{"id":"b328f13c-472b-48e7-b21d-c6ef743dfd78","arxiv_id":"2508.10613","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A new attack-radius metric and routing algorithm for QKD networks reduces the maximum number of key requests disrupted by a single fiber jamming attack.","lead":"This paper introduces a new metric, maxNAR, measuring the worst-case impact of a single fiber jamming attack on quantum key distribution (QKD) networks, and proposes an ILP plus a Tabu-search heuristic to minimize it. The heuristic matches the ILP on a small testbed and cuts attack impact versus a simple baseline on a larger network.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'ILP' formulation is not an ILP: Eq. (3) uses logical OR and Eq. (15) uses logical AND with no linearization, so the reported exact optimum for PoliQi is not reproducible from the stated model.","rationale":"The reader's verdict CONDITIONAL is appropriate. I identified the nonlinearity of the ILP as the most load-bearing concern because the paper's exactness claim and the heuristic's optimality benchmark depend on it. The reader's weakest_assumption focuses on the attack model; while that is a conceptual concern, the ILP flaw is a formal, checkable defect that currently invalidates the 'optimal' claim. The concrete test can resolve whether the model is merely missing a linearization or fundamentally misformulated. If the linearized model reproduces the reported optimum, the paper becomes conditionally acceptable with a required revision; if not, the evaluation section's central comparison is unsupported.","tokens_in":8598,"tokens_out":6266,"duration_ms":68565,"concrete_test":"Implement the model exactly as written in Gurobi/CPLEX, including the OR in Eq. (3) and the AND in Eq. (15). If the solver rejects the constraints as nonlinear (or requires an indicator/bilinear setting), the 'ILP' claim is falsified. Then apply the standard linearization: for OR, f ≤ q+p, f ≥ q, f ≥ p; for AND, C ≤ A, C ≤ B, C ≥ A+B-1. Re-solve the PoliQi 5-node ring instance with the same parameters (7 requests, 10 modules/node, 10 kb/s per request). If the resulting optimal maxNAR differs from 2, or if the solution time is not in the reported range (>10h with the original solver), the reported match between heuristic and ILP is not reproducible. If the linearized model gives maxNAR=2, the issue is a missing derivation rather than a wrong result, and the paper should include the linearization.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim rests on an Integer Linear Programming (ILP) model that supposedly defines RWA-MAR and provides the optimal maxNAR benchmark (Section II-D). However, constraints (3) and (15) are not linear as written. Eq. (3) states f_{e,w}^{p,t} = q_{e,w}^{p,t} ∨ p_{e,w}^{p,t}, a logical OR of two binary variables. Eq. (15) states C_{ϕ}^{d,t} ≥ (xx_{e'}^{d,t} · h_{ϕ,e'}) ∧ B_{ϕ}^{t}, a logical AND between a binary expression and a binary variable. Neither is a linear inequality, and the paper provides no linearization (e.g., standard big-M substitutions). Therefore the model is not an ILP; it is a mixed-integer nonlinear model or relies on undocumented solver-level logical constraint handling. Consequently, the claim that the Min-maxNAR heuristic 'achieves the same (optimal) maxNAR of the ILP' (Section III) is unsupported: we do not know what objective the solver actually optimized, nor whether the reported >10-hour run corresponds to the stated constraints. This is not cosmetic; the formal guarantee that the heuristic is within a known factor of optimum—or exactly optimal on the small instance—is the main evidence for the heuristic's correctness. The attack-propagation model (Section II-A) may also deserve scrutiny, but the immediate load-bearing flaw is the invalid ILP.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a new metric, maxNAR (Maximum Number of Affected Requests), to quantify the worst-case impact of a single physical-layer jamming attack in QKD networks. It formulates the Routing and Wavelength Assignment with Minimal Attack Radius (RWA-MAR) problem as an ILP, proposes a Tabu-search heuristic (Min-maxNAR), and evaluates the approach on a 5-node ring (PoliQi) and the 14-node NSF topology under three architectures: Optical Bypass (OB), Trusted Relay (TR), and their combination (OB-TR). The paper reports that the heuristic achieves the same optimal maxNAR as the ILP on the small ring, reduces maxNAR by about 27% over a depth-first shortest-path baseline on NSF, and analyzes the effect of a tunable priority parameter α and QKP caching across time slots.","tokens_in":8987,"tokens_out":5045,"duration_ms":60785,"significance":"If the technical issues are resolved, the paper addresses a relevant and under-explored problem: resilience of QKD networks to physical-layer attacks. The proposed maxNAR metric, the explicit modeling of OB/TR differences, and the use of QKP caching are useful contributions. The paper also builds on realistic QKD key-rate models from prior work. The main value would be in providing a network-design metric and heuristic algorithms for attack-aware routing. However, in its current form, the ILP formulation and the statistical support for the headline claims need correction before the results can be trusted.","major_comments":[{"comment":"Eq. (3) defines f_{e,w}^{p,t} = q_{e,w}^{p,t} ∨ p_{e,w}^{p,t}, and Eq. (15) defines C_{φ}^{d,t} ≥ (xx_{e'}^{d,t} · h_{φ,e'}) ∧ B_{φ}^{t}. Neither is a linear inequality, and no linearization (e.g., big-M) is provided. As written, the model is not an ILP. This is load-bearing because the claim that the Min-maxNAR heuristic 'achieves the same (optimal) maxNAR of the ILP' in Section III is the main evidence for correctness. Please provide the linearized constraints explicitly, or state that a solver's indicator/logical-constraint handling is used and justify that the model remains a valid ILP. Without this, the optimality result is not reproducible from the stated model.","section":"Section II-D, Eqs. (3) and (15)"},{"comment":"The definition of maxNAR in Section II-A says it is 'the maximum number of requests any one request is link-sharing with,' which is a pairwise property between requests. However, Eq. (16) defines maxNAR^t as a lower bound on the number of requests affected by an attack on a route/link, maximized over e and φ. These are different objects. The text and the equations must be aligned; otherwise the metric being minimized is ambiguous. This is central because all security claims are stated in terms of maxNAR.","section":"Section II-A vs. Eq. (16)"},{"comment":"The heuristic is randomized (Tabu search with random lightpath selection), but the results in Figs. 4-6 show no error bars, confidence intervals, or statistical tests. The conclusion that the heuristic 'significantly outperforms the baseline' is not supported by a single run of a stochastic algorithm. Please report multiple independent runs (or seed sensitivity) and provide statistical comparisons. Also, for the PoliQi case, give the ILP solver's optimality gap/time-limit information so the exact-optimality claim can be assessed.","section":"Section III, Figs. 4-6"},{"comment":"The attack-propagation model assumes that a high-power jamming signal on one link propagates through OB to downstream links, and that QKP-cached keys are completely immune. These assumptions are load-bearing for the validity of maxNAR as a security metric, but the paper does not validate them experimentally or with a sensitivity analysis. If the physical model is inaccurate, the optimized routing may not improve true resilience. Please add a discussion of the physical basis, cite relevant experimental work, and/or include sensitivity tests of the propagation assumptions.","section":"Section I and Section II-A"}],"minor_comments":[{"comment":"The quantifier in Eq. (10) reads '∀p∈E_a', but p should be a node pair, not an auxiliary link; it should likely be '∀p∈P'.","section":"Section II-D, Eq. (10)"},{"comment":"The variable z_{p,w}^t is described as 'equals t' if the QKD path uses channel w; this should be 'equals 1'.","section":"Table II"},{"comment":"There is a typo 'Fig. ref fig:14-2(a)' in the α-discussion; it should be a proper figure reference.","section":"Section III"},{"comment":"Several spacing and capitalization inconsistencies appear: 'RW A-MAR' in the abstract, 'maXNAR' in the contribution list. These should be corrected.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant problem in QKD networking and has a plausible heuristic, but the ILP issue and the lack of statistical validation are significant. I do not see evidence of intentional misrepresentation; the problems appear fixable with a careful revision. The authors should also consider releasing the evaluation code and data to support reproducibility, given the randomized heuristic and the long ILP runtimes. The topic fits a networking/optical communications venue rather than a pure quantum-information one."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper introduces a new resilience metric for QKD networks—maxNAR, the worst-case number of key requests disrupted by a single physical-layer jamming attack—and formulates the RWA-MAR problem to minimize it under OB/TR architectures and QKP caching. That combination is genuinely new; prior work on attack-radius in classical optical networks used LAR/NAR, and QKD routing has not considered jamming propagation. The authors also provide a tabu-search heuristic with a tunable priority parameter α, and show on a 5-node ring that it matches the ILP optimum in seconds versus over ten hours for the ILP. On NSF topology, they report ~27% maxNAR reduction over a shortest-path baseline. The treatment of QKP storage over timeslots is sensible, and the paper is clearly written.\n\nThe soft spots are real. Most importantly, the \"ILP\" is not an ILP as written. Constraint (3) uses a logical OR between two binaries (f = q ∨ p) and constraint (15) uses a logical AND between a product and a binary (C ≥ (xx·h) ∧ B). Neither is linear, and no linearization is given. That means the claimed optimal benchmark is not reproducible from the stated model; we don't know what the solver actually solved. This is fixable—standard big-M linearizations exist for both—but the paper needs to show them. Second, the maxNAR definition in Section II-A (\"maximum number of requests any one request is link-sharing with\") doesn't match Eq. (16), which defines it as the max over routes of the number of affected requests. I think the text is just sloppy, not the math. Third, the baseline is a depth-first shortest-path algorithm, which is weak, and there are no error bars or statistical tests; the 27% figure is a single run comparison. The heuristic is evaluated on the metric it optimizes, which is fine for an optimization paper, but the average-NAR comparison is more informative and shows smaller gaps.\n\nThe attack propagation model—that a jamming signal on one link propagates downstream only for OB paths—is a simplification, but the authors state it clearly and it's a reasonable first-order model. The α parameter is a design knob, not fitted, so no circularity issue.\n\nWho is this for? People working on QKD network design or resilience, and to a lesser extent classical optical network attack-aware routing. It deserves a serious referee: the idea is useful, the flaws are addressable, and even with the ILP problem, the heuristic and the metric are worth publishing after revision.\n\nRecommendation: send it to peer review, but require the authors to linearize constraints (3) and (15), fix the maxNAR wording, and add more robust baseline comparisons or at least multiple runs.\n\nBest.","headline":"Useful new metric and heuristic for QKD attack resilience, but the ILP as written is not linear and the optimality claim needs a fix.","tokens_in":9462,"tokens_out":2743,"would_cite":true,"duration_ms":27747,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P94","90C10","90C59"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper introduces maxNAR, a metric for the worst-case impact of a single physical-layer jamming attack on a QKD network, and a routing heuristic that minimizes it.","keywords":["quantum key distribution","network resilience","physical-layer attack","maxNAR","routing and wavelength assignment","trusted relay","optical bypass","tabu search"],"falsifier":"A controlled experiment on a deployed QKD testbed: inject a high-power jamming signal into one link of a path using optical bypass and measure which key requests actually fail; if requests beyond the first bypassed node survive because the jamming signal drops below the damage threshold, maxNAR is not a faithful worst-case count. Alternatively, compute maxNAR on a topology where OB paths share only a single upstream link; if the model predicts large downstream disruption but observed outages are confined to the directly jammed link, the propagation assumption is falsified.","tokens_in":8552,"feed_emoji":"🔐","tokens_out":6968,"duration_ms":60805,"temperature":0.7,"pith_summary":"Quantum key distribution (QKD) networks can exchange keys with information-theoretic security, but a single high-power jamming attack on one fiber can disrupt many key requests at once, especially when optical bypass lets interference propagate downstream. This paper defines maxNAR, the Maximum Number of Affected Requests under the worst single-link attack, and uses it to formulate a new problem: routing, wavelength, and key-rate assignment that minimizes maxNAR over time slots. The authors solve the problem exactly with an integer linear program for small networks and with a scalable tabu-search heuristic for larger ones. On a five-node ring, the heuristic matches the ILP optimum (maxNAR = 2 for trusted-relay and hybrid architectures) in seconds rather than hours; on a 14-node topology it reduces maxNAR by about 27% versus a shortest-path baseline. A tunable parameter lets operators trade off between optical bypass (low module cost, larger attack radius) and trusted relays (more modules, smaller radius).","feed_headline":"27% smaller worst-case attack radius for QKD routing","feed_subtitle":"New maxNAR metric and tabu-search heuristic minimize how many key requests one jammed fiber can disrupt.","key_machinery":"The central object is the maxNAR metric together with the attack-propagation model it encodes: a jamming attack on a physical link disrupts every request sharing that link, and for optical-bypass (OB) paths the attack signal propagates through the bypass to all downstream links in the same direction, while trusted-relay (TR) paths terminate the propagation and QKP-cached keys are unaffected. The RWA-MAR problem turns this metric into an optimization: minimize the sum over timeslots of maxNAR, subject to flow conservation, module counts, channel capacity, and key-rate/QKP storage constraints. The Min-maxNAR heuristic is a tabu-search algorithm that starts from a shortest-path solution (with a","core_discovery":"The central claim is that the worst-case security exposure of a QKD network to a physical-layer attack can be captured by a single number, maxNAR, and that this number can be substantially reduced by routing choices that exploit key caching, trusted relays, and optical bypass. The attack-propagation model behind maxNAR treats a jammed link as disrupting every request traversing it; for optical-bypass paths the interference propagates to all downstream links in the same direction, while trusted-relay paths stop at the relay and QKP-cached keys are completely immune. The paper supports this claim by formulating the RWA-MAR problem as an ILP, showing the heuristic matches the ILP where the ILP","pith_inferences":["If maxNAR is adopted as a planning metric, the same attack-propagation model could be applied to classical encrypted optical networks that use regenerators or optical bypass, not just QKD.","The paper's attack model assumes a single link is jammed; a natural extension is a worst-case over multiple simultaneous jammers, which would turn the min-max problem into a min-max over attack sets and likely require different combinatorial tools.","The heuristic's performance depends on the assumption that the jamming signal propagates deterministically through optical bypass; a field measurement of jamming attenuation over multiple bypassed nodes would calibrate whether maxNAR should be weighted by distance or signal power rather than counting requests uniformly.","QKP capacity is sized against AES-256 key rotation; if QKD keys were used for one-time-pad encryption of large traffic, the QKP buffer would be a scarcer resource and the routing trade-offs in this paper would shift."],"forward_implications":["Network operators could quantify a QKD deployment's security exposure by a single worst-case number, maxNAR, and compare architectures (OB, TR, hybrid) on that basis.","Key caching in QKPs can be exploited by routing to sharply reduce maxNAR over timeslots; the paper shows maxNAR dropping from 37 to 32 in the first timeslot and stabilizing near 1–2 once pools are stocked.","A tunable OB/TR preference parameter spans the resource–security trade-off: higher TR preference cuts maxNAR substantially (23% between α=80 and α=0 in the hybrid architecture) at the cost of more QKD modules.","The tabu-search heuristic is fast enough for operational use—about five seconds versus over ten hours for the ILP on the small testbed—and scales to network sizes where the ILP is infeasible.","The approach adapts the classical lightpath attack radius idea from all-optical WDM networks to QKD-specific features like key pools, trusted relays, and optical bypass."],"supporting_citations":[{"why":"Supplies the achievable key-rate model and the QKD module accounting for OB and TR paths that the RWA-MAR constraints are built on.","marker":"[5]"},{"why":"Defines physical-layer attack models and the lightpath attack radius metric that maxNAR extends to QKD-specific features.","marker":"[6]"},{"why":"Introduces the auxiliary-graph representation of quantum-node logical adjacencies that the paper uses to model optical bypass.","marker":"[14]"},{"why":"The authors' prior work, used for QKP capacity estimation and the staged key-caching dynamics.","marker":"[15]"},{"why":"The tabu-search tutorial on which the Min-maxNAR heuristic is based.","marker":"[17]"},{"why":"Comparative study justifying tabu search for min-max optimization problems like RWA-MAR.","marker":"[18]"}],"fun_headline_variants":["Shrink QKD attack impact with smarter routing","Minimize worst-case damage from QKD fiber jams","New metric and heuristic cut QKD attack radius","QKD routing strategy reduces worst-case attack impact","Cut network attack risk with maxNAR-aware routing"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The paper treats a high-power jamming attack on one fiber as disrupting every request that shares the fiber, with the disruption propagating unchanged through every downstream link when the path uses optical bypass; if real jamming signals attenuate, spread differently, or fail to propagate through certain hardware, maxNAR may overstate or understate the true worst case.","fun_headline_variants_meta":{"raw":{"variants":["Shrink QKD attack impact with smarter routing","Minimize worst-case damage from QKD fiber jams","New metric and heuristic cut QKD attack radius","QKD routing strategy reduces worst-case attack impact","Cut network attack risk with maxNAR-aware routing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000234,"raw_usage":{"total_tokens":1326,"prompt_tokens":732,"completion_tokens":594,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":476,"completion_tokens_details":{"reasoning_tokens":520}},"tokens_in":476,"tokens_out":594,"duration_ms":6540,"temperature":1.0,"reasoning_tokens":520,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T20:18:29.614259+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled experiment on a deployed QKD testbed: inject a high-power jamming signal into one link of a path using optical bypass and measure which key requests actually fail; if requests beyond the first bypassed node survive because the jamming signal drops below the damage threshold, maxNAR is not a faithful worst-case count. Alternatively, compute maxNAR on a topology where OB paths share only a single upstream link; if the model predicts large downstream disruption but observed outages are confined to the directly jammed link, the propagation assumption is falsified.","supporting_citations":[{"cited_title":"Routing, channel, key-rate, and time-slot assignment for qkd in optical networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the achievable key-rate model and the QKD module accounting for OB and TR paths that the RWA-MAR constraints are built on."},{"cited_title":"Physical-layer attacks in all-optical wdm networks,","cited_arxiv_id":null,"evidence_quote":"Defines physical-layer attack models and the lightpath attack radius metric that maxNAR extends to QKD-specific features."},{"cited_title":"Auxiliary graph based routing, wavelength, and time- slot assignment...,","cited_arxiv_id":null,"evidence_quote":"Introduces the auxiliary-graph representation of quantum-node logical adjacencies that the paper uses to model optical bypass."},{"cited_title":"Drl-based progressive recovery for quantum-key- distribution networks,","cited_arxiv_id":null,"evidence_quote":"The authors' prior work, used for QKP capacity estimation and the staged key-caching dynamics."},{"cited_title":"Tabu search: A tutorial,","cited_arxiv_id":null,"evidence_quote":"The tabu-search tutorial on which the Min-maxNAR heuristic is based."},{"cited_title":"Evolutionary algorithms, simulated annealing and tabu search: a comparative study,","cited_arxiv_id":null,"evidence_quote":"Comparative study justifying tabu search for min-max optimization problems like RWA-MAR."}],"review_version":1}