Beamsplitting attack to the revised KKKP protocol and a possible solution

Core claim

On the paper's own terms, the central discovery is that a classically invisible observable—the parity $s_0 \oplus s_1$ of Bob's two random rotation choices—is encoded in the relative polarization of the two coherent pulses Bob sends back to Alice, and Eve can read that relative polarization with a beam splitter. Because the two returning pulses have equal amplitude and are phase-locked, splitting a fraction of each and mixing them at a 50:50 beam splitter produces a definite output-port pattern if the polarizations are parallel and a different pattern if they are orthogonal; choosing the amplitude $\gamma$ lets Eve match Bob's expected detection rate. Once Eve knows $s_0 \oplus s_1$, she impersonates Bob to Alice and Alice to Bob: she intercepts Alice's initial pulses, substitutes her own, later substitutes her own results for Bob, and composes rotations so that Bob's final measurement gives the bit value Alice intended, with Eve holding the key. The paper then revises the protocol by adding a random $\delta(\pi/4)$ rotation and test rounds, in order to make this specific parity-learning step cause detectable errors. The claimed conclusion is that the revised KKKP protocol fails for coherent-state keys unless a modification of this kind is used.