{"id":"2d6098c6-67ca-4141-a54e-2c0bc400e4e0","arxiv_id":"2501.11074","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A GCN-DQN framework that weights routing rewards by attack detection results is shown to produce attack-avoiding paths on CICIoT2022, outperforming an attack-unaware MPNN-DQN baseline.","lead":"This paper proposes a security-aware GNN-DRL framework that uses graph neural networks to detect attacks and deep reinforcement learning to route traffic around them. The authors demonstrate the idea on an IoT traffic dataset, showing higher rewards than a baseline that ignores attacks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported 'superiority' is not established: the baseline optimizes a different reward, and the evaluation metric is the same engineered reward, so the comparison is potentially circular.","rationale":"In good faith, the paper presents a plausible GNN-DRL framework and a case study intended to illustrate resilience against encrypted-traffic attacks. However, the central claim of 'demonstrated effectiveness and superiority' rests on the experimental comparison in Section IV.C, and that comparison is flawed in a way that is independent of the GCN detector's accuracy. The reward function R is both the training objective and the reported evaluation metric, while the baseline MPNN-DQN is described as not accounting for attacks—i.e., it is not trained on the security term w. Comparing average R under these conditions is not a controlled test of resilience; it can largely reflect the addition of a penalty term the baseline never optimizes. The reader's weakest_assumption focused on GCN detection accuracy, which is a valid concern, but even perfect detection would not salvage the superiority claim without a fair baseline and a metric that directly measures attack avoidance (e.g., number of attack nodes on selected paths). Therefore the most load-bearing concern is the experimental design itself, not merely the unvalidated detector. Since this reinforces the reader's rejection while identifying a more fundamental issue, the verdict should remain unchanged.","tokens_in":11556,"tokens_out":5759,"duration_ms":60576,"concrete_test":"Retrain the MPNN-DQN baseline using the identical reward function R (including the security weights w) and re-run the Fig. 6 experiment. If the baseline then matches or exceeds the proposed method's average R, the claimed superiority is an artifact of baseline objective mismatch.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV.A defines R = -Σ_{n∈p1}(w_n+c_n) - Σ_{m∈p2}(w_m+c_m) - α·O, with security weights w derived (unspecified) from attack traffic ratio. Section IV.B states the baseline MPNN-DQN 'considers only communication cost without accounting for attacks.' Yet the results in Fig. 6 report average R, i.e., the same training reward that includes the w-term. Thus the proposed method is compared on the objective it was explicitly trained to optimize, while the baseline is evaluated on an objective it never saw. The red curve's higher R may simply reflect the added penalty term, not better network resilience. No detection accuracy, no ablation, and no direct metric of attack avoidance (e.g., fraction of attack nodes on chosen paths) is reported. The claim that the framework 'can converge quickly and generate the optimal routing policy' and 'demonstrated effectiveness and superiority' is therefore unsupported, independent of whether the GCN detector is accurate. If the detector itself is unvalidated, the reward signal may also be misaligned, further undermining the resilience claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework that combines graph neural networks (GNNs) with deep reinforcement learning (DRL) to improve network resilience against attacks. After a tutorial-style overview of GNN-DRL methods and their applications in communication networks, the authors present a case study in which a GCN-based attack detector (adapted from TFE-GNN) is integrated with a DQN agent that selects routing paths. The reward function penalizes communication cost, path overlap, and node-level security weights derived from attack traffic. Experiments are performed on two reconstructed topologies (20 and 50 nodes) from the CICIoT2022 dataset, and the proposed GCN-DQN is compared against an MPNN-DQN baseline. The paper claims that the proposed method 'can converge quickly and generate the optimal routing policy' and that the results 'demonstrated the effectiveness and superiority of our framework' (Section IV.C, Abstract).","tokens_in":11798,"tokens_out":4230,"duration_ms":43349,"significance":"If the empirical claims were fully supported, the paper would offer a useful integration of attack detection and defense-oriented routing in a GNN-DRL setting, and its tutorial component provides a structured overview of GNN-DRL for network security. However, the validation as reported is not conclusive. The comparison with the baseline is circular because both methods are evaluated on the same engineered reward that only the proposed method is trained to optimize, and the attack-detection component is never validated with any accuracy metric. The paper ships no code, no formal convergence or optimality analysis, and no statistical confidence measures. Consequently, the central contribution is only partially supported, and the 'superiority' claim should be substantially moderated or backed by additional evidence. The tutorial and taxonomy sections are of interest to readers new to the area, but the case study needs rework before the paper can be accepted.","major_comments":[{"comment":"The main empirical comparison is circular. The proposed GCN-DQN is compared with MPNN-DQN using the average reward R, which is the SAME reward the proposed method is trained to optimize. Section IV.B states that the baseline MPNN-DQN 'considers only communication cost without accounting for attacks,' so the baseline never optimizes the security-weight terms w_n and w_m in R. The higher R of the red curve is therefore expected by construction: it reflects the added penalty term, not better network resilience. To substantiate superiority, the authors should report a direct attack-avoidance metric (e.g., the fraction of selected paths that contain attack nodes) or evaluate both methods under an identical reward function.","section":"Section IV.B, IV.C, Fig. 6"},{"comment":"The security weight w is defined as an integer in [1,200] assigned 'by considering the attack traffic ratio,' but no formula, threshold, or algorithm is provided. This omission is load-bearing because w is the only attack-related signal in the reward R = -Σ_{n∈p1}(w_n+c_n) - Σ_{m∈p2}(w_m+c_m) - α·O. Furthermore, the GCN detector designed following TFE-GNN [2] is never validated: no accuracy, precision, recall, or F1 scores are reported on the CICIoT2022 traffic. Without evidence that w correctly identifies attack nodes, the reward may be misaligned with true network security, and the resilience claim becomes unsupported.","section":"Section IV.A"},{"comment":"All results are reported as averages over ten experiments with no error bars, standard deviations, or statistical significance tests. The smoothed curves in Fig. 6 cannot support the statements 'can converge quickly' and 'generate the optimal routing policy,' especially because the 20-node red and green curves are numerically close after early episodes (e.g., -130.1 vs -156.5 for episodes 2500-5000, and -50.5 vs -50.3 for episodes 10000-12500). The authors should provide variance or confidence intervals and specify exactly how the table entries were computed from the ten runs.","section":"Section IV.C, Fig. 6"},{"comment":"The reconstruction of the two network topologies from CICIoT2022 is described too briefly for reproducibility. The paper does not say how the 20-node and 50-node graphs are derived from the dataset, how edges are assigned, how per-node traffic is 'kept consistent' with the dataset, or how attack labels in CICIoT2022 are mapped to graph nodes. Without these details, the reported reward values are difficult to interpret, and the case study cannot be independently reproduced.","section":"Section IV.B"}],"minor_comments":[{"comment":"The phrase 'conductive learning' should be 'transductive learning' when describing GCNs.","section":"Section II.A.2"},{"comment":"The values of the penalty factor α and the communication costs c are not specified beyond the range [1,200]; please state the actual choices used in the experiments.","section":"Section IV.A"},{"comment":"The action-space sizes 464 and 39,038 are mentioned but not explained; please specify how these counts are derived from the path sets for the two routing requirements.","section":"Section IV.A"},{"comment":"The time-consumption numbers (0.49s per 1000 packets, 0.031s and 0.257s per episode) lack context; please report the hardware/software environment and whether these are training or inference times.","section":"Section IV.C"},{"comment":"There is a typo: 'has became more and more popular' should be 'has become more and more popular.'","section":"Section II.B"},{"comment":"The table caption says 'the bolded values representing the optimal ones,' but the text does not explain which values are bolded or why; ensure the table formatting is consistent with the caption.","section":"Section IV.C"}],"recommendation":"major_revision","confidential_remarks":"The paper is essentially a survey/tutorial with a case study appended. The case study claims are not adequately supported, and the circular evaluation is the main obstacle. The authors could repair this by adding detection metrics, direct attack-avoidance metrics, error bars, and a fair baseline; such fixes are within the scope of a revision. However, the current abstract and conclusions overstate the evidence, so the editor may also consider whether an IEEE Network magazine contribution with a weaker 'illustrative study' framing would be more appropriate than the strong superiority claims made here."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The survey portion is genuinely useful. The taxonomy of GNN-DRL combinations (GCN/GraphSAGE/GAT with DQN/PPO/DDPG/SAC) is clearly organized, and the framework description in Section III is coherent. The specific idea of using GCN-based encrypted traffic detection to compute security-aware node weights, then feeding those into a DQN routing reward, is a legitimate extension that I haven't seen in the cited prior work. Credit where due: the paper does a decent job of positioning itself relative to existing GNN-DRL literature.\n\nThe soft spots are concentrated in the case study, and they are load-bearing. The comparison against MPNN-DQN is potentially circular. The reward function R includes the security weight w derived from attack traffic ratio, and both methods are evaluated on that same R. The baseline, by the paper's own description, optimizes only communication cost and never sees the w term. So the red curve's higher accumulated reward may simply reflect that the proposed method is being scored on an objective it was explicitly trained to optimize, while the baseline is scored on an objective it never saw. That does not establish better network resilience. The paper needs a direct metric like fraction of attack nodes on chosen paths, or a per-path security cost, evaluated independently of the training reward.\n\nSecond, the GCN detector is never validated. No detection accuracy, precision, or recall is reported, and the computation of w from attack traffic ratio is left unspecified. If the detector is inaccurate, the reward signal is misaligned and the whole resilience claim unravels. Third, ten runs with no error bars is thin; the smoothing window in Fig. 6 hides variance. Fourth, no code or data is shared, which makes it hard to check the reconstruction of the CICIoT2022 topologies.\n\nThese are not minor quibbles; they directly undermine the abstract's claim of demonstrated effectiveness and superiority. But the survey content and the framework proposal are not invalidated by the weak experiments. A reader interested in a compact overview of GNN-DRL for networks will get value from Sections II and III. A reader looking for evidence that this framework improves resilience should not rely on the case study as it stands.\n\nMy recommendation: major revision before it deserves to appear as a research contribution. The authors should either fix the baseline (train it on the same reward, or evaluate both on an attack-avoidance metric), report detection accuracy and error bars, and release code, or reframe the paper as a position/tutorial without empirical superiority claims. I would send it to peer review because the survey is competent and the idea is worth testing properly, but I would not cite the current empirical results.","headline":"The survey half of this paper is solid and worth reading; the case study, however, does not support the 'superiority' claim because the baseline is evaluated on a reward it never optimized and the attack detector is never validated.","tokens_in":12295,"tokens_out":1793,"would_cite":false,"duration_ms":20701,"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":"This paper claims that a GNN-DRL framework, combining GCN-based encrypted-traffic detection with a security-weighted DQN reward, produces routing policies that avoid attack nodes and thereby improve network resilience.","keywords":["network resilience","graph neural networks","deep reinforcement learning","encrypted traffic classification","routing optimization","IoT security","CICIoT2022","GCN-DQN"],"falsifier":"Run the GCN detector on held-out CICIoT2022 samples and measure its precision and recall at the operating point used in the case study; then recompute the final routing policies and check how many selected paths contain ground-truth attack nodes. If the detector misses most attacks, or if the chosen paths do not contain fewer labeled attack nodes than the MPNN-DQN baseline, the resilience claim fails.","tokens_in":11400,"feed_emoji":"🛡️","tokens_out":7618,"duration_ms":77445,"temperature":0.7,"pith_summary":"Deep reinforcement learning can optimize network routing, but standard DRL agents are blind to security: they keep using attacked nodes and their policies oscillate under attack. This paper argues that grafting a graph neural network onto the DRL agent closes that gap: a GCN reads the network topology and encrypted traffic, marks attack-prone nodes, and a DQN chooses paths that avoid them. The authors propose a three-phase framework (graph modeling, GNN-DRL decision-making, policy execution) and back it with a case study on CICIoT2022 IoT traffic, where their method converges faster and to a higher average reward than a GNN-DRL baseline that ignores attacks. The value, if the claim holds, is a single data-driven control loop that couples attack detection with defensive routing instead of treating them as separate problems.","feed_headline":"AI router learns to steer IoT traffic around attack nodes","feed_subtitle":"Case study on real encrypted IoT traffic reports faster convergence and higher reward than an attack-blind baseline.","key_machinery":"Two components carry the argument. The first is a two-layer GCN detector, designed following TFE-GNN, that classifies encrypted traffic and assigns each node a security weight $w$ from its attack traffic ratio; the second is a DQN whose reward $R = -\\sum_{n \\in p_1}(w_n + c_n) - \\sum_{m \\in p_2}(w_m + c_m) - \\alpha O$ couples those weights with communication costs $c$ and a penalty $\\alpha O$ for overlapping paths $p_1$ and $p_2$. The GCN gives the DRL agent a security-aware state representation, and the reward makes attack avoidance part of the optimization objective, so the learned policy routes around nodes the detector distrusts.","core_discovery":"On the paper's own terms, the central discovery is that security-weighted GNN-DRL can turn encrypted-traffic classification into an actionable routing policy. The framework models the network as a graph, uses a two-layer GCN following TFE-GNN to extract attack features from encrypted traffic, summarizes the environment state, and lets a DQN agent select pairs of routing paths under a reward that penalizes communication cost, node overlap, and attack-node usage. In 20-node and 50-node topologies reconstructed from CICIoT2022, the authors report that this GCN-DQN stabilizes at higher average rewards than MPNN-DQN, which oscillates and degrades when attacks are present. They interpret this as evidence that attack-aware GNN-DRL improves network resilience against flood and brute-force attacks.","pith_inferences":["The paper does not report detection accuracy for the GCN; a direct reader test would measure precision and recall on the CICIoT2022 labels and then count how many chosen paths contain ground-truth attack nodes.","A natural stress test is to vary the fraction of attack nodes in the reconstructed topology and check whether the reward gap versus MPNN-DQN grows with attack severity; the paper only uses one attack configuration.","Security weighting could also be applied to moving-target defense, where the DQN would randomize routes over a set of low-risk paths; the paper discusses defense broadly but does not instantiate that variant."],"forward_implications":["Attack-aware routing becomes a byproduct of the same learning loop: once the GCN identifies risky nodes, the DQN adjusts routes without hand-coded avoidance rules.","Reported per-episode times (0.031 s for the 20-node network, 0.257 s for the 50-node network) suggest the combined detector and router can operate at control-loop speeds rather than only offline.","Because the reward separates node cost and security weight, the framework can be extended to latency-, energy-, or throughput-aware routing by adding terms to the node cost.","The three-phase architecture gives a template for adding attack detection to other DRL network-control tasks such as virtual network embedding and resource allocation."],"supporting_citations":[{"why":"This GNN-DRL routing method supplies the routing-optimization setting and the practice of using a GNN to extract network features for the DRL agent; the paper positions its framework as an extension of this line.","marker":"[1]"},{"why":"TFE-GNN supplies the temporal-fusion GCN architecture that the paper adapts for encrypted-traffic attack detection, i.e., the source of the security weights in the reward.","marker":"[2]"},{"why":"CICIoT2022 supplies the real IoT encrypted-traffic data and attack scenarios from which the 20-node and 50-node test topologies are reconstructed.","marker":"[14]"},{"why":"MPNN-DQN is the GNN-DRL routing baseline whose reward curves and convergence behavior anchor the comparison that supports the superiority claim.","marker":"[15]"}],"fun_headline_variants":["GNN-DRL routes IoT traffic away from attack nodes","Security-aware GNN-DRL boosts network resilience","Graph-learning RL sidesteps network attacks in IoT","Attack-aware DRL reroutes around cyber threats","Resilient networks via GNN-powered deep RL"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything hinges on the GCN detector correctly identifying attack nodes, yet the paper reports no detection accuracy, precision, or recall; if the security weights do not track real attacks, the DQN optimizes a reward that looks security-aware without actually avoiding compromised nodes.","fun_headline_variants_meta":{"raw":{"variants":["GNN-DRL routes IoT traffic away from attack nodes","Security-aware GNN-DRL boosts network resilience","Graph-learning RL sidesteps network attacks in IoT","Attack-aware DRL reroutes around cyber threats","Resilient networks via GNN-powered deep RL"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1339,"prompt_tokens":913,"completion_tokens":426,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":351}},"tokens_in":529,"tokens_out":426,"duration_ms":5198,"temperature":1.0,"reasoning_tokens":351,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:39:30.283635+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the GCN detector on held-out CICIoT2022 samples and measure its precision and recall at the operating point used in the case study; then recompute the final routing policies and check how many selected paths contain ground-truth attack nodes. If the detector misses most attacks, or if the chosen paths do not contain fewer labeled attack nodes than the MPNN-DQN baseline, the resilience claim fails.","supporting_citations":[{"cited_title":"Towards the development of a realistic multi- dimensional iot profiling dataset,","cited_arxiv_id":null,"evidence_quote":"CICIoT2022 supplies the real IoT encrypted-traffic data and attack scenarios from which the 20-node and 50-node test topologies are reconstructed."},{"cited_title":"G-routing: Graph neural networks-based flexible online routing,","cited_arxiv_id":null,"evidence_quote":"This GNN-DRL routing method supplies the routing-optimization setting and the practice of using a GNN to extract network features for the DRL agent; the paper positions its framework as an extension of this line."},{"cited_title":"Tfe-gnn: A temporal fusion encoder using graph neural networks for fine-grained encrypted traffic classification,","cited_arxiv_id":null,"evidence_quote":"TFE-GNN supplies the temporal-fusion GCN architecture that the paper adapts for encrypted-traffic attack detection, i.e., the source of the security weights in the reward."},{"cited_title":"Deep reinforcement learning meets graph neural networks: Exploring a routing optimization use case,","cited_arxiv_id":null,"evidence_quote":"MPNN-DQN is the GNN-DRL routing baseline whose reward curves and convergence behavior anchor the comparison that supports the superiority claim."}],"review_version":1}