Functional encryption (FE) gives the power to retain control of sensitive information and is particularly suitable in several practical real-world use cases. Using this primitive, anyone having a specific functional decryption key (derived from some master secret key) could only obtain the evaluation of an authorized function f over a message m, given its encryption. For many scenarios, the data owner is always different from the functionality owner, such that a classical implementation of functional encryption naturally implies an interactive key generation protocol between an entity owning the function f and another one managing the master secret key. We focus on this particular phase and consider the case where the function needs to be secret. In this paper, we introduce the new notion of blind functional encryption in which, during an interactive key generation protocol, the master secret key owner does not learn anything about the function f. Our new notion can be seen as a generalisation of the existing concepts of blind IBE/ABE. After a deep study of this new property and its relation with other security notions, we show how to obtain a generic blind FE from any non-blind FE, using homomorphic encryption and zero-knowledge proofs of knowledge. We finally illustrate such construction by giving an efficient instantiation in the case of the inner product functionality. This work was supported by the European Union H2020 PAPAYA Innovation Program Grant 786767.