{"id":"63b42b6b-ebd0-4d19-9442-833a1b644c27","arxiv_id":"2608.08123","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Artifacts in 5G NR sensing are modeled as multiplicative-noise-induced periodic range extensions from Zadoff-Chu pilots, and a physics-informed masked encoder-decoder suppresses them, boosting detection probability to 98.88% in field tests.","lead":"This paper derives a mathematical model of striped artifacts that appear in 5G radio sensing images, then builds a neural network that removes them. Real measurements with commercial equipment show target detection rising from about 73% to nearly 99% while false alarms drop sharply.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed detection gain is unreliable because Remark 7 says an ISI ghost is treated as a second ground-truth target, so the CFAR hit criterion and/or training labels count artifact peaks as true detections; the 98.88% figure needs re-scoring on the single RTK target.","rationale":"The strongest part of the paper is Theorem 2 and its real-world validation: Scenario A shows predicted artifact spacings (40.3 / 80.5 / 241.5 m for μ = 51 / 102 / 306) matching measured peaks at 64.0 / 104.3 / 265.5 m, which is genuine independent evidence for the ZC artifact model. Theorem 1's proof is narrower than its statement (Appendix A assumes g_l[n] = g_l), but that does not by itself undermine the mask positions in Eq. (32), which rely only on range periodicity. The performance evaluation is the load-bearing weakness. Remark 7 explicitly says a single UAV is scored as two targets after insufficient-CP ISI. Because X^(0) in Eq. (38) is the training target and CFAR scoring uses RTK-derived peaks, adding the ghost as a second true target converts a false alarm into a true positive, directly inflating P_d and deflating N_fa. The baseline inconsistency (79.58% vs 73.33%) further obscures the actual gain. These issues are correctable, so they do not prove the network is ineffective; they mean the current numbers cannot be trusted. Thus the reader's CONDITIONAL verdict remains appropriate, with the experimental section needing a corrected re-evaluation before the 98.88% claim is cited.","tokens_in":21785,"tokens_out":12236,"duration_ms":128489,"concrete_test":"Re-run the Section IV-C experiment with the single RTK-derived UAV peak as the only ground truth: exclude the ISI ghost from X^(0) training labels and from the ±2-bin CFAR hit criterion, and report P_d, N_fa, and PSNR for raw RV, median filter, ECA, U-Net, Ra-SPD, and PIAENet; also reconcile the baseline discrepancy (79.58% vs 73.33%) under this protocol. If PIAENet still reaches roughly 98.88% P_d with the false-target CDF left-shifted, the concern is resolved; if P_d falls materially or false-target counts rise, the headline detection gain is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-C, Remark 7 states: \"although only a single UAV was used, two target points appeared... we do not address this phenomenon and treat the ground truth as corresponding to two targets.\" This is load-bearing for the headline empirical claim. Ground truth for PIAENet training (X^(0) in Eq. 38) and for CFAR scoring (Section IV-C; the ±2-bin RTK hit rule) is supposed to be derived from the single UAV's RTK track (Fig. 8). If the ISI ghost is added as a second \"target,\" detections at the ghost are counted as true positives, inflating P_d in Fig. 12 and deflating false-target counts in Fig. 13. This biases every learning-based method, but PIAENet's masked reconstruction is specifically designed to restore artifact-like structures, so its 98.88% P_d cannot be taken at face value. The related baseline inconsistency (abstract: 79.58%; body/Fig. 12: 73.33%) is unresolved and compounds the problem. The artifact-spacing validation in Scenario A is independent and supports Theorem 2, but it does not validate the mitigation numbers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper addresses striped artifacts observed in 5G NR integrated sensing and communication (ISAC) range-velocity spectra when polyphase pilot sequences such as Zadoff-Chu (ZC) codes are used. The authors model non-ideal hardware impairments as time-domain multiplicative noise, prove that this noise causes Doppler-domain spreading (Theorem 1) and code-dependent range-domain extension with periodic offsets c0*mu*m/(2K*delta_f) (Theorem 2), and then propose PIAENet, a multi-frame selective-masked encoder-decoder trained to reconstruct a clean RV spectrum. The throughput of the paper is twofold: a theoretical artifact-formation model, and a data-physics-driven mitigation framework. The model is evaluated against rooftop measurements with three ZC root indices, and the mitigation performance is evaluated on UAV field measurements with RTK ground truth, plus simulations over varying root indices and trajectories. The paper claims detection probability improvement from 79.58% to 98.88% and a substantial false-target-count reduction.","tokens_in":21981,"tokens_out":3950,"duration_ms":43381,"significance":"The artifact-spacing prediction of Theorem 2, expressed in Eq. (55) as c0*mu/(2K*delta_f), is a genuine, parameter-free theoretical result. It is backed by locally consistent derivations in Appendices A-C and by direct experimental confirmation for three root indices (mu = 51, 102, 306) in Section IV-B, where the observed artifact offsets match the predicted periodicities. This part of the paper is a useful contribution to 5G NR ISAC signal processing. The PIAENet architecture is also interesting, and the release of the dataset is a practical strength. However, the central empirical claim — that PIAENet raises detection probability to 98.88% and suppresses false targets — is weakened by the treatment of an ISI ghost as a second ground-truth target in Remark 7, by an unresolved inconsistency between the abstract's baseline (79.58%) and the body's baseline (73.33%), and by a robustness simulation that generates test data from the same theoretical model used to design the network. These issues affect the significance of the mitigation results but do not invalidate the theoretical artifact characterization.","major_comments":[{"comment":"The detection-probability and false-target-count results are scored with a CFAR rule that counts a peak as correct if it lies within a ±2-bin neighborhood of the RTK-derived ground-truth peak. Remark 7 states that although only one UAV was present, two target points appeared and the ground truth is treated as corresponding to two targets. Consequently, detections of the ISI ghost are counted as true positives, inflating the P_d values in Fig. 12 and deflating the false-target counts in Fig. 13. This bias affects all methods but is particularly problematic for PIAENet, whose masked reconstruction is specifically designed to restore artifact-like structures. The 98.88% headline figure is therefore not a trustworthy single-target detection probability. Please re-score all detection metrics against the single RTK-derived UAV target and report both the single-target and two-target interpretations explicitly.","section":"Section IV-C, Remark 7 and CFAR scoring"},{"comment":"The abstract claims the detection probability is improved from 79.58% to 98.88%, while Section IV-C and Fig. 12 report a raw-baseline P_d of 73.33%. This is a direct factual inconsistency in a central quantitative claim. The paper must state which baseline is correct, or explain the difference (e.g., different test sets, different CFAR settings), and ensure the abstract, body, and figures agree before the mitigation claims can be assessed.","section":"Abstract vs. Section IV-C, Fig. 12"},{"comment":"The simulation study is described as follows: 'the simulated data are generated based on the theoretically derived model, they inherently preserve the same range extension and Doppler spreading characteristics as the measured artifacts.' Since the mask in Eq. (32) and the network training are also derived from Theorem 2's artifact model, the simulation in Tables IV and V validates robustness only within the same assumed model family. It does not demonstrate robustness to non-multiplicative impairments (I/Q imbalance, nonlinear amplification, timing drift) or to delay/Doppler variations that violate the assumptions of Section II-B. The simulation is useful for interpolation across root indices and trajectories, but the claim that it provides 'broader evidence' of generalization should be tempered or supplemented with a mismatched-model test.","section":"Section IV-E, robustness simulation"},{"comment":"Theorems 1 and 2 are stated as general results about time-domain multiplicative noise, but the proofs in Appendices A and B impose substantial restrictions: Appendix A assumes g_l[n] = g_l for all n within a symbol and ignores delay (tau = 0), and Appendix B assumes N = K and ignores Doppler (f_d = 0). These restrictions are not stated in the main-text theorem statements, and their practical import is unclear for impairments that vary within a symbol or that are non-multiplicative. Please state the exact conditions under which Theorems 1 and 2 hold in the main text, and indicate which components of the artifact characterization are heuristic outside those conditions.","section":"Section II-B and Appendices A-B"}],"minor_comments":[{"comment":"The text says the trajectory-robustness table entries are 'temporarily left blank and will be filled after the corresponding experiments are finalized,' but Table V already contains numerical entries. Please remove the placeholder sentence or mark the table as preliminary if the results are not final.","section":"Section IV-E, Table V"},{"comment":"The statement that the artifacts from the static reflector at 23.7 m are 'primarily attributable to residual CFO or local-oscillator phase noise rather than insufficient-CP-induced truncation' is an inference from the reflector being within the CP-supported range, not a direct measurement of the impairment. Please present it as an interpretation and describe any supporting diagnostic if available.","section":"Section IV-B"},{"comment":"The mask definition uses d_0 = arg max_d |X[d,v]| with the velocity index v undefined. Please clarify whether the maximum is taken over all velocity bins, a fixed Doppler bin, or some other reduction, since the mask range depends on this choice.","section":"Equation (32)"},{"comment":"In Eq. (29b), the subscript p is dropped from eta, so the two equations are not notationally consistent. Use eta_p in both equations.","section":"Equation (29)"},{"comment":"The caption states 'The target is obscured by the artifacts' for all panels, but in several panels (e.g., (f)) the target is clearly visible. Adjust the caption to describe the varying degree of obscuration across methods.","section":"Figure 10 caption"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has one result that stands on its own and one that does not. The range-domain artifact spacing formula c0*mu/(2K*delta_f) for ZC pilots is new, derived cleanly in Appendix B with no fitted parameters, and supported by rooftop measurements at three root indices. That part is worth taking seriously. The PIAENet detection numbers, on the other hand, are not reliable as reported, because Remark 7 says an ISI ghost is treated as a second ground-truth target. That means CFAR scoring and training labels count artifact peaks as true detections, and the headline 98.88% P_d is inflated. The abstract/body baseline inconsistency (79.58% vs 73.33%) is unresolved and compounds the problem. The robustness simulation also generates test data from the same theoretical model used to design the network, so it does not validate generalization; it checks self-consistency. The pointwise multiplicative noise assumption in Eq. (10) is reasonable for CFO and phase noise, but it may miss I/Q imbalance and nonlinear distortion; the paper does not claim otherwise, so this is a limitation rather than a flaw. The ablation study is informative: the multi-frame mechanism drives most of the P_d gain, which is plausible. The network architecture is a modest extension of U-Net, but the selective mask derived from the theorem is a legitimate physics-informed design. The dataset release is a plus. Overall, the artifact characterization deserves a serious referee. The detection claims need re-scoring on the single RTK target, the baseline numbers need reconciliation, and the simulation section should be repositioned as a sanity check rather than robustness evidence. I would send it to review with a request for major revision.","headline":"The artifact-spacing derivation is a real contribution worth citing, but the PIAENet detection numbers are inflated by counting an ISI ghost as a second ground-truth target, so the mitigation claims need re-scoring before they can be believed.","tokens_in":22534,"tokens_out":1478,"would_cite":true,"duration_ms":16500,"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":"A physics-aware network lifts 5G sensing detection from 79.58% to 98.88%.","keywords":["integrated sensing and communication","5G NR","Zadoff-Chu sequence","range-velocity spectrum","multiplicative noise","artifact mitigation","physics-informed neural network","target detection"],"falsifier":"Measure a single strong static reflector with a known ZC root while sweeping CFO continuously: if artifact peaks do not appear at range offsets $c_0\\mu m/(2K\\Delta f)$ with spacing $c_0\\mu/(2K\\Delta f)$ for each ICI order $m$, the multiplicative-noise derivation fails. Conversely, injecting a non-multiplicative distortion such as frequency-dependent I/Q imbalance should produce artifacts that violate the Eq. (32) mask, exposing the model's boundary.","tokens_in":21548,"feed_emoji":"📡","tokens_out":4197,"duration_ms":42659,"temperature":0.7,"pith_summary":"This paper tries to establish where the striped artifacts in 5G NR integrated sensing and communication (ISAC) range-velocity spectra come from and how to remove them without modifying the standardized waveform. The paper models non-ideal hardware impairments as time-domain multiplicative noise and proves two consequences: the noise spreads target energy across Doppler bins, and, because the SRS pilot is a Zadoff-Chu polyphase code, it creates periodic range-domain ghost peaks at offsets $c_0\\mu m/(2K\\Delta f)$. These physical priors are then encoded into a multi-frame selective-masked encoder-decoder network, PIAENet, that reconstructs the clean spectrum. On measured commercial mmWave data, the network raises detection probability from 79.58% to 98.88% while reducing false targets. If correct, it gives a waveform-design degree of freedom that controls where artifacts appear and a network architecture that targets exactly those locations.","feed_headline":"Physics-aware network lifts 5G sensing detection to 98.88%","feed_subtitle":"Paper derives where Zadoff-Chu artifacts appear in range-velocity spectra and masks exactly those bins.","key_machinery":"The central object is the time-domain multiplicative noise $g_l[n]$ in Eq. (10), coupled with the Zadoff-Chu pilot structure. The key identity is the pilot ratio $S^\\mathrm{ZC}_{k-m}/S^\\mathrm{ZC}_k = e^{j(\\pi\\mu m^2/K + \\pi\\mu m/K)} e^{j2\\pi\\mu m k/K}$, which is a single linear-phase term in the subcarrier index $k$; each ICI order $m$ therefore becomes a shifted delta-like kernel in the delay domain at $\\tau_0 \\pm \\mu m/(K\\Delta f)$. This identity is what converts a generic multiplicative impairment into code-dependent, periodically spaced range artifacts. The paper's Eq. (32) mask places network attention at range bins $d_0 \\pm k c_0\\mu/(2K\\Delta f)$, so the architecture inherits the derived physics directly.","core_discovery":"The central claim is that non-ideal transceiver effects in 5G NR sensing behave as pointwise multiplicative noise in the time domain, and for polyphase-coded pilots this creates two coupled artifacts in the range-velocity spectrum. Theorem 1 states that time-domain multiplicative noise causes Doppler-domain spreading through convolution with the noise spectrum, raising the noise floor across velocity bins. Theorem 2 states that each inter-carrier-interference order $m$ produces a pair of delay-domain kernels centered at $\\tau_0 \\pm \\mu m/(K\\Delta f)$, corresponding to range offsets $c_0\\mu m/(2K\\Delta f)$ with spacing $c_0\\mu/(2K\\Delta f)$. The paper validates this periodicity with measured data for ZC roots $\\mu=51,102,306$, including a false peak at 64.0 m from a 23.7 m reflector for $\\mu=51$. It then shows that a selective mask placed on these predicted range bins, combined with multi-frame stacking, yields PIAENet, which reconstructs the ideal RV spectrum and outperforms U-Net and Ra-SPD in peak signal-to-noise ratio, detection probability, and false-target suppression.","pith_inferences":["The multiplicative-noise model likely captures CFO and insufficient-CP truncation well, but non-multiplicative distortions such as I/Q imbalance, nonlinear amplification, or timing drift could create artifacts at positions outside the predicted grid; a controlled injection experiment would reveal where the model's boundary lies. ","Appendix C suggests the same shifted-kernel machinery applies to matrix-polyphase codes like Frank, P1, and P2, so the selective-mask strategy could transfer to other pilot families with different artifact amplitudes rather than different candidate delay grids. ","The artifact-position formula also offers a diagnostic tool: comparing measured ghost locations against $c_0\\mu/(2K\\Delta f)$ could identify which hardware impairment dominates in a given deployment, since CFO and truncation produce the same range periodicity but different frequency-domain weightings. "],"forward_implications":["Choosing the ZC root index $μ$ can control artifact placement, for example overlapping artifacts from different ICI orders with $μ=K/2$ or $μ=K/4$, or keeping dominant artifacts close to the strong reflector. ","Because the artifacts are tied to the pilot code and not to a particular device, the mitigation works across commercial mmWave and sub-6 GHz hardware without changing the standardized 5G NR waveform. ","Artifact elimination in the RV domain can be treated as a supervised reconstruction problem whose ground truth is generated from RTK-measured target positions and velocities. ","The network's two physical priors contribute differently: the selective mask identifies where reconstruction is needed, while multi-frame stacking supplies temporal information that suppresses residual artifacts without removing weak targets. "],"supporting_citations":[{"why":"Supplies the Zadoff-Chu polyphase code definition and its periodic correlation properties, which the artifact derivation relies on.","marker":"[11]"},{"why":"Defines the 5G NR SRS pilot structure and ZC-based reference signal generation, fixing the pilot phase model used in Eqs. (1) and (48).","marker":"[35]"},{"why":"Provides the CA-CFAR detection procedure used to compute detection probability and false-target counts for all compared methods.","marker":"[38]"},{"why":"Supplies the U-Net encoder-decoder architecture that is a direct deep-learning baseline and the structural basis for PIAENet.","marker":"[37]"},{"why":"Provides the spectral-spatial decomposition baseline Ra-SPD, which PIAENet is compared against in artifact elimination.","marker":"[33]"},{"why":"Supplies the CNN classification approach used by the subregion classifier that filters out target-free regions before artifact elimination.","marker":"[36]"},{"why":"Motivates the temporal-continuity prior by linking target motion dynamics to consecutive radar frames.","marker":"[34]"}],"fun_headline_variants":["Physics-aware masking clears 5G sensing artifacts","Selective mask on predicted bins lifts 5G detection to 98.88%","ZC artifacts mapped; PIAENet masks them for 5G sensing","Why 5G sensing sees ghosts: theoretical artifact model","Physics-driven masking improves 5G detection from 79.58% to 98.88%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire artifact model assumes every non-ideal impairment acts as pointwise multiplication in the time domain, with delay and Doppler constant over the coherent processing interval and the pilot phase structure unchanged across symbols; if real hardware also produces non-multiplicative distortions, the predicted artifact positions and the PIAENet mask in Eq. (32) will be misplaced.","fun_headline_variants_meta":{"raw":{"variants":["Physics-aware masking clears 5G sensing artifacts","Selective mask on predicted bins lifts 5G detection to 98.88%","ZC artifacts mapped; PIAENet masks them for 5G sensing","Why 5G sensing sees ghosts: theoretical artifact model","Physics-driven masking improves 5G detection from 79.58% to 98.88%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001106,"raw_usage":{"total_tokens":4679,"prompt_tokens":1083,"completion_tokens":3596,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":699,"completion_tokens_details":{"reasoning_tokens":3498}},"tokens_in":699,"tokens_out":3596,"duration_ms":26870,"temperature":1.0,"reasoning_tokens":3498,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:22:25.236497+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure a single strong static reflector with a known ZC root while sweeping CFO continuously: if artifact peaks do not appear at range offsets $c_0\\mu m/(2K\\Delta f)$ with spacing $c_0\\mu/(2K\\Delta f)$ for each ICI order $m$, the multiplicative-noise derivation fails. Conversely, injecting a non-multiplicative distortion such as frequency-dependent I/Q imbalance should produce artifacts that violate the Eq. (32) mask, exposing the model's boundary.","supporting_citations":[{"cited_title":"Polyphase codes with good periodic correlation properties (Corresp.),","cited_arxiv_id":null,"evidence_quote":"Supplies the Zadoff-Chu polyphase code definition and its periodic correlation properties, which the artifact derivation relies on."},{"cited_title":"NR; Physical channels and modulation,","cited_arxiv_id":null,"evidence_quote":"Defines the 5G NR SRS pilot structure and ZC-based reference signal generation, fixing the pilot phase model used in Eqs. (1) and (48)."},{"cited_title":"Radar CFAR thresholding in clutter and multiple target situations,","cited_arxiv_id":null,"evidence_quote":"Provides the CA-CFAR detection procedure used to compute detection probability and false-target counts for all compared methods."},{"cited_title":"U-Net: Convolutional networks for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the U-Net encoder-decoder architecture that is a direct deep-learning baseline and the structural basis for PIAENet."},{"cited_title":"Ra-SPD: Radar sig- nal interference mitigation using spectral–spatial decomposition,","cited_arxiv_id":null,"evidence_quote":"Provides the spectral-spatial decomposition baseline Ra-SPD, which PIAENet is compared against in artifact elimination."},{"cited_title":"Gradient-based learning applied to document recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies the CNN classification approach used by the subregion classifier that filters out target-free regions before artifact elimination."},{"cited_title":"Convo- lutional long short-term memory networks for doppler-radar based target classification,","cited_arxiv_id":null,"evidence_quote":"Motivates the temporal-continuity prior by linking target motion dynamics to consecutive radar frames."}],"review_version":1}