There several issues related to your question.
First one is that you have to distinguish between viewshed (visibility from a point) and reverse viewshed (visibility of a point). The difference is that in the first case the OFFSETA refers to height of observer above surface and OFFSETB refers to height of target above surface. But if you are interested in reverse viewshed, then offsets have to be also switched. OFFSETA will represent height of the target (tower in your case) above surface and OFFSETB will be height of observer (most likely human in you case). Under these assumptions you can calculate the binary reverse viewshed with the use of Viewshed tool.
The second issue is how to actually archieve the result that you want. Ideally you could calculated so called extended viewshed, where you would be intrested in knowing the visibile height of target. From the outcome you would just reclassify the pixel with value of at least 3 and you would get the result that you want. While such extented viewshed are described in literature, they are not implemented in any currently available software. So you are stuck with the Viewshed tool. Fortunatelly, you can do it by calculation of viewshed, where OFFSETA will be equal to height of the tower minus 3 meters and OFFSETB will be height of observer.
If you would be interested only in visibility of the tower from specific locations (points) then it is possible to use "Line of Sight Analyst", which is a toolbox that calculates some extented viewshed for line of sights between observing and target points. You can take a look at the documentation. The process for you would be following: Create Lines of Sight -> Analyze of Lines of Sight -> Select results with attribute ElevDiff_H higher than 3.
I should probably note that I am the author of "Line of Sight Analyst".