semantic security and key privacy hold even if the mediator is dishonest. We give security de nitions for this new primitive and constructions meeting the formalized requirements. Our generic construction is fairly e cient, with ciphertexts that have logarithmic size in the number of group members, while preventing collusions. The security analysis requires classical complexity assumptions in the standard model.