{"id":"16b83e80-daa2-48c5-a47e-b3f601c8957c","arxiv_id":"2505.14027","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A dual-module intrusion detection model using GAN oversampling plus cost-sensitive attention CNN achieves 84.55% five-class and 91.09% binary accuracy on NSL-KDD.","lead":"This paper combines two deep-learning tricks: a self-attention GAN that creates synthetic network traffic to balance rare attack classes, and a CNN with cost-sensitive attention that classifies traffic. On the NSL-KDD benchmark it reports about 84.6% five-class accuracy and 91.1% binary accuracy, a small step above several earlier models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported accuracy/F1 are dominated by Normal/DoS classes; rare-class (R2L/U2R) performance is never reported, so the claimed imbalance-handling advantage is not established.","rationale":"The reader identified the undisclosed cost-sensitive weights as the weakest assumption, and I agree that is a serious reproducibility gap. However, the more load-bearing concern is that the reported global metrics cannot substantiate the paper's central imbalance-handling claim even if the weights were disclosed. Because the weighted F1 in Eq. (7) is dominated by Normal and DoS classes, high overall numbers are compatible with very poor rare-class detection. This matters especially because SC-CGAN and CSL are specifically designed to target the rare classes. The paper contains no per-class results for KDDTest+, so the contribution of the two novel modules to the classes they are intended to help is untested. I do not see internal inconsistency: the SC-CGAN/CNN numbers in Table 4 match the CNN-Only row of Table 6, which is reassuring. But the absence of class-wise evaluation leaves the headline result ambiguous. Requiring per-class metrics and the exact cost-weight vector is the appropriate condition for accepting the empirical claim.","tokens_in":12603,"tokens_out":9922,"duration_ms":99514,"concrete_test":"Ask the authors to release the exact cost-weight vector w_i from Eq. (5) and to report per-class precision, recall, F1, and a confusion matrix for CSCA-CNN, CNN-Only, w/o CSL, and w/o CAM on KDDTest+. Then independently rerun the Table 6 ablation with those weights and compute the class-wise metrics. If U2R and R2L F1 are near zero, or if neither module improves these classes over CNN-Only, the 'unequivocally demonstrated effectiveness' claim and the imbalance-handling contribution are not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"CSAGC-IDS's stated purpose is to improve detection of rare attacks under class imbalance, but the paper reports only global accuracy and the weighted F1 defined in Eq. (7). On KDDTest+, the two rarest classes are U2R (200 of 22,544) and R2L (2,754 of 22,544); together they contribute just 13.1% of the weighted F1. Meanwhile, SC-CGAN produces 67,291 of the 67,343 U2R training samples and 66,348 of the 67,343 R2L samples (Table 2), and CSCA-CNN's cost-sensitive weights in Eq. (5) are supposed to prioritize these classes. A model that classifies Normal/DoS/Probe well but fails on U2R/R2L can still reach the reported weighted F1 of about 0.845 and accuracy of about 0.846. Without per-class precision/recall/F1 on KDDTest+, neither the ablation in Table 6 nor the comparisons in Table 10 demonstrate that the proposed modules improve rare-class detection; the headline numbers could reflect majority-class performance only. This concern is compounded by the undeclared cost weights, but the absence of per-class metrics is the more fundamental missing evidence.","agreement_with_reader":"partial"},"referee_report":null,"author_rebuttal":null,"desk_editor":null,"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-07T15:41:33.112064+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}