{"id":"787480e8-1ea3-46db-86f4-8f33925f9a5c","arxiv_id":"2502.09194","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A semi-supervised contractive autoencoder with a fast SHAP explainer detects O-RAN traffic anomalies and identifies the responsible radio features.","lead":"This paper pairs a contractive autoencoder with a fast SHAP explainer to detect traffic anomalies in open radio access networks and to show which radio signal metrics triggered each alarm. It reports strong accuracy with very few labeled examples and faster explanations, which matters for real-time network management.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The semi-supervised loss in Eq. (12) cannot be computed from the architecture in Table IV: no classification head is defined, so the reported UAR gains from labeled samples are not tied to the described method.","rationale":"The reader's weakest assumption identifies the same mismatch between the loss in Eq. (12) and the architecture in Table IV, plus the undefined unlabeled-set construction. I find this to be the most load-bearing concern because the paper's headline contribution is a semi-supervised method that obtains high UAR with very few labels. If the cross-entropy term is not actually computed, the model is just a contractive autoencoder with no use of labels, and the reported advantage over DeepAE at 100 labeled samples has no described mechanism. The manuscript is otherwise a plausible incremental application of known components: contractive autoencoders and fastSHAP are prior work, the public O-RAN dataset is real, and the resource-utilization experiment is not inherently implausible. But the central anomaly-detection claim cannot be verified from the text because the training objective is not implementable as specified. The second issue, the unlabeled-set construction, matters because it determines whether the semi-supervised framing is meaningful at all. A code release or a minimal two-model reimplementation would settle the question; absent that, the current evidence does not support acceptance, and the reader's REJECT verdict stands.","tokens_in":27388,"tokens_out":6141,"duration_ms":60923,"concrete_test":"Obtain or minimally reimplement the training code and instrument the forward pass to check whether a classification head exists between the bottleneck (dense4, dim 16) and the binary label. Then train two models on the same 100-labeled-sample split: model A with Eq. (12) exactly as written, including the cross-entropy term via that head, and model B with alpha_i = 0 for all samples so no supervised signal is used. If B reproduces Table V's 80.17 +/- 0.6 UAR within its confidence interval, the semi-supervised term is not what produces the headline result; if A cannot be implemented without changing Table IV, the manuscript's method description is internally inconsistent.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central minimal-labels claim rests on the supervised cross-entropy term in Eqs. (11)-(12). Computing it requires a head that maps the latent representation z to a binary prediction y-hat, but Table IV contains only reconstruction layers (input 20 -> 64 -> 32 -> 16 -> 16 -> 32 -> 64 -> output 20) and no such head. No code or supplementary material is supplied. The paper also never states how the unlabeled set Du is constructed from a dataset with about 25% anomalies: if Du is sampled from the full data it contains many anomalies and contaminates the reconstruction and contractive objectives, while if anomalies are removed using labels then labels are used to build the so-called unlabeled set. Either way, the UAR values in Table V (e.g., 80.17 +/- 0.6 with 100 labels) cannot be attributed to the described SS-DeepCAE, and the claimed advantage over DeepAE is unsupported by the manuscript as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes XAInomaly, a framework combining a semi-supervised deep contractive autoencoder (SS-DeepCAE) with a fastSHAP-C explainability method for traffic anomaly detection in O-RAN. The authors claim that SS-DeepCAE achieves high unweighted average recall (UAR) with very few labeled samples (e.g., 80.17±0.6 with 100 labels), outperforming baselines including DeepAE, and that fastSHAP-C provides real-time SHAP values with a 34% runtime improvement. The paper includes a description of the model architecture, loss functions, training procedure, XAI algorithm, and experiments on a public O-RAN dataset.","tokens_in":27503,"tokens_out":7203,"duration_ms":56952,"significance":"If the reported results were reproducible, the framework would address a practically relevant problem: interpretable, resource-efficient anomaly detection in O-RAN with limited labeled data. The paper also makes a useful empirical comparison across several autoencoder baselines and two SHAP-based explanation methods, using a public dataset. However, the significance is severely limited by internal inconsistencies in the method description: the supervised loss cannot be computed from the specified architecture, the unlabeled data construction is ambiguous, and the contractive Jacobian formula is incorrect for deep encoders. These issues prevent verification of the central claims.","major_comments":[{"comment":"The semi-supervised loss in Eq. (12) includes the cross-entropy term α_i ℓ_cro(y_i, ŷ_i), which requires a classification head that maps the latent representation z to a binary prediction ŷ. However, the architecture in Table IV contains no such head: the final dense layer outputs a 20-dimensional reconstruction, not a scalar probability. Thus, as described, the supervised term cannot be computed, and the reported UAR gains from labeled samples (Table V, e.g., 80.17±0.6 with 100 labels) are not attributable to the stated method. Please specify the classification head (or reformulate the loss) and update the architecture and parameter counts accordingly.","section":"Sec. III-B, Eq. (12), Table IV"},{"comment":"The construction of the unlabeled set D_u is never specified. The dataset has approximately 25% anomalies; if D_u is built by sampling from the full dataset, it contains a substantial fraction of anomalies, contradicting the statement in Sec. III-A that the unlabeled data consist of 'normal traffic patterns.' If anomalies are instead removed using the available labels, then labels are used to construct the so-called unlabeled set, undermining the semi-supervised claim. The paper must state how D_u is sampled and how the model's reconstruction and contractive objectives treat anomalous samples in D_u.","section":"Sec. III-A, Eqs. (2)-(3)"},{"comment":"Equation (10) gives the Jacobian of the deep encoder as J_i = W^{(l)} diag(f'^{(l-1)}(...)) W^{(l-1)}, but the Jacobian of a multi-layer encoder is a product of per-layer Jacobians (one factor per layer), not the two-matrix expression shown. As written, Eq. (10) cannot be used to evaluate the contractive penalty in Eq. (9), and the mathematical description of the method's core mechanism is incorrect. Please provide the correct Jacobian formula for the full encoder (or state that automatic differentiation is used and remove the explicit incorrect formula).","section":"Sec. III-B, Eq. (10)"},{"comment":"The Confidence Score (CS) is defined inconsistently. In Algorithm 1 (lines 12-13), CS is the average absolute residual f_x,y(s_i) - f_x,y(0) - s_i^T φ̂ over sampled subsets, while in Sec. V-C, Eq. (34), CS is the average absolute difference between f(x_i) and Σ φ_i(x_i) over inputs. The verbal interpretations also conflict: Sec. IV-B states 'A lower CS indicates higher confidence,' whereas Sec. V-C states 'A high CS indicates that the explanations are accurate.' Please harmonize the definition and the interpretation across Sections IV-B and V-C.","section":"Sec. IV-B and Sec. V-C"},{"comment":"Contribution F2 states that 'fastSHAP-C provides 34% advance over its competitors in terms of runtime performance.' The 34% figure reported in Sec. VI-B3 is actually the runtime reduction from fastSHAP-C on DeepAE (33,627 ms) to fastSHAP-C on SS-DeepCAE (22,145 ms) — i.e., an improvement due to changing the underlying autoencoder, not a comparison with competitor XAI methods. Against kernelSHAP (320.4 s) and fastSHAP (48.007 s), the reductions are approximately 93% and 54%, respectively. The claim should be reworded to reflect the actual comparison.","section":"Sec. I (F2) and Sec. VI-B3"}],"minor_comments":[{"comment":"The header 'Hidden layer s.' is garbled, and the selected hidden layer size '3.0 xInput Size' is inconsistent with the 64-unit first hidden layer in Table IV (60 would be 3×20). Please correct the table and reconcile the text.","section":"Sec. V-D, Table III"},{"comment":"The 'normalize' flag in line 7 is undefined, and the computation of CS and EM on lines 12-13 does not specify how the n samples are accumulated or when they are finalized. Please clarify the pseudocode.","section":"Sec. IV-B, Algorithm 1"},{"comment":"The sensitivity metric uses an ε-neighborhood B_ε(x_i) without specifying a value of ε or how the maximization is performed; no sensitivity results are reported, so the metric's role is unclear.","section":"Sec. V-C, Eq. (35)"},{"comment":"The log-odds metric in Eq. (36) is not clearly connected to the exclusion/inclusion AUC values reported in Table VII; please define how the AUC is computed from log-odds.","section":"Sec. V-C, Eq. (36) and Table VII"},{"comment":"The UAR for SS-DeepCAE is 91.33±(1.4) at 1,000 labeled samples and 91.17±(0.6) for 'All' labeled samples; the slight decrease with more labels merits a brief explanation (e.g., different random splits).","section":"Table V"}],"recommendation":"reject","confidential_remarks":"The paper does not provide code or supplementary material, and the described architecture cannot realize the supervised loss, making the central results non-reproducible as written. If the authors can supply a corrected method (classification head, Jacobian, and unlabeled data construction) and rerun the experiments, a resubmission could be considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper combines a contractive autoencoder with a fastSHAP variant for O-RAN anomaly detection. The headline number – 80% UAR with only 100 labeled samples – is exactly what needs scrutiny, because as written the semi-supervised loss cannot be computed from the architecture described.\n\nWhat is genuinely useful: the application is timely (O-RAN, xApps, Near-RT RIC), the baseline comparison is broad (Vanilla-AE, LSTM-AE, SLA-VAE, DeepAE), and the XAI benchmarking with exclusion/inclusion curves is more thorough than most. The CS and EM additions to fastSHAP are minor but reasonable ideas. The paper is readable and the integration into the O-RAN architecture is a nice systems contribution.\n\nSoft spots, in increasing severity. The Jacobian formula in Eq (10) is wrong for a deep encoder – it looks like a single-layer expression misapplied. The CS definition in Algorithm 1 says lower is better, but in Section V-C it says high CS indicates accuracy; that is a direct contradiction. The “34% runtime improvement” is not over competitors; it is fastSHAP-C on SS-DeepCAE vs fastSHAP-C on the authors’ own DeepAE. The unlabeled set construction is never specified, and with about 25% anomalies in the full data, the claim that unlabeled data are mostly normal needs an explanation.\n\nThe load-bearing issue is the missing classification head. Eq (12) includes a cross-entropy term for labeled samples, but the architecture in Table IV has no layer that outputs a prediction. Without that head, the semi-supervised loss is undefined, and the UAR gains from added labels cannot be attributed to the described method. This could be a simple omission in the model summary, but as written it is a serious gap. The paper also ships no code or appendix, so there is no way to check.\n\nWho is this for? Someone working on O-RAN anomaly detection will want to know this literature exists, but I would not trust the headline numbers without code. It deserves a serious referee – the topic is relevant and the experiment design is repairable – but the referee should ask for the classification head, the unlabeled-set construction, and the corrected XAI metric definitions before any acceptance. My recommendation: send to review with a strong request for revisions, not desk reject, but I would not cite it in its current form.","headline":"A promising O-RAN anomaly detection combination undermined by a missing classification head in the semi-supervised loss, plus several smaller but real technical inconsistencies.","tokens_in":28152,"tokens_out":4409,"would_cite":false,"duration_ms":40507,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A semi-supervised contractive autoencoder detects O-RAN traffic anomalies with as few as 100 labeled samples, and a fast SHAP-based explainer makes the detections interpretable in real time.","keywords":["anomaly detection","O-RAN","contractive autoencoder","semi-supervised learning","explainable AI","SHAP values","network management"],"falsifier":"Reproduce the training with the architecture given in Table IV; if no classification head exists, the supervised term in Eq. (12) cannot be evaluated, and removing it would show whether the 80.17 UAR figure actually depends on labeled data. Separately, construct a variant of the unlabeled training pool whose anomaly share approaches 25% (matching the full dataset) and measure UAR; if it collapses toward the vanilla-autoencoder level, the mostly-normal-unlabeled premise is load-bearing.","tokens_in":27076,"feed_emoji":"📡","tokens_out":10155,"duration_ms":82660,"temperature":0.7,"pith_summary":"XAInomaly claims that a semi-supervised deep contractive autoencoder (SS-DeepCAE) can detect anomalies in O-RAN traffic much more accurately than standard deep autoencoders when labels are scarce, and that the accompanying fastSHAP-C explainer makes these detections interpretable in real time. The paper reports unweighted average recall of 80.17 ± 0.6 with only 100 labeled samples, against 71.72 ± 1.5 for the DeepAE baseline, and a 34% runtime reduction for fastSHAP-C compared with the same explainer on DeepAE. This matters because O-RAN's disaggregated, multi-vendor architecture needs anomaly detection that does not depend on large labeled datasets and can explain its alerts to operators. The framework is designed to run as xApps on the Near-RT RIC, connecting detection to traffic-steering decisions.","feed_headline":"Autoencoder detects O-RAN anomalies with just 100 labels","feed_subtitle":"It pairs 80% balanced accuracy at minimal labeling with real-time, interpretable alerts for open radio access networks.","key_machinery":"The central object is the SS-DeepCAE loss, which combines mean-squared reconstruction error, a contractive term $\\lambda_c \\| \\partial E(x_i;\\theta_e)/\\partial x_i \\|_F^2$ that penalizes how much the encoder's latent representation changes under small input perturbations, and a supervised cross-entropy term $\\alpha_i \\ell_{\\text{cro}}(y_i, \\hat{y}_i)$ applied only to labeled samples. The contractive penalty encourages smooth, robust latent features that separate normal from anomalous traffic. The supporting object is fastSHAP-C, a learned explainer that approximates Shapley-value attributions through one forward pass, augmented by a Confidence Score and an Error Metric that quantify how faithfully the explanation reconstructs the model output.","core_discovery":"On its own terms, the paper establishes XAInomaly as an explainable, resource-efficient anomaly detector for open radio access networks. The SS-DeepCAE learns compressed representations of normal traffic by minimizing a combined loss of reconstruction error, a contractive penalty on the encoder Jacobian, and a cross-entropy term applied to the few labeled samples; at inference, reconstruction error and latent norm feed an anomaly score. Experiments on a 10,000-sample O-RAN dataset with roughly 25% anomalies report UAR of 80.17 ± 0.6 at 100 labeled samples, rising to 91.17 ± 0.6 with all labels, beating DeepAE, SLA-VAE, LSTM-autoencoder, and vanilla autoencoder at every label count. fastSHAP-C, a learned explainer that estimates Shapley values with a single forward pass, reduces runtime to 22,145 ms from 33,627 ms on DeepAE (a 34% improvement) and from 320,400 ms for kernelSHAP (about 93% faster), while identifying RSRP, RSRQ, and RSSINR as the dominant features behind anomaly flags.","pith_inferences":["Editorial inference: the reported 34% runtime gain is measured between fastSHAP-C on SS-DeepCAE and fastSHAP-C on DeepAE, so part of the gain may come from the smaller backbone rather than the explainer itself; a controlled benchmark holding the backbone fixed would separate the two effects.","Editorial inference: the contractive penalty is generic, so the same architecture should transfer to other O-RAN telemetry streams (for example, fronthaul or core interfaces) provided the unlabeled training pool actually reflects normal behavior.","Editorial inference: if RSRP, RSRQ, and RSSINR remain the top attributions across cells and time, operators could pre-filter monitoring to these KPIs and shrink the feature space, potentially lowering labeling and compute costs further.","Editorial inference: the semi-supervised gains depend on a mostly-normal unlabeled pool, so an operator applying the method to a heavily contaminated stream should first validate that premise empirically."],"forward_implications":["Deploying SS-DeepCAE as an anomaly-detection xApp on the Near-RT RIC lets operators monitor O-RAN traffic with a model whose accuracy plateaus after roughly 800 labeled samples, easing the labeling burden.","fastSHAP-C can produce per-sample explanations in real time, so the Traffic Steering xApp can factor explanation confidence into handover decisions.","The finding that RSRP, RSRQ, and RSSINR dominate the attributions gives operators a concrete, model-driven prioritization of RF KPIs for monitoring and troubleshooting.","The low parameter count of 8,180 trainable parameters supports deployment in resource-constrained edge or micro-data-center settings.","Because UAR stays high as labels increase from 100 to the full dataset, the model offers a scalable path from label-scarce to label-rich operations."],"supporting_citations":[{"why":"Supplies the contractive autoencoder penalty that is the core regularizer in SS-DeepCAE.","marker":"[34]"},{"why":"The DeepAE baseline and prior model whose limitations motivate SS-DeepCAE and whose runtime comparison produces the 34% figure.","marker":"[21]"},{"why":"The fastSHAP method that fastSHAP-C extends with confidence and error metrics.","marker":"[45]"},{"why":"The kernelSHAP baseline and Shapley-value computation framework used for comparison.","marker":"[44]"},{"why":"The LSTM-autoencoder anomaly-detection baseline benchmarked in the UAR table.","marker":"[42]"},{"why":"The semi-supervised VAE baseline benchmarked against SS-DeepCAE.","marker":"[43]"},{"why":"Introduces fastSHAP-C in the prior xURLLC work and is cited as the method's origin.","marker":"[14]"}],"fun_headline_variants":["O-RAN anomaly detection: 80% UAR with only 100 labels","Explainable autoencoder spots O-RAN anomalies with 100 labels","Semi-supervised deep CAE detects O-RAN faults, 93% faster explanations","XAInomaly: interpretable O-RAN anomaly detection, 100-label accuracy","O-RAN anomaly flags 93% faster with fastSHAP-C explainer"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The semi-supervised gains rest on the assumption that the cross-entropy term in the stated loss actually reaches the network through a classification head on the latent representation, and that the unlabeled training samples are overwhelmingly normal traffic.","fun_headline_variants_meta":{"raw":{"variants":["O-RAN anomaly detection: 80% UAR with only 100 labels","Explainable autoencoder spots O-RAN anomalies with 100 labels","Semi-supervised deep CAE detects O-RAN faults, 93% faster explanations","XAInomaly: interpretable O-RAN anomaly detection, 100-label accuracy","O-RAN anomaly flags 93% faster with fastSHAP-C explainer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000939,"raw_usage":{"total_tokens":4042,"prompt_tokens":999,"completion_tokens":3043,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":2937}},"tokens_in":615,"tokens_out":3043,"duration_ms":20543,"temperature":1.0,"reasoning_tokens":2937,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T22:19:20.995485+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce the training with the architecture given in Table IV; if no classification head exists, the supervised term in Eq. (12) cannot be evaluated, and removing it would show whether the 80.17 UAR figure actually depends on labeled data. Separately, construct a variant of the unlabeled training pool whose anomaly share approaches 25% (matching the full dataset) and measure UAR; if it collapses toward the vanilla-autoencoder level, the mostly-normal-unlabeled premise is load-bearing.","supporting_citations":[{"cited_title":"Contractive auto-encoders: explicit invariance during feature extraction,","cited_arxiv_id":null,"evidence_quote":"Supplies the contractive autoencoder penalty that is the core regularizer in SS-DeepCAE."},{"cited_title":"Deep Autoencoder Design for RF Anomaly Detection in 5G O-RAN Near-RT RIC via xApps,","cited_arxiv_id":null,"evidence_quote":"The DeepAE baseline and prior model whose limitations motivate SS-DeepCAE and whose runtime comparison produces the 34% figure."},{"cited_title":"FastSHAP: Real-Time Shapley Value Estimation,","cited_arxiv_id":null,"evidence_quote":"The fastSHAP method that fastSHAP-C extends with confidence and error metrics."},{"cited_title":"A Unified Approach to Inter- preting Model Predictions,","cited_arxiv_id":null,"evidence_quote":"The kernelSHAP baseline and Shapley-value computation framework used for comparison."},{"cited_title":"Anomaly Detection Using Autoencoders in High Performance Computing Systems,","cited_arxiv_id":null,"evidence_quote":"The LSTM-autoencoder anomaly-detection baseline benchmarked in the UAR table."},{"cited_title":"A Semi-Supervised V AE Based Active Anomaly Detection Framework in Multivariate Time Series for Online Systems,","cited_arxiv_id":null,"evidence_quote":"The semi-supervised VAE baseline benchmarked against SS-DeepCAE."},{"cited_title":"XAInomaly: Explainable, Inter- pretable and Trustworthy AI for xURLLC in 6G Open-RAN,","cited_arxiv_id":null,"evidence_quote":"Introduces fastSHAP-C in the prior xURLLC work and is cited as the method's origin."}],"review_version":1}