{"id":"8b117a26-5281-4f61-ba61-f31056623c75","arxiv_id":"2606.07804","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A lightweight cost-sensitive DQN with polar encoding and latent cosine reward shaping reaches ~98%/91% accuracy on CICIDS2017/UNSW-NB15 at 32–46 μs per sample.","lead":"The paper presents QIRL, a small deep Q-network with polar feature encoding, adaptive action angles, and cosine-similarity reward shaping for network intrusion detection aimed at vehicles. It reports high accuracy on standard IDS datasets with tens-of-microseconds CPU inference, far faster than tree ensembles.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Ablations do not establish that quantum-inspired modules are necessary for the dual accuracy-latency result; simpler DQN variants match or beat Full QIRL.","rationale":"The reader correctly isolates the weakest link: the dual-constraint empirical win is real for a carefully regularized tiny DQN, but the paper’s marketing of quantum-inspired encoding/exploration/interference as the enabling mechanism is not backed by the ablations or by consistent main-vs-ablation numbers. That is precisely the load-bearing concern; everything else (dataset framing mismatch with the arXiv abstract, missing code, general NIDS rather than true V2X traces) is secondary. The concrete multi-seed ablation re-run would settle the issue cleanly. No stronger internal contradiction appears once the quantum-necessity claim is set aside, so the CONDITIONAL verdict stands unchanged: accept the latency-aware small-DQN systems result only after the quantum contribution is either dropped or re-proven.","tokens_in":21035,"tokens_out":638,"duration_ms":23457,"concrete_test":"Re-train Baseline, Cost-Sensitive Only, Interference Only and Full QIRL for 5 independent seeds under the exact main-text protocol (train-only SMOTE, validation TSS threshold selection, single-sample CPU latency). Report mean±std Acc/TSS/F1/latency on both datasets; if Full QIRL does not significantly outperform Cost-Sensitive Only and Baseline on TSS (paired t-test p<0.05) while preserving sub-50 μs, the quantum-necessity claim is falsified.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that QIRL (amplitude-phase encoding + rotation-gate exploration + cosine-interference reward with Λ=0.1) simultaneously delivers competitive TSS/AUC and sub-50 μs inference (Eq. 1, Tables 1–3). Yet the ablations in §7.4.4 (Tables 4–5) show this is not demonstrated. On CICIDS2017, Cost-Sensitive Only already reaches Acc 0.9793 / TSS 0.9558 while Full QIRL is 0.9762 / 0.9586; main-text Acc 97.89 % / TSS 0.9443 also fails to match the ablation Full numbers. On UNSW-NB15, Baseline DQN (no quantum, symmetric reward) records Acc 0.9434 / TSS 0.8597, strictly better than Full QIRL’s 0.9388 / 0.8094, while main-text Acc collapses to 91.04 %. Latencies remain ~30–50 μs across every variant, confirming the speedup is simply the tiny 64-wide MLP, not the quantum machinery. Because the quantum components add no inference cost and do not reliably improve the constrained objective, the claim that they are load-bearing for the dual result is unsupported by the paper’s own evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript proposes QIRL, a lightweight Deep Q-Network with three quantum-inspired components (amplitude-phase feature encoding, rotation-gate exploration, and cosine-similarity interference on consecutive latent states) plus cost-sensitive rewards and training-only SMOTE, framed as a cost-sensitive MDP for IoV intrusion detection. The central claim is that this architecture simultaneously satisfies competitive detection robustness and sub-millisecond inference (Eq. 1): 97.89% accuracy / TSS 0.9443 / AUC 0.9945 at 32.5 μs on CICIDS2017 and 91.04% / TSS 0.8244 / AUC 0.9713 at 45.7 μs on UNSW-NB15, with large speedups over ensemble baselines. Evaluation uses a careful train/val/test protocol, single-sample CPU latency measurement, ablations, comparative tables, and SHAP/PCA/t-SNE explainability.","tokens_in":21296,"tokens_out":1553,"duration_ms":22909,"significance":"If the dual accuracy–latency result is solid under realistic IoV constraints, the work is practically significant: sub-50 μs CPU inference with competitive TSS/AUC would be deployable on edge nodes where ensemble IDS methods exceed the ~1 ms safety window. Strengths include an explicit dual objective (Eq. 1), training-only SMOTE to avoid leakage, single-sample latency protocol, asymptotic complexity analysis showing QIM adds no inference cost, and a structured XAI section. The quantum-inspired modules are presented as the distinctive technical contribution; their necessity for the dual claim is therefore load-bearing for the paper’s novelty relative to a small cost-sensitive DQN.","major_comments":[{"comment":"§7.4.4, Tables 4–5: The ablations do not establish that the quantum-inspired modules are necessary for the dual objective in Eq. 1. On CICIDS2017, Cost-Sensitive Only already reaches Acc 0.9793 / TSS 0.9558 while Full QIRL is 0.9762 / 0.9586; on UNSW-NB15, Baseline DQN (no quantum modules, symmetric reward) records Acc 0.9434 / TSS 0.8597, strictly better than Full QIRL’s 0.9388 / 0.8094. Latencies remain ~30–50 μs across all variants, so the reported 51–68× speedups are explained by the tiny 64-wide MLP, not by encoding/rotation/interference. The manuscript should either (i) show statistically that Full QIRL improves the constrained objective (TSS subject to τ < 1 ms) over the best simpler variant under matched seeds/thresholds, or (ii) reframe the contribution as a lightweight cost-sensitive DQN and demote the quantum machinery to optional regularisation.","section":null},{"comment":"Table 3 vs Tables 4–5: Main-text CICIDS2017 results (Acc 97.89%, TSS 0.9443) do not match the Full QIRL ablation row (Acc 0.9762, TSS 0.9586); UNSW main-text Acc 91.04% / TSS 0.8244 likewise diverges from ablation Full (0.9388 / 0.8094) and from Baseline (0.9434 / 0.8597). Different episode counts (20 vs 10), thresholds, or SMOTE settings appear to be mixed without reconciliation. Please report a single protocol (episodes, τ*, seeds, SMOTE) for main and ablation tables, with means±std over multiple runs, so the dual-claim numbers are reproducible and comparable.","section":null},{"comment":"§1–§5 and §7.1: The paper is framed as V2X/IoV cyber defense with multi-stage temporal attacks and edge safety constraints, yet evaluation uses only CICIDS2017 (DDoS-focused flows) and UNSW-NB15 under binary (or effectively binary) classification of general network traffic. The MDP state is an encoded packet-flow vector and the action space is binary detect/benign (§5.2); there is no mobility, RSU load, channel busy ratio, or mitigation action beyond classification. Either add a vehicular/CAN or mobility-aware evaluation that exercises the claimed IoV threat model (§4), or narrow the claims to low-latency NIDS with IoV as motivation rather than demonstrated deployment setting.","section":null},{"comment":"§5.3, Eqs. (11)–(12) and §7.3.3: The Quantum Interference Module adds Λ·cos_sim(ψ_t, ψ_{t+1}) to the Bellman target with fixed Λ=0.1 and no extra parameters. Given that Interference Only does not dominate Baseline on either dataset (Tables 4–5), the paper needs a clearer causal account—e.g., learning curves, TD-error variance, or sensitivity over Λ—showing when interference stabilises training versus when it is neutral. Without that, the quantum analogy remains branding rather than a demonstrated mechanism for the dual constraint.","section":null}],"minor_comments":[{"comment":"Figure 6 caption notes AUC 0.9712 vs main-text 0.9713 and mentions an “initial model” ROC; align figure, caption, and Table 3 to the final thresholded model.","section":null},{"comment":"§5.2 reward weights: the text states both w_fn=5.0, w_fp=1.0 and w_fn=1.5, w_fp=1.0 in one sentence; clarify dataset-specific assignment as later done in §7.3.3.","section":null},{"comment":"Table 6: several baselines lack latency; the independent re-measurements (§, ‡) are useful—state hardware and code parity more explicitly so readers can reproduce the 2248 μs ensemble figure.","section":null},{"comment":"Notation: QIRL vs “Deep QIRL” in figures; ψ for both latent embeddings and interference similarity; standardise.","section":null},{"comment":"Minor language/typos throughout (e.g., “not but fixed”, “TrueSkillStatisticsof”); a copy-edit pass would help.","section":null},{"comment":"§6 complexity: d∈{40} after feature selection is stated, but §7.3.2 does not detail which 40 features or the selection method; add a short note for reproducibility.","section":null}],"recommendation":"major_revision","confidential_remarks":"The arXiv-facing abstract/title block supplied with the submission (CA-QIRL, CAN-MIRGU 80.31%, CBR/delay reductions, ROAD/VeReMi) does not match the full manuscript body (QIRL on CICIDS2017/UNSW-NB15 only). Please confirm the authors submitted a single consistent version; a metadata/body mismatch would confuse readers and indexing. Scope-wise the work is a solid systems/ML security paper if claims are recalibrated; the quantum framing is the main overclaim relative to the ablations."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The thing worth knowing is that this is a careful small-DQN systems paper with a real latency story, not a new quantum defense principle. A 64-wide feedforward Q-network with train-only SMOTE, asymmetric FN/FP costs, and single-sample CPU timing hits ~32–46 μs/sample and stays near ensemble accuracy on CICIDS2017 and UNSW-NB15. That dual accuracy–latency package is the actual contribution, and the measurement protocol is better than most IDS papers that only report accuracy.\n\nWhat they do well: leakage-aware preprocessing, validation-set threshold selection, explicit cost-sensitive rewards, ablations, SHAP plus PCA/t-SNE, and a complexity argument that matches the measured speed. The comparison to re-timed ensembles (~2 ms) is fair under the same single-sample protocol. For edge IDS people who care about sub-millisecond response, that empirical package is useful.\n\nSoft spots, in proportion. The ablations undercut the branding: on UNSW, baseline DQN beats Full QIRL on TSS; on CICIDS, cost-sensitive-only is already competitive, and main-text numbers do not line up cleanly with the ablation Full row. Latencies are ~30–50 μs across every variant, so the speedup is the tiny MLP, not polar encoding, rotation angles, or cosine interference. Calling those modules “quantum-inspired” is packaging of known reparameterizations and a latent cosine term; they are not shown to be necessary for Eq. 1. Second, the arXiv abstract (CA-QIRL, CAN-MIRGU, CBR/delay, ROAD/VeReMi) does not match the manuscript body (QIRL on CICIDS/UNSW only). That is a real integrity/framing problem, not a nit. Third, the IoV safety story is mostly general NIDS data; no code release.\n\nMath and citations look ordinary for the area—no circularity beyond normal hyperparameter tuning. This is for systems/IDS readers who want a latency-aware recipe, not for people hunting a new RL or quantum result. I would send it to peer review with a clear demand to drop or re-prove the quantum necessity claim, reconcile abstract vs body, and either add true V2X evidence or reframe the scope. Engage the systems numbers; do not buy the quantum story on the present ablations.","headline":"Solid latency-first small-DQN systems result on NIDS data; the quantum-inspired modules are not shown to be load-bearing, and the abstract/body framing do not match.","tokens_in":22005,"tokens_out":601,"would_cite":false,"duration_ms":9552,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A lightweight quantum-inspired deep Q-network can detect IoV intrusions at competitive accuracy while staying under 50 microseconds per sample—fast enough for vehicle safety windows.","keywords":["Internet of Vehicles","intrusion detection","reinforcement learning","quantum-inspired computing","Deep Q-Network","edge latency","smart cities","cost-sensitive learning"],"falsifier":"Re-run the same train-only SMOTE, cost weights, and tiny three-layer DQN with those three quantum modules fully disabled; if accuracy, TSS, and sub-50 μs latency remain statistically unchanged on the same CICIDS2017 and UNSW-NB15 splits, the claim that the quantum-inspired machinery is necessary for the dual objective fails.","tokens_in":21812,"feed_emoji":"🚗","tokens_out":961,"duration_ms":17581,"temperature":0.7,"pith_summary":"Connected vehicles and smart-city traffic systems need intrusion detection that is both accurate and fast enough to act inside sub-millisecond safety budgets. This paper argues that existing ensemble detectors hit high accuracy but take milliseconds per flow, while smaller models struggle to adapt to multi-stage, imbalanced attacks. It introduces QIRL, a compact deep Q-network that treats detection as a cost-sensitive sequential decision problem and adds three quantum-inspired pieces: polar amplitude-phase encoding of features, rotation-gate action exploration, and a cosine-similarity interference term that stabilizes learning. On CICIDS2017 and UNSW-NB15, under train-only SMOTE and validation-chosen thresholds, the agent reports about 97.9% and 91% accuracy with True Skill Statistics well above chance, and inference of 32.5 and 45.7 microseconds per sample—tens of times faster than reimplemented ensemble baselines. A sympathetic reader cares because delayed detection in autonomous braking and steering is a physical-safety problem, not only a cybersecurity metric.","feed_headline":"IoV intrusion detection in under 50 microseconds","feed_subtitle":"A compact quantum-inspired DQN matches strong accuracy while beating ensemble latency by 50–70×.","key_machinery":"QIRL: a lightweight Deep Q-Network whose Bellman target is augmented by a Quantum Interference Module (cosine similarity of consecutive latent states, scaled by Λ) and whose exploration uses per-action rotation angles, trained under a cost-sensitive MDP reward that heavily penalizes missed attacks.","core_discovery":"The paper claims that embedding amplitude-phase state encoding, rotation-based exploration, and interference-augmented rewards inside a small DQN, together with asymmetric false-negative/false-positive costs and training-only SMOTE, lets one agent jointly meet competitive detection robustness and sub-millisecond inference on standard network intrusion benchmarks relevant to IoV, closing a dual accuracy-latency gap that prior static and ensemble detectors do not simultaneously satisfy.","pith_inferences":["If the quantum modules add little beyond the tiny net and cost shaping, the same dual objective might be reachable with ordinary DQN plus asymmetric rewards—suggesting a simpler production path for vehicle ECUs.","Extending the same cost-sensitive MDP to true multi-agent V2V/V2I settings could expose whether interference rewards still stabilize learning when neighboring vehicles are also attackers or sensors.","Adversarial packet crafting aimed at the learned Q-boundary would be a natural next stress test, because the paper defers model and physical-layer attacks."],"forward_implications":["Edge nodes on vehicles or RSUs could run intrusion decisions inside the same sub-millisecond budget as braking and steering loops without waiting on cloud ensembles.","Cost-sensitive sequential policies become a practical alternative to static packet classifiers when multi-stage attacks and severe class imbalance dominate IoV traffic.","Training-only SMOTE plus validation-chosen thresholds becomes a required reporting baseline so that reported IoV IDS accuracy is not inflated by leakage.","Latency, not only accuracy, becomes a first-class published metric for any IDS claimed to protect Level 4/5 autonomous driving."],"fun_headline_variants":["Quantum-inspired DQN detects IoV attacks 50-70x faster than ensembles","CA-QIRL reaches 97.9% detection with microsecond-scale V2X inference","Quantum RL cuts vehicular delay 95% while mitigating multi-stage threats","Amplitude-phase encoding lets tiny DQN secure IoV without latency trade-off"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The dual accuracy-plus-speed result is assumed to depend on the quantum-inspired encoding, rotation exploration, and interference reward, rather than mainly on the tiny network, cost-sensitive rewards, and careful train-only balancing.","fun_headline_variants_meta":{"raw":{"variants":["Quantum-inspired DQN detects IoV attacks 50-70x faster than ensembles","CA-QIRL reaches 97.9% detection with microsecond-scale V2X inference","Quantum RL cuts vehicular delay 95% while mitigating multi-stage threats","Amplitude-phase encoding lets tiny DQN secure IoV without latency trade-off"]},"model":"grok-4.5","effort":"low","cost_usd":0.010984,"raw_usage":{"total_tokens":2458,"prompt_tokens":845,"num_sources_used":0,"completion_tokens":91,"cost_in_usd_ticks":109840000,"prompt_tokens_details":{"text_tokens":845,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1522,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":845,"tokens_out":91,"duration_ms":14959,"temperature":1.0,"reasoning_tokens":1522,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T18:16:35.887992+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the same train-only SMOTE, cost weights, and tiny three-layer DQN with those three quantum modules fully disabled; if accuracy, TSS, and sub-50 μs latency remain statistically unchanged on the same CICIDS2017 and UNSW-NB15 splits, the claim that the quantum-inspired machinery is necessary for the dual objective fails.","supporting_citations":[],"review_version":2}