Learning domain-invariant representation for generalizing face forgery detection