{"id":"b426ebf4-516d-48c3-84e8-748f6f38c419","arxiv_id":"2412.10290","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A heterodyne detection method characterizes the phase de-randomization caused by injection locking and derives a lower bound on the required optical isolation for QKD transmitters.","lead":"This paper measures how much an attacker can reduce the phase randomness of a QKD laser by injecting light into it, using a laboratory setup with two lasers and an interferometer. It reports that the tested telecom laser needs about 140 dB of optical isolation to stay safe against this injection-locking attack.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 140 dB isolation figure is conditional on the chosen master laser emulating Eve's optimal injection-locking source; a better-matched Eve source could lower qmin_rel at the same injected power, invalidating the bound as a security guarantee.","rationale":"The reader's weakest assumption correctly identifies the master laser's ability to emulate Eve's optimal attack as the main load-bearing point. The paper is honest about this limitation, and the proposed method is genuinely useful for black-box evaluation, but the abstract's claim of 'providing a lower bound' on required isolation goes beyond what the data can support: the bound is conditional on the specific ML and setup. This is not an internal inconsistency but a correctness risk in extrapolating the 140 dB figure to arbitrary Eve sources. The secondary model-dependence (wrapped Voigt fit with finite samples) also deserves attention, but the ML-dependence is more fundamental because it limits the very definition of qrel as a bound on Eve's influence. The proposed test, whether experimental or via rate-equation simulation, would settle whether a better-matched Eve source changes the threshold. Since the paper already frames its result carefully in several places, the appropriate verdict remains CONDITIONAL rather than REJECT or ACCEPT, and the reader's verdict does not need to change.","tokens_in":13797,"tokens_out":6976,"duration_ms":69467,"concrete_test":"Repeat the qmin_rel versus injected-power measurement (Fig. 7) with a family of alternative master lasers that include (i) a narrower-linewidth laser (<100 kHz), (ii) a pulsed source synchronized to the gain-switching of the slave laser, and (iii) a source with finely tunable detuning; at each injected power, optimize the polarization state and record the minimum qmin_rel. If any of these configurations yields qmin_rel below the original curve at powers below -90 dBm, then the 140 dB isolation estimate does not hold as a worst-case security bound and the result should be presented as equipment-specific.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim is a lower bound on required isolation (~140 dB), derived from qmin_rel as a function of injected optical power measured with one specific continuous-wave DFB master laser. The authors state in Sec. IIIA that the characterization is only as good as the ML simulates Eve's best possible attack, and Sec. IV explicitly leaves open the problem of determining the maximum degree of injection locking Eve can induce. Because injection locking depends on spectral matching, linewidth, detuning, polarization, and temporal modulation, a real adversary could use a source that couples more strongly into the slave laser than the ML used here. If such a source yields a smaller qmin_rel at injected powers below the -90 dBm threshold, then the 140 dB attenuation estimate would be too low and would not be a valid lower bound. Additionally, the qrel metric is defined relative to the ML phase; any finite linewidth or phase drift of the ML broadens the measured relative phase distribution, inflating qmin_rel and overestimating the degree of phase randomization. No raw data or analysis code are provided, so the figure cannot be independently re-derived or scrutinized.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper reports an experimental method to characterize the phase de-randomization of a gain-switched slave laser under injection locking from a master laser, using a heterodyne detection setup and a qrel metric adapted from Refs. [21,27]. The authors measure phase histograms with and without injection, optimize the injected polarization, and extract qmin_rel as a function of injected optical power. For their DFB lasers, qmin_rel remains close to the no-injection value below approximately -90 dBm; assuming a 100 W fiber damage threshold, they conclude that about 140 dB of isolation is required to protect against the attack. The paper also discusses how the method could be used for black-box testing of complete QKD transmitters.","tokens_in":14028,"tokens_out":8373,"duration_ms":83950,"significance":"The work addresses a real gap: experimental evaluation of injection-locking attacks on QKD phase randomization. The polarization-optimization procedure and the pulse-resolved heterodyne characterization are useful building blocks for certification workflows, and the authors are commendably explicit that the characterization is limited by the master laser's ability to emulate Eve. The bootstrap error bars and the detailed apparatus description are also strengths. However, the central quantitative claim, stated as a lower bound on the required isolation of about 140 dB, is not supported as stated: the experiment can only establish a bound for the particular master laser and slave laser used, and the reported qmin_rel values are minima of a fitted wrapped Voigt model rather than statistically certified lower bounds on the true phase distribution. The qualitative observation that the phase distribution localizes under strong injection and remains flat below about -90 dBm is convincing, but the security-oriented conclusions need re-scoping and additional conservative analysis.","major_comments":[{"comment":"The central claim that 'at least about 140 dB' of isolation is required is logically inverted relative to the experiment. The data show that, for this particular master laser, 140 dB of attenuation keeps the injected power below -90 dBm, at which the measured qmin_rel is close to the no-injection value; this demonstrates that 140 dB suffices for this Eve model, i.e., it gives an upper bound on the required isolation for that model, not a lower bound. Since a better-matched or otherwise stronger Eve source could induce stronger de-randomization at the same injected power, the experiment cannot establish a lower bound on the required isolation for an arbitrary adversary. The authors acknowledge this in Sec. IIIA ('the characterization may only be as good as the ML simulates Eve's best possible attack') and in Sec. IV, but the abstract and Sec. IIIE nevertheless present the result as a lower bound and as 'requires at least about 140 dB.' The claims should be re-scoped to a characterization for the tested ML/DUT pair, or the logical direction should be fixed.","section":"Abstract; Sec. IIIE; Sec. IV"},{"comment":"The reported qmin_rel is the minimum of a wrapped Voigt profile fitted to each histogram, but the definition in Eq. (14) requires a lower bound on the true PDF. Least-squares fitting does not guarantee f_w ≤ f_true; a model that is too narrow or that overestimates the minimum would inflate qmin_rel and therefore understate the required isolation. The bootstrap error bars in Fig. 7 only quantify sampling uncertainty of the fitted parameters, not model misspecification. In addition, the time window used for qmin_rel is selected in App. D based on the goodness of fit, which can bias the result toward time regions where the model happens to perform well. Since the paper proposes the method for certification-style evaluation, the qmin_rel values should be conservative lower bounds, obtained either from a validated model with a safety margin or from distribution-free concentration inequalities, as the authors themselves mention in Sec. IIIC.","section":"Sec. IIIC, Eqs. (12)-(14); App. D; Fig. 7"},{"comment":"The security interpretation of qmin_rel is weakened by the inequality qrel ≥ q stated in Sec. IIIB. A high measured qmin_rel implies only an upper bound on the q-parameter that enters the security proofs of Refs. [21,27]; it does not by itself certify that Alice's phase distribution is close to uniform. In the present experiment, the no-injection value q≈0.963 supports the assumption of good intrinsic randomization, but for black-box testing of an unknown transmitter, a high qrel could in principle be produced by a non-uniform Alice phase combined with master-laser phase fluctuations. The manuscript should state more carefully what qmin_rel can and cannot certify, or provide an additional characterization of the master-laser phase stability on the pulse-to-pulse timescale.","section":"Sec. IIC; Sec. IIIB"}],"minor_comments":[{"comment":"The text says 'the ML will be refereed to as the local oscillator'; 'refereed' should be 'referred'.","section":"Sec. IIIA"},{"comment":"In the sentence citing decoy-state security proofs, the reference marker '28?' appears with a stray question mark; this citation should be cleaned up.","section":"Sec. IIA, references"},{"comment":"Equation (7) writes A_S(n)(τ_n) = (A_S(n)(τ_n))*, which is a tautology as written; the notation for the complex amplitude and its conjugate should be clarified.","section":"Sec. IIIA; Eq. (7)"},{"comment":"The term 'source-agnostic' is used in the abstract and conclusions, but only a DFB slave laser is demonstrated; the method may be adaptable to other sources, but the claim of source-agnosticism should be softened unless a non-DFB device is tested.","section":"Abstract; Sec. IIIE"},{"comment":"No data availability statement or analysis code is provided; for a method whose quantitative output is intended for security certification, making the phase-extraction and fitting code available would substantially improve reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest about its main limitation, and the experimental demonstration appears competently executed. The problem is that the title, abstract, and conclusion overstate the result as a general lower bound on required isolation, while the body correctly notes the dependence on the master laser. The second major issue is that the qmin_rel extraction from a fitted Voigt profile is not conservative in the direction needed for a security bound. Both issues are fixable by re-scoping the claims and adding a conservative statistical treatment, so I see this as a major revision rather than a rejection. The paper fits the scope of the journal as an experimental quantum-cryptography implementation-security study."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is the first source-agnostic experimental recipe for measuring injection-locking-induced phase de-randomization in QKD transmitters and turning it into an isolation requirement. The heterodyne setup, polarization optimization, and qmin_rel vs injected power curve are solid enough to be useful. The authors also state the central caveat themselves: the result is only as good as the master laser's ability to mimic Eve's best attack.\n\nWhat's actually new: Pang et al. and Lovic et al. were attack demonstrations or laser-specific models. Here the method is designed for black-box testing: replace the slave laser with a QKD transmitter, scan polarization, vary the VOA, get a qmin_rel curve. This is most useful for QKD evaluation labs and people designing transmitter isolation. The connection to the qrel metric from security proofs is reasonable, and the claim that qrel >= q is explicitly made, so the authors know the metric is an upper bound on the security-relevant q.\n\nWhere I have doubts:\n- The quantitative curve rests on a wrapped Voigt fit, and the time window is chosen partly by fit quality (App. D). Bootstrap errors (50 resamples) don't include model uncertainty or the ML/SL frequency drift they mention. So the error bars understate the real uncertainty.\n- The 140 dB figure assumes a 100 W fiber damage threshold and a -90 dBm empirical threshold for one DFB pair. If a real Eve uses a source that couples better (narrower linewidth, better spectral overlap, faster modulation), the threshold could move down and the required isolation up. The paper says this too; my issue is that the abstract and conclusion phrase it as 'a lower bound,' which is true as a minimum but too easily read as a sufficiency guarantee.\n- More fundamentally, qrel is defined relative to the ML phase. Any ML phase noise or drift broadens the measured histogram and inflates qmin_rel. The authors acknowledge qrel >= q, but that means the measured qmin_rel is not directly usable as a conservative estimate of q for a security proof. For system comparison it's fine; for certification it needs more thought.\n- No data or code. For a method paper, that's a real gap.\n\nNone of this is fatal. The method is new, clearly explained, and the experimental data support the qualitative behavior. I'd send it to peer review with requests for a model-uncertainty analysis and data release. The 140 dB number should be presented as 'for the tested ML/SL pair' rather than as a general bound.","headline":"Useful black-box method for injection-locking evaluation; treat the 140 dB figure as a data point for one laser pair, not as a security bound.","tokens_in":14567,"tokens_out":5936,"would_cite":true,"duration_ms":662786,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper establishes an experimental method to measure injection-locking-induced phase de-randomization in QKD transmitters, yielding a minimum required isolation of about 140 dB for the tested DFB lasers.","keywords":["quantum key distribution","injection locking","phase randomization","side-channel attack","heterodyne detection","decoy-state BB84","laser isolation","black-box testing"],"falsifier":"Repeat the power scan of Fig. 7 using a master laser with a substantially narrower linewidth or a higher-power, modulated source and check whether $q^{\\min}_{\\rm rel}$ falls measurably below the measured curve at any injected power; a drop, especially near $-90$ dBm, would show that the reported isolation bound is not the worst case for a real attacker.","tokens_in":13595,"feed_emoji":"🔐","tokens_out":10738,"duration_ms":85062,"temperature":0.7,"pith_summary":"Practical quantum key distribution systems rely on the assumption that Alice's laser pulses have uniformly random phases, but an attacker can inject light into the transmitter to lock those phases and break the assumption. This paper develops an experimental procedure to measure how much such an injection-locking attack degrades phase randomization, using a heterodyne detection setup in which a master laser plays the role of Eve and a gain-switched slave laser plays the role of Alice. The measured quantity is a relative $q$-parameter, $q_{\\rm rel}$, a lower bound on the probability density of the phase of each pulse relative to the injected light; the authors report $q^{\\min}_{\\rm rel}$ as a function of injected optical power. For the DFB lasers used here, injected power below about $-90$ dBm leaves phase randomization effectively unchanged, which translates into a required attenuation of roughly $140$ dB assuming Eve can inject up to $100$ W before damaging the fiber. The procedure is source-agnostic and intended for black-box evaluation of QKD transmitters, with the caveat that the result is only as strong as the master laser's ability to simulate Eve's best possible attack.","feed_headline":"QKD needs ~140 dB of isolation to block injection-locking attacks","feed_subtitle":"New heterodyne test measures how Eve's light de-randomizes Alice's laser phases, giving a certifiable isolation target.","key_machinery":"The load-bearing object is the relative $q$-parameter $q_{\\rm rel}$, a lower bound on the probability density of each pulse's phase relative to the master laser's reference phase, defined so that $q_{\\rm rel}=1$ for perfect uniform randomization and $q_{\\rm rel}\\to 0$ for a fully localized phase. It is extracted from a heterodyne detection setup: a continuous-wave DFB master laser injects light into a gain-switched DFB slave laser, a $90^\\circ$ optical hybrid produces in-phase and quadrature signals, and the two-argument arctangent of those signals yields the per-pulse relative phase; a histogram of $N=8000$ phases is fit with a wrapped Voigt profile to obtain $q_{\\rm rel}(\\tau_n)$, and $q^{\\min}_{\\rm rel}=\\min_{\\tau_n} q_{\\rm rel}(\\tau_n)$ is the reported figure of merit. This quantity bridges experiment and security proof: current proofs for imperfect phase randomization require either the full phase density or a parameter like $q$, so measuring $q^{\\min}_{\\rm rel}$ as a function of injected power turns a device vulnerability into a concrete isolation requirement. The polarization of the injected light is optimized either by scanning the Poincaré sphere for the smallest $q^{\\min}_{\\rm rel}$, or by minimizing back-reflected light from the unpowered slave cavity, a shortcut that works for DFB lasers and shortens the optimization from hours to minutes.","core_discovery":"The central claim is that the degree of phase de-randomization caused by injection locking can be measured directly and converted into a minimum optical isolation requirement. The authors define $q_{\\rm rel}$ by requiring the conditional phase density to satisfy $f(\\Delta\\theta^{(n)}|\\Delta\\theta^{(n-1)}\\dots\\Delta\\theta^{(1)}) \\ge q_{\\rm rel}/(2\\pi)$, so a uniform phase distribution gives $q_{\\rm rel}=1$ and a phase-localized distribution gives values approaching zero. A heterodyne receiver with a $90^\\circ$ optical hybrid records $I_0$ and $I_{\\pi/2}$ for each pulse, and $\\arctan 2(I_0,I_{\\pi/2})$ gives the phase relative to the master laser; histograms of $N=8000$ pulses are fit with a wrapped Voigt profile to extract $q_{\\rm rel}(\\tau_n)$, and the minimum over the pulse window defines $q^{\\min}_{\\rm rel}$. Scanning injected power at the optimal polarization, the authors find that $q^{\\min}_{\\rm rel}$ stays near $0.96$ without injection and degrades smoothly as the injected power rises, with no significant influence below about $-90$ dBm; combining this threshold with a $100$ W fiber damage limit yields at least about $140$ dB of attenuation. The method also provides two polarization-optimization routes, a full Poincaré-sphere scan and a faster back-reflection measurement, which agree for the DFB lasers tested. In a black-box evaluation, the measured $q^{\\min}_{\\rm rel}$ is compared with the value assumed by the security proof, and since $q^{\\min}_{\\rm rel}\\ge q$, exceeding the implementer's claimed value certifies the transmitter against this attack.","pith_inferences":["If this procedure is adopted for certification, the quantitative $140$ dB number should not be read as a universal safety margin; it is specific to the tested devices and to the master laser's ability to approximate the worst realistic attacker.","A natural stress test would repeat the scan with a master laser that has a narrower linewidth, higher power, or modulation synchronized to the gain-switching; a drop in $q^{\\min}_{\\rm rel}$ at lower injected powers would show that the reported threshold is not worst-case.","The time-resolved nature of $q_{\\rm rel}(\\tau_n)$ suggests that security proofs treating the phase as constant over the pulse may need to be extended to intra-pulse phase variation, which the paper notes is absent from current analyses.","The same heterodyne methodology could be reused, with a suitable reference source, to characterize other phase-related side channels, such as Trojan-horse attacks that leave a detectable phase imprint in the transmitted pulses."],"forward_implications":["A complete QKD transmitter can be put through the same test in place of the slave laser, and its measured $q^{\\min}_{\\rm rel}$ can be checked against the phase-randomization value assumed in the security proof.","For the DFB lasers measured in this paper, an isolation chain providing less than about $140$ dB of attenuation leaves the transmitter vulnerable to a $100$ W injection-locking attacker.","Knowing the smallest $q_{\\rm rel}$ Eve can induce covers even pulse-to-pulse variation of the attack, because the bound in Eq. (4) holds for every conditional phase density.","The two polarization-optimization methods give equivalent results for DFB lasers, allowing fast characterization, while the Poincaré-sphere scan remains applicable to more complex transmitters with polarization-dependent losses.","A faster oscilloscope or a different histogram model would extend the usable time window and tighten the $q^{\\min}_{\\rm rel}$ estimate, which the authors identify as an experimental limitation."],"supporting_citations":[{"why":"Defines the q-parameter for imperfect phase randomization that this paper adapts into the relative $q_{\\rm rel}$ metric.","marker":"[27]"},{"why":"Provides the generalized decoy-state security analysis requiring a partial characterization of the phase distribution.","marker":"[21]"},{"why":"Demonstrates hacking QKD via injection locking, the attack class the method evaluates.","marker":"[23]"},{"why":"Studies the laser-seeding attack's effect on intensity, a related side channel the method complements.","marker":"[24]"},{"why":"Models injection locking's effect on phase de-randomization, which the present method extends beyond laser-specific modeling.","marker":"[26]"},{"why":"Supplies the 100 W fiber damage threshold used to convert the measured power threshold into the 140 dB isolation requirement.","marker":"[12]"},{"why":"Gives the decoy-state BB84 security proof that assumes uniform phase randomization, the assumption under attack.","marker":"[6]"},{"why":"Provides a finite-key decoy-state analysis whose phase-randomized source is the protocol example for the characterization.","marker":"[8]"}],"fun_headline_variants":["QKD needs 140 dB isolation to block injection-locking","Heterodyne test reveals QKD isolation target vs injection locking","Measure phase de-randomization to certify QKD transmitters","Injection-locking attack: how much isolation does QKD need?"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method stands on the assumption that the laboratory master laser can imitate Eve's best possible injection-locking source; if a real attacker's light couples to Alice's laser more effectively than the master laser does, the true de-randomization is stronger and the required isolation is higher than measured.","fun_headline_variants_meta":{"raw":{"variants":["QKD needs 140 dB isolation to block injection-locking","Heterodyne test reveals QKD isolation target vs injection locking","Measure phase de-randomization to certify QKD transmitters","Injection-locking attack: how much isolation does QKD need?"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1479,"prompt_tokens":1029,"completion_tokens":450,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":645,"completion_tokens_details":{"reasoning_tokens":378}},"tokens_in":645,"tokens_out":450,"duration_ms":4737,"temperature":1.0,"reasoning_tokens":378,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:58:56.625926+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the power scan of Fig. 7 using a master laser with a substantially narrower linewidth or a higher-power, modulated source and check whether $q^{\\min}_{\\rm rel}$ falls measurably below the measured curve at any injected power; a drop, especially near $-90$ dBm, would show that the reported isolation bound is not the worst case for a real attacker.","supporting_citations":[{"cited_title":"Security of quantum key distribution with imperfect phase randomisation","cited_arxiv_id":"2210.08183","evidence_quote":"Defines the q-parameter for imperfect phase randomization that this paper adapts into the relative $q_{\\rm rel}$ metric."},{"cited_title":"Imperfect Phase-Randomisation and Generalised Decoy-State Quantum Key Distribution","cited_arxiv_id":"2304.09401","evidence_quote":"Provides the generalized decoy-state security analysis requiring a partial characterization of the phase distribution."},{"cited_title":"Pang, A.-L","cited_arxiv_id":null,"evidence_quote":"Demonstrates hacking QKD via injection locking, the attack class the method evaluates."},{"cited_title":"Huang, Á","cited_arxiv_id":null,"evidence_quote":"Studies the laser-seeding attack's effect on intensity, a related side channel the method complements."},{"cited_title":"Lovic, D.G","cited_arxiv_id":null,"evidence_quote":"Models injection locking's effect on phase de-randomization, which the present method extends beyond laser-specific modeling."},{"cited_title":"Makarov, A","cited_arxiv_id":null,"evidence_quote":"Supplies the 100 W fiber damage threshold used to convert the measured power threshold into the 140 dB isolation requirement."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the decoy-state BB84 security proof that assumes uniform phase randomization, the assumption under attack."},{"cited_title":"Rusca, A","cited_arxiv_id":null,"evidence_quote":"Provides a finite-key decoy-state analysis whose phase-randomized source is the protocol example for the characterization."}],"review_version":1}