{"id":"04e657d7-0c57-476b-ade8-9a76270ac9c9","arxiv_id":"2507.17843","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A proposed 5G closed loop that estimates user latency at the UPF and classifies games with NWDAF-hosted ML, but the components are evaluated separately and the runtime feature gap is unresolved.","lead":"This paper proposes adding a packet-timing monitor to the User Plane Function of a 5G core and feeding those measurements to the network's analytics function to classify online games and manage latency. The authors report good results for the timing monitor and for game classification separately, but the closed-loop system itself is not implemented or tested.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed control loop is undermined by the UPF metric: L = tout − tin is a packet residence time inside the UPF, not UE-perceived latency, so the R² = 0.980 result does not support the latency-aware architecture.","rationale":"Good-faith reading: the paper contributes two artifacts—a UPF filter estimating a per-TEID latency and an offline game classifier—and assembles them into a claimed closed-loop architecture. What must be true for the central claim is (i) the filter measures a latency that is meaningful for UE QoE, and (ii) the classifier can be fed with data that the runtime path can supply. The first condition is not established by the paper's own equations and validation. L = tout − tin at the UPF is a residence time; the paper's only comparison is against a synthetic sinusoidal baseline that plausibly injects the same shape into both series. The second condition is also unmet: Section 4's classifier uses 16 features (source, destination, latitude, longitude, RTT) while Section 3's runtime supplies only per-TEID time-shift values. I chose the latency metric as the most load-bearing because even fixing the classifier cannot rescue a loop whose input quantity is mislabeled. The reader's REJECT verdict remains appropriate; no change needed.","tokens_in":6144,"tokens_out":7042,"duration_ms":77672,"concrete_test":"Deploy the described UPF filter in a testbed with independent E2E latency ground truth: a UE and a game server, with controlled additional delay inserted at different points (RAN, N3, N6, server). Vary the delay at each point separately and compare the filter's L = tout − tin against the ground-truth RTT. If L tracks only delay inserted inside the UPF and not delay inserted at the UE, RAN, or server, then the metric is not UE-perceived latency and the closed-loop claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the UPF user-space filter estimates UE-perceived latency on the N3 interface and feeds NWDAF/SMF to make the control plane latency-aware. The load-bearing step is the equation L = tout − tin, where tin/tout are packet arrival and departure timestamps at the UPF. This quantity is the time a packet spends transiting the UPF (or at most the sum of such residence times for request and response). It does not include RAN scheduling, air-interface transmission, core network transport, or server processing, so it is not the UE-perceived RTT that the baseline in Section 5 claims to match. The validation induces a sinusoidal load on the packet recipient and compares it to the filter's estimate, but nothing in the method ties the recipient-side delay to the UPF timestamps. The resulting R² = 0.980 can be an artifact of both series following the injected sine, rather than the filter measuring user latency. Without a ground-truth E2E latency measurement that is independent of the injected signal and of the UPF's own timestamps, the foundational metric of the closed loop is unvalidated. The classifier feature mismatch identified by the reader is a second serious gap, but it is downstream of this metric: even a perfectly trained classifier cannot make the loop latency-aware if the input latency is not user-perceived latency.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a closed-loop 5G edge-gaming architecture in which a user-space filter at the UPF N3 interface computes per-TEID \"latency\" from packet timestamps (L = tout - tin), reports the values to NWDAF, and NWDAF uses ML models to classify games (League of Legends, Teamfight Tactics, Valorant). Reported results include normalized MSE 0.019, MAE 0.085, and R² 0.980 for the latency estimate, and CatBoost accuracy 0.9483 with F1-score 0.9477 for game classification. The authors conclude that the scheme enables latency-aware SMF interventions and SLA management in beyond-5G edge gaming slices.","tokens_in":6388,"tokens_out":4881,"duration_ms":53335,"significance":"If the proposed metric genuinely captured UE-perceived latency and the classifier could operate on features measurable in the live UPF/NWDAF path, the architecture would be a practical, non-intrusive addition to 5G core analytics for edge gaming. The paper has some worthwhile components: it uses a public dataset, compares five ML algorithms with standard metrics, and the UPF filter concept is simple to replicate. However, the significance is conditional on two preconditions that the manuscript does not establish: that the UPF time-shift measurement is a valid proxy for UE-perceived RTT, and that the offline-trained classifier can run in the proposed runtime. The current validation addresses neither precondition end-to-end.","major_comments":[{"comment":"The load-bearing equation L = tout - tin is the residence time of a packet inside the UPF, not the UE-perceived RTT. It excludes RAN scheduling, air-interface transmission, core transport, and server processing. In Section 5, the baseline is \"UE-perceived RTT\" generated by a sinusoidal synthetic load applied to the packet recipient, but nothing in the method ties the recipient-side delay to the UPF timestamps. Because the injected load and the measured series share the same sinusoidal period, the reported R² = 0.980 could simply reflect that both follow the injected signal rather than that the filter measures user latency. An independent ground-truth E2E latency measurement at the UE, or a decomposition showing which components of E2E latency the UPF transit time captures, is required before the latency-aware control-loop claim is supported.","section":"Section 3, Eq. (1)-(2); Section 5, Table 1"},{"comment":"The game classifier is trained on a public dataset with 16 features, including source, destination, latitude, longitude, and RTT, but the runtime architecture in Section 3 supplies NWDAF only with per-TEID time-shift latency from the UPF. The paper never demonstrates that the other dataset features are available or measurable in the live free5GC/NWDAF path. Consequently, the high accuracy in Table 2 and the ROC/PR curves in Figures 6 and 7 are offline benchmarks on dataset features, not evidence that the proposed NWDAF classifier can classify games from the UPF-provided latency feature in the intended closed loop.","section":"Section 4 vs. Section 3"},{"comment":"No integrated closed-loop experiment is reported. The latency estimation is evaluated against a synthetic load, and the classifier is evaluated offline on a public dataset, but the two components are never connected: no experiment feeds actual UPF-generated per-TEID latency into the classifier, and no demonstration shows that NWDAF analytics reach the SMF or that any resulting intervention improves SLA compliance or QoE. The concluding claim of a latency-aware 5G control plane is therefore unsupported by the presented measurements.","section":"Section 5"},{"comment":"The Random Forest row is internally inconsistent: it reports Recall 0.0981 alongside Precision 0.9478 and F1-Score 0.9479, which is arithmetically impossible (the harmonic mean of 0.9478 and 0.0981 is about 0.178). Since the text states that Random Forest performed similarly to CatBoost, this inconsistency directly affects the stated model-comparison conclusion and requires correction or verification.","section":"Table 2"}],"minor_comments":[{"comment":"The sentence beginning \"The curves in Figure 7 exhibit strong class distinction...\" refers to AUC values, but Figure 7 shows precision-recall curves; the ROC AUC values appear in Figure 6, so the figure references need to be corrected.","section":"Section 5, Figures 6-7"},{"comment":"\"Games Distribuition by Class\" contains a typo; it should read \"Distribution.\"","section":"Figure 4 caption"},{"comment":"The difference between normalized and original MAPE is not explained; please define the normalization applied to MSE, MAE, and MAPE, and state the units of the original metrics.","section":"Section 5, Table 1"},{"comment":"The experimental section does not describe the train/test split or cross-validation scheme used for the classification experiments, so it is unclear whether the 10-run averages in Table 2 are over the same test folds or different random splits.","section":"Section 4"},{"comment":"The Hassanein et al. dataset reference is incomplete (no publication venue, URL, or version), which makes it difficult for readers to reproduce the classification experiments.","section":"References"}],"recommendation":"reject","confidential_remarks":"For the editor: the central problem is not presentation but the validity of the proposed latency metric and the transferability of the classifier to the runtime path. Fixing these would require a new end-to-end validation campaign with true UE-side ground truth and a redesign of the classifier input schema to match what the UPF can actually measure; that is beyond a routine revision. The manuscript's current claims are therefore not supported by its evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe one thing to know: the paper's central control-loop claim does not survive contact with its own equation. They define L = tout − tin at the UPF N3 interface and call it \"UE-perceived latency,\" then report R² = 0.980 against a baseline. But that quantity is just packet residence time inside the UPF. It excludes RAN scheduling, air-interface transmission, and server processing—so it is not the end-to-end RTT a gamer experiences. The validation makes it worse: they inject a sinusoidal load on the packet recipient and compare the filter's timestamps to that same injected signal. With both series following the sine, the high R² can be an artifact. The stress-test note is right; the metric is unvalidated for its intended use.\n\nWhat's genuinely there: a real testbed (free5GC on Kubernetes), a user-space filter that extracts per-TEID timings, and a standard but thorough classifier evaluation on a public gaming dataset. Five algorithms, confidence intervals, ROC and PR curves. That is work you can point to, and the NWDAF-UPF integration idea is a reasonable extension of 3GPP Release 16 analytics. The writing is clear and the related work is appropriately placed.\n\nThe soft spots are serious and multiple. First, the classifier was trained on 16 features—source, destination, latitude, longitude, RTT—but the runtime only supplies per-TEID time-shift latency. The paper never shows the rest are available in the live UPF/NWDAF path, so the game classification cannot run in the proposed loop as built. Second, Table 2 contains an impossible entry: Random Forest recall 0.0981 with F1-Score 0.9479. Those two numbers cannot coexist, which makes me doubt the rest of the table's accuracy. Third, no code or dataset is shipped, so the \"empirical\" results are not independently checkable. The closed-loop SMF/PCF actions are only described, not implemented.\n\nWho is this for? A reader interested in 5G core analytics might read it to see the architecture idea, but they should not cite the latency numbers. As submitted, the central claim is not supported. A serious revision could rename the metric as \"UPF residence time,\" add a real end-to-end ground truth independent of the injected load, fix the feature gap, and correct the table. But that is a big rework.\n\nRecommendation: I would not accept this for peer review as-is; desk reject seems appropriate, though you could invite a resubmission if the authors rework the metric and validation. It is not a paper to spend referee time on in its current form.","headline":"The paper's control-loop claim collapses on its own metric: L = tout − tin at the UPF is packet residence time, not UE-perceived latency, so the R² = 0.980 result does not support the architecture.","tokens_in":6938,"tokens_out":5077,"would_cite":false,"duration_ms":48894,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims a 5G core can estimate player latency from packet timestamps and classify the game being played from that signal.","keywords":["5G core","user plane function","network data analytics function","time-shift latency estimation","edge gaming","network slicing","game classification","TEID"],"falsifier":"Run the proposed N3 time-shift monitor on live LOL, TFT, and Valorant sessions, feed NWDAF only the per-TEID latency values, and retrain and evaluate CatBoost on that single signal; if accuracy falls well below the reported 0.9483, the closed loop as described cannot perform the claimed classification.","tokens_in":5885,"feed_emoji":"🎮","tokens_out":10708,"duration_ms":92112,"temperature":0.7,"pith_summary":"The paper proposes making the 5G control plane latency-aware for mobile edge gaming by instrumenting the User Plane Function (UPF) with a user-space packet filter. The filter reads timestamps of GTP packets crossing the N3 interface, computes per-flow latency as $L = t_{\\mathrm{out}} - t_{\\mathrm{in}}$, and labels each measurement with the Tunnel Endpoint Identifier (TEID) of the user session. These per-TEID latency values go to the Network Data Analytics Function (NWDAF), where pretrained machine-learning models classify the game being played and can notify the Session Management Function (SMF) to adjust resources. In experiments the estimate reaches normalized MSE 0.019 and $R^2$ 0.980 against a synthetic load, and a CatBoost classifier trained on the dataset distinguishes League of Legends, Teamfight Tactics, and Valorant with 0.9483 accuracy. If correct, operators get a non-intrusive closed loop for SLA enforcement and QoE protection in gaming slices.","feed_headline":"A packet filter estimates gamer latency in 5G with R² 0.98","feed_subtitle":"UPF timestamps feed a CatBoost classifier that tells LOL, TFT, and Valorant apart, enabling latency-aware slicing.","key_machinery":"The load-bearing object is the time-shift latency monitor: a user-space filter on the UPF N3 interface that captures GTP packets, extracts the 32-bit TEID, and computes per-tunnel latency from $L = t_{\\mathrm{out}} - t_{\\mathrm{in}}$, with a bidirectional variant $L_{\\mathrm{total}} = (t^{\\mathrm{request}}_{\\mathrm{out}} - t^{\\mathrm{request}}_{\\mathrm{in}}) + (t^{\\mathrm{response}}_{\\mathrm{out}} - t^{\\mathrm{response}}_{\\mathrm{in}})$. It converts raw packet timestamps into a per-UE signal without injecting probe traffic. The second piece is the analytics stage inside NWDAF: pretrained ML models, of which CatBoost, a gradient-boosted decision-tree model, performs best, mapping monitored traffic features to a game class so the SMF can decide whether an intervention is needed.","core_discovery":"The central claim is that a closed loop between an instrumented UPF and NWDAF can turn passive packet observation into actionable user-latency awareness. The paper argues that the time-shift between a packet's arrival at and departure from the N3 interface, accumulated per TEID, is a faithful proxy for UE-perceived latency, with normalized MSE 0.019, MAE 0.085, and $R^2$ 0.980 on a synthetic sinusoidal load of 1–600 ms. It further claims that an ML model embedded in NWDAF can classify the active game class from traffic features, with CatBoost reaching 0.9483 accuracy and per-class AUC above 0.99 for LOL, TFT, and Valorant. On that basis the paper asserts that SMF can react to degradation by reallocating resources, adjusting routing policy, or adapting network parameters, making the control plane latency-aware for edge gaming.","pith_inferences":["This paper's own text leaves a gap between the Section 4 classifier, trained on 16 features, and the Section 3 runtime, which feeds NWDAF only per-TEID latency; a reader should treat the 0.9483 accuracy as an upper-bound demonstration, not as proof of the closed loop in its described form.","A direct test would retrain CatBoost on the UPF-derived latency signal alone; if the accuracy drop is small, the loop becomes credible without relying on unevaluated features.","The same per-TEID time-shift monitor could transfer to other latency-sensitive GTP services, such as AR/VR streaming or industrial remote control, with the game-class label replaced by a service-class label."],"forward_implications":["An operator can estimate UE-perceived latency from passive UPF observations with normalized MSE 0.019 and $R^2$ 0.980, so no client-side probes or game-specific instrumentation are required.","NWDAF can distinguish LOL, TFT, and Valorant with 0.9483 accuracy and per-class AUC above 0.99, making game class a usable input for slice policy decisions.","The SMF gains a trigger for corrective actions such as resource reallocation, routing-policy changes, or dynamic parameter adaptation when a gaming session degrades.","The same architecture connects non-intrusive latency measurement to SLA compliance monitoring for gaming slices in beyond-5G networks."],"supporting_citations":[{"why":"Supplies the 69,395-instance, 16-feature dataset with LOL, TFT, and Valorant classes used for the ML classification experiments.","marker":"[Hassanein et al. 2025]"},{"why":"Frames the GPU and edge-AI capacity gains that motivate latency and QoE demands for online gaming in 5G.","marker":"[Shankar 2024]"},{"why":"Provides prior 5G cloud-gaming QoE estimation work that this paper extends by adding NWDAF and UPF in a closed loop.","marker":"[Soares et al. 2024]"},{"why":"Supplies previous QoE prediction for gaming video streaming in 5G/O-RAN that motivates AI-driven QoS control in the radio and edge path.","marker":"[Kougioumtzidis et al. 2024]"},{"why":"Provides a 4G/5G network and service metrics dataset used to ground end-to-end performance estimation for video and cloud gaming.","marker":"[Baena et al. 2023]"}],"fun_headline_variants":["UPF timestamps estimate gamer latency with R² 0.98","CatBoost and UPF time-shifts classify games and track latency","5G slicing adapts to games via UPF-NWDAF closed loop","Predict gamer latency from N3 time-shifts in 5G","UPF packet timing drives latency-aware game slicing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the per-TEID time-shift latency delivered by the UPF monitor carries enough information for the game classifier, yet the classifier is trained and evaluated on a 16-feature dataset whose source, destination, latitude, longitude, and RTT features are never shown to be available in the proposed live UPF-to-NWDAF path.","fun_headline_variants_meta":{"raw":{"variants":["UPF timestamps estimate gamer latency with R² 0.98","CatBoost and UPF time-shifts classify games and track latency","5G slicing adapts to games via UPF-NWDAF closed loop","Predict gamer latency from N3 time-shifts in 5G","UPF packet timing drives latency-aware game slicing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000638,"raw_usage":{"total_tokens":2894,"prompt_tokens":857,"completion_tokens":2037,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":1944}},"tokens_in":473,"tokens_out":2037,"duration_ms":14025,"temperature":1.0,"reasoning_tokens":1944,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:39:01.968859+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the proposed N3 time-shift monitor on live LOL, TFT, and Valorant sessions, feed NWDAF only the per-TEID latency values, and retrain and evaluate CatBoost on that single signal; if accuracy falls well below the reported 0.9483, the closed loop as described cannot perform the claimed classification.","supporting_citations":[{"cited_title":"R., and Shirmohammadi, S","cited_arxiv_id":null,"evidence_quote":"Supplies the 69,395-instance, 16-feature dataset with LOL, TFT, and Valorant classes used for the ML classification experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Frames the GPU and edge-AI capacity gains that motivate latency and QoE demands for online gaming in 5G."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides prior 5G cloud-gaming QoE estimation work that this paper extends by adding NWDAF and UPF in a closed loop."},{"cited_title":"K., Lazaridis, P","cited_arxiv_id":null,"evidence_quote":"Supplies previous QoE prediction for gaming video streaming in 5G/O-RAN that motivates AI-driven QoS control in the radio and edge path."},{"cited_title":"S., Camacho, L., Barco, R., and Fortes, S","cited_arxiv_id":null,"evidence_quote":"Provides a 4G/5G network and service metrics dataset used to ground end-to-end performance estimation for video and cloud gaming."}],"review_version":1}