Attribute-efficient learning of linear threshold functions under unconcentrated distributions