Is there a way to know if the sender of a message really exists before accepting his message?

The simple answer is no. The only way to verify if a sender exists is to do closed-loop validation, and that can't be done in realtime.

To do this you would have to accept the message, but defer it

while you do the validation. See TMDA for one implementation of this.