{"id":"a3076cc9-1bcd-4b95-b493-3b25a34d6a76","arxiv_id":"2505.21018","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A spectrum-blind operator can identify which OSaaS user caused interference with 90.3% accuracy using a positional 1D-CNN on its own probe channels.","lead":"This paper reports a machine learning system that detects which of three optical spectrum users is responsible for interference on a shared fiber, using only coarse power readings and the operator's own channel monitors. It reports 90.3% accuracy on a 190 km testbed, suggesting interference attribution may be possible without seeing the users' internal spectrum.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Temporal autocorrelation in 30-s polling may make the reported 90.31% accuracy an artifact of random train/test splitting; a time-based or event-grouped split is needed before the central claim is credible.","rationale":"The reader's CONDITIONAL verdict is appropriate, but I would prioritize a different load-bearing concern than the one named as the weakest assumption. The reader's rationale already mentions a time-based split as a desirable check, yet the weakest assumption is framed as ASE emulation. In my read, the temporal autocorrelation issue is more pressing because it can invalidate the reported accuracy even within the single testbed. If the same interference configuration is polled repeatedly every 30 seconds and a random stratified split is used, the model is likely to see near-duplicates of the same event in both training and testing; this would inflate accuracy and F1 scores without demonstrating detection of novel events. The ASE-emulation concern is real and should also be addressed, but it is a generalization limitation, whereas the temporal-split issue concerns the validity of the headline number itself. I am not claiming any misconduct; this is a common but serious validation gap. The proposed concrete test is a standard event-grouped or time-based split that would settle the issue. Until that is reported, the paper should remain conditional, with the time-based split added as an explicit acceptance condition.","tokens_in":11566,"tokens_out":6211,"duration_ms":75813,"concrete_test":"Re-run the exact training and hyperparameters but construct the split by time/event: assign all 30-s polls belonging to the same interference configuration (same user, interference type, channel positions, power level, contiguous time block) to either train or test, never both; for example train on the first 75% of configurations/events and test on the remaining 25%, maintaining class balance as much as possible. Report accuracy and per-user F1 for the three interference types. If the 90.31% accuracy and ≥87% minimum F1 survive this split, the temporal-leakage concern is refuted; if they drop materially, the published evaluation is not measuring generalization to new events.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a 1D-CNN can classify the source of OSaaS interference from operator-visible data alone. The most load-bearing unaddressed assumption is that the reported 90.31% accuracy is an unbiased estimate of performance on new interference events. Section 4 states that measurements are collected by polling ROADMs and transceivers every 30 seconds, and Section 5 states that the data were split 3:1 with stratified random splits across classes. No grouping by time or by interference event is reported. If the same configuration persists for multiple 30-s polls, the test set contains near-duplicate samples of interference states already seen in training. Under a random split, those duplicates are split across train and test, so the classifier can memorize the event rather than learn a transferable interference signature. This is internal to the evaluation and directly affects the central claim: an operator needs to detect an unseen interference event, not re-identify fragments of previously observed ones. The paper also limits generalization with only three fixed users and ASE-emulated user spectra, but the temporal-split issue is more fundamental because it affects even the single-testbed numbers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses interference attribution in multi-user Optical Spectrum as a Service (OSaaS) networks. The authors propose a 1D-CNN architecture with sinusoidal positional encoding that, using only operator-visible optical performance monitoring (OPM) data from four probe channels and ROADM power telemetry, classifies the source of three types of interference (power increase, ADD/DROP, and rogue OOK) among three OSaaS users. Experimental data were collected on a 190 km Open Ireland testbed with commercial equipment. The model reports 90.31% classification accuracy and a minimum per-user F1-score of 87%, outperforming an MLP baseline at 82.53%.","tokens_in":11790,"tokens_out":6238,"duration_ms":62220,"significance":"If the reported results hold under a properly event-independent evaluation, the paper makes a useful contribution by showing that interference source attribution is possible from operator-visible data alone, without user spectrum visibility. The use of a real optical testbed with commercial transceivers, a well-defined service handover sheet, and a comparison against an MLP baseline are strengths. The F1-score analysis for an imbalanced multi-class problem is appropriate. The main caveat is that the headline accuracy has not yet been demonstrated to generalize to unseen interference events, which is the operational scenario claimed in the abstract.","major_comments":[{"comment":"The evaluation uses a random stratified 3:1 split without any grouping by time or by interference event. Because measurements are polled every 30 seconds (Section 4), successive samples recorded under the same configuration are near-duplicates; a random split will place these duplicates in both training and test sets. The reported 90.31% accuracy may therefore reflect memorization of specific experimental states rather than the ability to classify a genuinely new interference event, which is the operational scenario the abstract claims. Please re-evaluate using a split that groups samples by interference event or by contiguous time window (e.g., all polls of a given configuration in one partition), and report the resulting accuracy and F1 scores. Also report performance over multiple training seeds with means and standard deviations.","section":"Sections 4 and 5 (Data Collection and Model Architecture)"},{"comment":"The text states that optimal hyperparameters \"were identified via a randomized-grid search using stratified K-fold cross validation\" immediately before stating that \"data splitting for performance evaluation maintained a 3:1 training-to-test ratio.\" It is not clear whether the final test partition was kept separate from the data used for hyperparameter selection. If cross-validation included the test samples, the reported test accuracy is optimistic. Please clarify the order of operations, and if necessary re-run the experiment with a held-out test set that is not used for model selection.","section":"Section 5 (Model Architecture, hyperparameter search)"},{"comment":"The user spectra are emulated by shaping ASE noise into 50 GHz channels, with only a single real coherent Cassini channel per user. Nonlinear crosstalk and XPM signatures from ASE-shaped \"channels\" are not necessarily representative of those generated by fully loaded coherent transponders, so the learned interference signatures may not transfer to real OSaaS users. The paper does not acknowledge this limitation in the conclusions, where it states the framework \"can accurately locate the source of power/PSD violations and rogue OOK channels\" in a general spectrum-blind scenario. Please add a careful discussion of this proxy and, ideally, a validation against a fully coherent multi-channel user setup, or soften the generalization claims accordingly.","section":"Section 2.B (OSaaS Experimental Setup)"}],"minor_comments":[{"comment":"In the sentence \"Firstly we aim to understand ... Secondly, the channel-related features ...\", the second reason appears incomplete; please rephrase to clearly state why the aggregate OSaaS window power is unsuitable.","section":"Section 4 (Data Collection)"},{"comment":"The padding value of 40 is introduced without explanation; please clarify the dimensionality of the concatenated feature representation and how the padding constant is derived.","section":"Section 5 (Model Architecture)"},{"comment":"The phrase \"a decaying learning rate of 0.001\" is ambiguous; please specify the initial learning rate and the schedule used during training.","section":"Section 5 (Model Architecture)"},{"comment":"In the text, \"the F1-scores for each user was calculated\" should be \"were calculated.\"","section":"Section 6 (Results)"}],"recommendation":"major_revision","confidential_remarks":"The temporal-leakage concern is the key risk in this paper. If a re-run with an event-grouped or time-based split still yields reasonably high accuracy, the paper is likely publishable after revision; if not, the central claim would be substantially weakened. The ambiguity of the hyperparameter-search protocol should also be resolved explicitly. The ASE emulation is a secondary but important external-validity concern that should at least be discussed in the final version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a real, incremental extension of the authors' ECOC 2024 work — longer 190 km topology, a rogue OOK interference scenario, and a 1D-CNN with positional encoding — aimed at a genuine operational gap: attributing interference to a specific user when the operator is blind to the user's internal spectrum. Second, the reported 90.31% accuracy is plausible on the testbed, but I would not trust it as a measure of how the model will do on a new interference event until the temporal structure of the data is handled.\n\nWhat the paper does well: it tackles a problem that matters in open line systems, with commercial ROADMs, transceivers, and real fiber. The idea of using the operator's own channels as dedicated probe/guard channels is sensible, and the three interference types — power increase, ADD/DROP, and OOK — are the right ones to test. The model architecture is not exotic: 1D convolutions plus sinusoidal positional encoding, which is a standard combination. But the paper argues clearly why the CNN should beat a flattened MLP, and the comparison with an MLP baseline supports that. The confusion matrices and F1 scores are internally consistent, and the paper is honest about the scalability limits.\n\nThe soft spots: the stress-test note is correct. Measurements are polled every 30 seconds, and the train/test split is random and stratified by class, not grouped by time or interference event. If the same configuration persists for multiple polls — which it does in these experiments, since the power/drop/OOK settings are held while data is collected — then the test set almost certainly contains near-duplicates of training samples. That inflates the accuracy and directly undermines the central claim that the model can detect an unseen interference event. This is not a minor quibble; it affects the validity of the numbers even on the single testbed. The paper also reports no error bars or repeated-seed analysis, so we have no sense of variance. The ASE-based emulation of user spectra is a limitation, though the authors explicitly state it; real coherent transponders may leave different fingerprints, and generalization to other spectral layouts or user counts is unvalidated.\n\nBottom line: a useful, sincere experimental contribution that deserves a serious referee, but the evaluation needs a time-based or event-grouped split — and ideally released data and code — before the 90.31% claim is convincing. I would send this to peer review and ask for those fixes.","headline":"Solid experimental paper on a real OSaaS problem, but the random split without temporal grouping makes the headline accuracy untrustworthy as a measure of detecting unseen events.","tokens_in":12330,"tokens_out":2927,"would_cite":true,"duration_ms":30339,"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 neural network can attribute optical interferences to a specific spectrum-sharing tenant using only operator-visible data.","keywords":["interference detection","optical spectrum as a service","spectrum-blind monitoring","1D convolutional neural network","positional encoding","elastic optical network","cross-phase modulation","optical performance monitoring"],"falsifier":"Train the model on the paper's ASE-based setup, then test it on a physically identical line in which each user's 400 GHz window is loaded with actual coherent transponders across all eight 50 GHz channels; if per-user F1 for add/drop or OOK attribution falls below 87%, the result is specific to the emulator rather than to OSaaS traffic in general.","tokens_in":11345,"feed_emoji":"🔎","tokens_out":5247,"duration_ms":53645,"temperature":0.7,"pith_summary":"The paper sets out to show that a network operator can identify which tenant is causing an optical interference in an Optical Spectrum as a Service (OSaaS) network—where users lease spectral windows on a shared fiber—even though the operator cannot see any tenant's internal spectrum configuration. It focuses on three realistic impairments: total power rising above the agreed limit, coherent channels being added or dropped while total power stays flat, and a rogue 10 Gbit/s OOK channel inserted among coherent traffic. The proposed 1D-CNN, fed only with the operator's own channel performance telemetry and coarse optical power readings, classifies the interfering user with 90.31% accuracy and a minimum per-user F1-score of 87% on a 190 km three-user experimental line. If this holds, an operator can attribute faults to a specific tenant and take mitigation action without deploying spectrum analyzers or tapping into user equipment.","feed_headline":"Neural network names the tenant behind optical interference","feed_subtitle":"Trained only on operator-visible power and performance data, it hits 90% accuracy on three interference types.","key_machinery":"The load-bearing object is a 1D-CNN applied to an $(m,n)$ measurement matrix whose rows are the operator channels ordered by frequency and whose columns are ordered optical performance monitoring parameters. Two parallel branches add sinusoidal positional encodings, one along the channel-frequency index and one along the component index, so the convolution keeps local interference patterns while retaining absolute position in the spectrum and along the lightpath. A padding of 40 after concatenation keeps the fully connected layer's input size fixed when more channels or components are added later. This combination is what lets a model attribute a shared-fiber impairment to a specific spectral window using only aggregate operator-side observables.","core_discovery":"The central claim is that interference source attribution in a spectrum-blind OSaaS network is learnable from operator-visible data alone. The operator's own 400G 64-QAM channels act as guard probes interleaved between the spectral windows of three users; their per-channel OPM features (pre-FEC BER, Q-factor, OSNR, polarization and dispersion parameters) plus ROADM power readings form the only input. A 1D-CNN with sinusoidal positional encodings along the frequency axis and along the component axis classifies each measurement as no-interference, User-1, User-2, or User-3, reaching 90.31% test accuracy and a minimum 87% F1 across users and interference types. The same model outperforms a fully connected MLP baseline (82.53%), and it is most stressed by the OOK case, where fast polarization fluctuations are hardest to track.","pith_inferences":["Because each user's 400 GHz window was emulated with ASE-shaped noise plus a single Cassini transceiver at the window center, the learned signatures may not include the richer nonlinear interactions of a window fully loaded with coherent transponders; a model trained this way should be revalidated against real tenant traffic before deployment.","The positional-encoding design suggests a testable generalization: a model trained on one spectral plan might transfer to other bandwidth allocations if the encoding is tied to physical frequency rather than channel index, but the paper does not test this.","A natural extension is severity regression or multi-user attribution: instead of naming a single interferer, predict the launch-power offset of the rogue source, or detect two tenants interfering at once, since the measured telemetry encodes gradual power ramps."],"forward_implications":["Operators can localize an interfering OSaaS tenant from their own transceiver telemetry and ROADM power monitors, without any visibility into tenant spectral loading.","The three studied impairments—power-limit violation, channel add/drop under constant total power, and rogue OOK insertion—are all attributable to the right user at F1 above 87%.","The architecture's fixed final-layer input, via padding, is designed to absorb additional operator channels or monitoring components without reinitializing the classifier.","Adding new users changes the output space and requires fine-tuning of the final layer, and the authors point to existing service-installation windows as the opportunity to collect such data.","The OOK case is the hardest; the paper attributes this to rapid polarization fluctuations from cross-polarization modulation in the intensity-modulated channel."],"supporting_citations":[{"why":"Introduces the OSaaS service concept and the installation window used later for fine-tuning new users.","marker":"[1]"},{"why":"Supplies the operator survey showing power/PSD management concerns and the service requirement sheet used in the experiment.","marker":"[6]"},{"why":"Provides the prior ML fault-detection work from the user perspective that this paper contrasts with the operator perspective.","marker":"[7]"},{"why":"Is the authors' preliminary ECOC analysis that this longer, three-user study extends.","marker":"[8]"},{"why":"Motivates the high-power jamming attack mechanism behind the power-increase interference type.","marker":"[13]"},{"why":"Gives the nonlinear interference theory explaining why higher per-channel power in the add/drop case degrades neighbors.","marker":"[14]"},{"why":"Establishes cross-phase modulation from OOK channels as the dominant interference mechanism for coherent channels.","marker":"[15]"},{"why":"Shows performance degradation when legacy 10G OOK signals are overlaid with coherent terminals, supporting the OOK use case.","marker":"[16]"},{"why":"Provides theory and measurement of cross-polarization modulation, explaining the fast SOP fluctuations in the OOK case.","marker":"[17]"},{"why":"Supplies the sinusoidal positional encoding technique used in the two network branches.","marker":"[20]"}],"fun_headline_variants":["ML blames the right user for optical interference, blind to spectra","Spectrum-blind AI identifies interference source with 90% accuracy","Neural net names the interfering tenant without seeing user spectra","90% accuracy: ML attributes optical interference from blind spectrum data","Operator-visible data alone lets ML finger the interfering user"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The threshold assumption is that user spectra loaded as shaped ASE noise behave, for interference purposes, like real OSaaS tenant transceivers; if real multi-wavelength coherent signals produce different monitoring signatures, the trained classifier may not transfer to production.","fun_headline_variants_meta":{"raw":{"variants":["ML blames the right user for optical interference, blind to spectra","Spectrum-blind AI identifies interference source with 90% accuracy","Neural net names the interfering tenant without seeing user spectra","90% accuracy: ML attributes optical interference from blind spectrum data","Operator-visible data alone lets ML finger the interfering user"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001996,"raw_usage":{"total_tokens":7789,"prompt_tokens":942,"completion_tokens":6847,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":6762}},"tokens_in":558,"tokens_out":6847,"duration_ms":48721,"temperature":1.0,"reasoning_tokens":6762,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:42:19.690338+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the model on the paper's ASE-based setup, then test it on a physically identical line in which each user's 400 GHz window is loaded with actual coherent transponders across all eight 50 GHz channels; if per-user F1 for add/drop or OOK attribution falls below 87%, the result is specific to the emulator rather than to OSaaS traffic in general.","supporting_citations":[{"cited_title":"Employing channel probing to derive end-of-life service margins for op- tical spectrum services,","cited_arxiv_id":null,"evidence_quote":"Introduces the OSaaS service concept and the installation window used later for fine-tuning new users."},{"cited_title":"Optical spectrum services in open disaggregated transport networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the operator survey showing power/PSD management concerns and the service requirement sheet used in the experiment."},{"cited_title":"Machine Learn- ing enabled Fault-Detection Algorithms for Optical Spectrum-as-a- Service Users,","cited_arxiv_id":null,"evidence_quote":"Provides the prior ML fault-detection work from the user perspective that this paper contrasts with the operator perspective."},{"cited_title":"In- terference identification in multi-user optical spectrum as a service using convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"Is the authors' preliminary ECOC analysis that this longer, three-user study extends."},{"cited_title":"High- power jamming attack mitigation techniques in spectrally-spatially flexi- ble optical networks,","cited_arxiv_id":null,"evidence_quote":"Motivates the high-power jamming attack mechanism behind the power-increase interference type."},{"cited_title":"Nonlinear interference mitigation: Methods and potential gain,","cited_arxiv_id":null,"evidence_quote":"Gives the nonlinear interference theory explaining why higher per-channel power in the add/drop case degrades neighbors."},{"cited_title":"Cross-phase mod- ulation induced by ook channels on higher-rate dqpsk and coherent qpsk channels,","cited_arxiv_id":null,"evidence_quote":"Establishes cross-phase modulation from OOK channels as the dominant interference mechanism for coherent channels."},{"cited_title":"Overlaying 10 gb/s legacy optical networks with 40 and 100 gb/s coherent terminals,","cited_arxiv_id":null,"evidence_quote":"Shows performance degradation when legacy 10G OOK signals are overlaid with coherent terminals, supporting the OOK use case."},{"cited_title":"Cross-polarization modulation: Theory and measurement in subcarrier-modulated wdm systems,","cited_arxiv_id":null,"evidence_quote":"Provides theory and measurement of cross-polarization modulation, explaining the fast SOP fluctuations in the OOK case."},{"cited_title":"Attention is All you Need,","cited_arxiv_id":null,"evidence_quote":"Supplies the sinusoidal positional encoding technique used in the two network branches."}],"review_version":1}