Anytime-Valid Eavesdropper Monitoring for Low-Resource Semi-Quantum Links

Low-resource quantum communication protocols try to deliver as much message data as possible per photon or entangled pair. The best of them, such as the Tele-Fetch protocol, let a nearly classical client talk securely to a quantum server while using almost every qubit for data. They still need a separate supply of "check" qubits to detect eavesdroppers, and today that supply is a fixed block sized in advance. This proposal replaces the fixed block with a running statistical test that watches randomly interleaved check rounds throughout the session. The test stops the session the moment the evidence of tampering is strong enough. It also raises or lowers how often check rounds are inserted, based on how ambiguous the evidence currently is. Its false-alarm guarantee holds no matter when it is consulted, which is what "anytime-valid" means.

This matters because the fixed block wastes qubits in two ways. On a clean channel it spends the full block even though a strong attack would have been obvious after a handful of rounds. Against a weak attack it is too small to detect anything, and it only reaches a verdict at the end, after the message has already leaked. Rare-event statistics has a toolkit for this exact problem: deciding as early as possible from scarce counts while keeping error rates honest. The goal here is to bring that toolkit into the security layer of semi-quantum protocols. The yardstick is the one that actually matters for a keyless protocol, namely how many message bits an attacker gets before the protocol aborts.

Core Claim

In semi-quantum direct communication, a fixed-sample hypothesis test on a pre-allocated block of probing bits is the wrong tool. A betting martingale over interleaved check rounds, with the check-insertion rate adapted from the martingale's own state, should do better than the fixed block at the same false-abort rate on honest noisy channels. It should use fewer total probing qubits across a range of unknown attack strengths. It should also sharply cut the number of message bits leaked before abort against weak or delayed attacks.

The claim is not that sequential testing saves a few percent of samples on a single known alternative. The overshoot-avoidance work on Wald's test reports savings in that range: 3.4–11.8% for simple and 6.3–13.9% for composite alternatives (Improving Wald's (Approximate) Sequential Probability Ratio Test by Avoiding Overshoot). Gains that small would not justify a new protocol layer. The expected gains come from two things a fixed block cannot do: adapt to an attack strength nobody knows in advance, and cap leakage over the whole session rather than only at a checkpoint.

Background and Where This Sits

The motivating protocol is Lu, Barbeau and Nayak's keyless Tele-Fetch scheme and its rate-estimation variant (Keyless Semi-Quantum Point-to-point Communication Protocol with Low Resource Requirements). In that scheme, the server sends one half of an entangled pair, and the classical user either measures it or reflects it back. The server then recovers the user's result with a Bell measurement. This design keeps only one qubit of quantum memory and drives qubit efficiency toward 100%, compared with the 12.5–50% of earlier semi-quantum protocols.

To detect a measure-and-replay attack, where an eavesdropper measures each passing qubit and sends a replacement, the scheme runs a null-hypothesis test on $r$ probing bits. It reports about 99.5% detection with 15 probing bits at attack probability $p=0.6$ on a noiseless channel, and about 60 probing bits on a noisy one. That fixed-sample test is exactly the component this proposal replaces. The authors themselves flag that the test reserves a predetermined block.

The smart-grid semi-quantum protocol of Lou et al. (Privacy-Preserving Bidirectional Data Transmission of Smart Grid Via Semi-Quantum Computation) motivates the same move from another side. It fixes a signal-to-check ratio to reach 94.1% communication efficiency and reports roughly 2.1 s execution on edge-class hardware. That is evidence both that check overhead is the remaining efficiency lever and that the monitoring logic must be cheap.

The method comes from sequential testing. Wald-style tests minimise expected sample size, but their boundaries overshoot, and the overshoot distorts real error rates. The boosted, clipped test martingales of the overshoot paper restore guaranteed error control. We reuse this construction, but with a composite null (unknown but bounded channel noise) and a predictable, adaptively chosen sampling rate, which that paper does not treat.

The quantum coin-flipping analysis of Zhou et al. (Biased decoy-state measurement-device-independent quantum key distribution with finite resources) constrains the design. It shows that responding to noise-versus-cheating ambiguity by restarting the protocol lets a cheater's success probability climb from 0.5 to 0.75 after one restart and toward 1 with more. Any adaptive scheme must therefore avoid giving the adversary free retries. In our construction the evidence accumulates in one martingale that is never reset, which keeps a single time-uniform error budget.

Finite-key decoy-state analyses (Decoy States for Eavesdropping Detection in Quantum Communication; Finite-key analysis for quantum conference key agreement with asymmetric channels) use Chernoff and Hoeffding bounds on pre-allocated blocks. They are the natural fixed-block comparison for any extension to weak-pulse sources.

Research Program

The object we build is a monitoring layer that sits on top of the Tele-Fetch rounds. At each round $t$, the server privately decides, with probability $\pi_t$, whether the round is a check round. If it is, the round yields an error indicator $X_t \in \{0,1\}$; otherwise it carries a message bit.

The honest-channel null is that the per-check error rate is at most a calibrated noise ceiling $\eta_0$. The alternative is that an attack raises it by at least some unknown amount. The server maintains the wealth process

$$M_t = \prod_{s \le t,\ s \text{ a check}} \big(1 + \lambda_s (X_s - \eta_0)\big),$$

where $\lambda_s \in [0, 1/(1-\eta_0)]$ is a betting fraction chosen from data before round $s$. The protocol aborts when $M_t \ge 1/\alpha$, with the wealth clipped at that boundary as in the boosted-test construction. Under the null, $M_t$ is a nonnegative supermartingale, so Ville's inequality bounds the probability of ever aborting by $\alpha$. This holds whatever the stopping time and however $\pi_t$ and $\lambda_t$ were chosen, as long as both depend only on the past. This time-uniform guarantee is what makes adaptive check allocation legitimate.

The allocation rule is the main new ingredient. When $M_t$ is near 1, the evidence is uninformative and $\pi_t$ stays at a low floor. When $M_t$ rises into an ambiguous middle band, $\pi_t$ is increased so that a suspected attack is confirmed or dismissed quickly. When $M_t$ decays, $\pi_t$ falls back. The floor matters for security: it stops an adversary from waiting for a quiet period, and it ties the attacker's expected undetected leakage to $1/\pi_{\min}$.

We will also analyse what the adversary learns if check positions are announced after the fact. Seeing the check rate rise tells the attacker it is suspected. Its best response is then to stop attacking, which lowers leakage rather than raising it. We will state this argument formally and test it with an adaptive simulated attacker.

The analysis will derive the expected leaked message bits before abort, $\mathbb{E}[L]$, as a function of attack probability $p$, noise $\eta$, $\pi_{\min}$ and $\alpha$. It will also derive total probing qubits per delivered message bit. Both will be compared with the fixed-block test sized either for a single design attack strength or for the weakest attack of interest.

Baselines

The naive reference point is no monitoring at all. Every qubit carries data, so efficiency is maximal and detection is zero. It fixes the ceiling on efficiency and the floor on security, and it is run on the same simulated channels so that leakage can be reported relative to it.

The strongest existing method on this exact question is the fixed-block null-hypothesis test of the Tele-Fetch protocol (Keyless Semi-Quantum Point-to-point Communication Protocol with Low Resource Requirements). Its reported numbers are about 99.5% detection success at $p=0.6$ with 15 probing bits (noiseless) and about 60 probing bits in the noisy setting. It is the strongest competitor because it is the only method in the file that already operates in the one-qubit-memory, keyless regime we target.

We will reimplement it in two versions. The first uses a front-loaded block and the second uses uniformly interleaved checks evaluated at the end. Both use the paper's $r$ values and are then re-sized for weaker attacks such as $p=0.1$ and $p=0.2$. All runs share the same channel noise draws, the same $\alpha$, and the same attack schedules as our method. Beating it means that, at matched empirical false-abort rate on honest channels and matched or higher detection probability, our method uses fewer probing qubits averaged over a spread of $p$, and has lower $\mathbb{E}[L]$ at every $p$ tested.

A second baseline is the static signal-to-check ratio of the smart-grid protocol (Privacy-Preserving Bidirectional Data Transmission of Smart Grid Via Semi-Quantum Computation), which reports 94.1% communication efficiency. We will run it at that ratio with a fixed-sample test and compare our method at equal efficiency. At equal efficiency, the question is whether our method detects more attacks and leaks fewer message bits.

A third baseline is off-the-shelf sequential testing: Wald's test with textbook boundaries, the MSPRT R package (A Modified Sequential Probability Ratio Test (MSPRT)), and the boosted clipped test (Improving Wald's (Approximate) Sequential Probability Ratio Test by Avoiding Overshoot). All three run at a fixed check rate with the same $\alpha$. Two ablations isolate our ingredients. The first keeps our martingale but fixes $\pi_t$, which separates the value of anytime-valid monitoring from the value of adaptive allocation. The second resets the martingale on "ambiguous" outcomes, reproducing the restart failure mode of the nested coin-flipping construction, to confirm that never resetting is what preserves the error budget.

Why not just…?

Why not just use a conservatively large fixed check block? This is the real threat. For a single known attack strength on a stable channel, a well-sized fixed block is close to optimal, and the savings from sequential testing on simple alternatives are only 3.4–11.8% in the overshoot study. We do not argue around this. The first experiment measures savings at the design point, and if they are in that range, we report that the case for sequential monitoring rests entirely on unknown $p$ and on leakage before abort.

Those two properties are where a fixed block is structurally weak. A block sized for weak attacks wastes qubits against strong ones, and a block evaluated only at its end cannot bound how much message leaks before the verdict. The experiments sweep $p$ from 0.05 to 0.6 and include delayed attackers to show whether this structural gap is large in practice.

Why not just run Wald's test or MSPRT unchanged? Classic Wald boundaries are approximate, and overshoot changes the realised error rates, as the boosted-test paper documents. Our null is also composite, because the true channel noise is anywhere below $\eta_0$, which standard likelihood-ratio tests do not handle with exact guarantees. Standard tests also assume a fixed sampling design, so changing the check rate mid-stream voids their guarantees. The off-the-shelf baselines are run precisely to measure this. If their empirical false-abort rate under drifting noise and adaptive rates stays at or below $\alpha$ and their efficiency matches ours, the simpler tool wins, and we will say so.

Why not rely on privacy amplification instead of tighter monitoring? Privacy amplification, which hashes a partly leaked key down to a shorter secret one, needs a key to act on. The Tele-Fetch protocol is keyless: the user's message bits travel directly, so bits read by an eavesdropper are simply lost to it. The finite-key decoy literature also shows that loose parameter estimates from small samples force worst-case leakage assumptions and can drive the key rate to zero. Amplification therefore inherits, rather than removes, the small-sample problem studied here. The smart-grid timing results suggest heavy block post-processing is also a poor fit for edge clients. An experiment that tacks privacy amplification onto a key-based variant would be a different protocol, so we treat this as out of scope rather than as a competing baseline.

Proposed Experiments

The main experiment is a Monte Carlo simulation of Tele-Fetch sessions of $10^4$ to $10^5$ rounds. Channel noise $\eta$ is drawn in a band below $\eta_0$, and the measure-and-replay attacker has probability $p \in \{0.05, 0.1, 0.2, 0.4, 0.6\}$. Attacks are either constant, delayed (starting after a chosen round), or adaptive (stopping when check density rises). Each configuration runs at least $10^5$ honest sessions, so a false-abort rate near $\alpha = 10^{-3}$ can be estimated with useful precision. The metrics are empirical false-abort rate, detection probability, probing qubits per delivered message bit, and $\mathbb{E}[L]$.

The first decisive check is calibration. Our method's false-abort rate must stay at or below $\alpha$ under drifting noise and adaptive rates. We expect textbook Wald's test to exceed $\alpha$ in some configurations; if it does not, that weakens the case for our construction.

The second decisive check is the $\mathbb{E}[L]$ curve against the interleaved fixed-block baseline under delayed and weak attacks. If our method does not reduce leakage by a clear margin at $p \le 0.2$, the proposal's central claim fails.

For realistic noise, we will rerun the honest-channel and attack sweeps in NetSquid with the cavity-interconnect extension (Accurate and Scalable Simulation of Cavity-Based Networks in Modular Quantum Architectures). There we use its amplitude-damping and dephasing models calibrated to the reported $T_1 \approx 292\,\mu$s and $T_2 \approx 184\,\mu$s. The question is whether correlated, time-varying noise from decoherence breaks the null model, and whether a slowly updated $\eta_0$ still keeps false aborts bounded.

A small hardware sanity check will run the honest-channel test on a cloud superconducting processor of the kind Lou et al. used. The aim is only to confirm that real device error statistics do not violate the per-round bounded-noise assumption; no efficiency claim will rest on hardware.

As a secondary extension, the same construction applies naturally to multi-link networks. There, per-link martingales can be combined with a multiple-testing correction on e-values to avoid the look-elsewhere inflation seen in the cosmic-ray hotspot analysis, where 5.1σ pre-trial fell to 3.4σ post-trial (Indications of Intermediate-Scale Anisotropy of Cosmic Rays...). We would evaluate this on a five-node NetSquid chain by measuring false network teardowns. It is framed as an extension and is not part of the core claim.