But given that Alice randomly chooses the linear basis for each measurement on each photon, she is effectively generating a bit-wise one-time pad. Bob–by either converting the linear states to circular states or not–is either returning the pad to her randomized, or intact. Because Alice and only Alice can know the original pad (both the basis and the measurement outcome) only Alice can read Bob’s message.
The use of a 0->0 EPR-type source, which are very nearly off-the shelf devices these days, makes the whole thing pretty easy to implement, as follows:
The source Σ emits correlated photon pairs. Alice measures one member of the pair in some linear basis B, which she changes randomly for each pair. The other member goes to Bob, who reflects it back to Alice via the mirrors M and inserts (or not) a 1/4 λ plate into the beam. Alice then measures the linear polarization of the returning photon in basis B’ == B, and generates the correlation function C. If Bob has left the beam undisturbed the correlation will be perfect. If Bob has converted the linear polarization states to elliptical ones the correlation will be degraded.
The whole thing works because there is no way to distinguish an unpolarized beam of linearly polarized photons from an unpolarized beam of elliptically polarized photons (if there were it would be possible to sent superluminal signals via EPR correlations.) Only Alice, who knows the basis states of the original beam, can tell the difference.
I’m reasonably sure that this is not an original idea, but it’s been fun to work it out.