{"id":"cf12ee86-f71d-4e86-bb74-f3d32faf4d8f","arxiv_id":"2506.07532","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper combines a cross-modal neural network for jamming-type recognition with a DQN-based anti-jamming waveform selector, reporting 95.45% accuracy on three simulated jamming classes and faster reward convergence than SARSA.","lead":"A radar anti-jamming system that first classifies the type of jammer from radar echoes and then uses a deep reinforcement learning network to choose a frequency-hopping waveform to escape it. The two-stage design is tested only in simulation, and the reported gains are small and partly inconsistent.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fig. 12's DQN-vs-SARSA comparison is uncontrolled: reward-parity and state-information parity are unspecified, and the only reported metric is the hand-crafted reward that DQN is trained to maximize.","rationale":"The reader's CONDITIONAL verdict is apt. My concern overlaps with but goes beyond its stated weakest assumption: the reward functions are unvalidated proxies, and the DQN-vs-SARSA comparison is also missing state/action parity and an external performance measure. This is load-bearing because the only quantitative evidence for the anti-jamming half of the abstract is Fig. 12. The proposed test would settle it by isolating the algorithm/fusion contribution from the information-access confound and by checking reward alignment with real radar metrics. I do not reject the paper: recognition results, despite their reporting problems, are plausible and the framework is coherent. The issues are addressable, so the conditional verdict stands. I would, however, also ask that the authors correct the abstract/Table VII percentage mismatch and the confusion-matrix arithmetic; those are independent reporting inconsistencies, not the basis of my main concern.","tokens_in":14781,"tokens_out":8807,"duration_ms":110495,"concrete_test":"Obtain or re-implement the DQN and SARSA agents with identical state vectors (including the FDJR prediction and passive-radar parameters), identical action spaces, and identical rewards, and evaluate both learned policies on an external anti-jamming metric that is not part of Eqs. (12)-(14), e.g., detection probability or monopulse angle RMSE over many jamming episodes. Additionally, train a DQN variant with a modified reward scale (e.g., different 30/-100 thresholds and hop cost c) and check whether the selected waveforms and external metric remain stable. If DQN no longer beats SARSA when states are matched, or if the external metric does not track the hand-crafted reward, the Fig. 12 claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Fig. 12 is the only evidence for the anti-jamming half of the central claim (faster convergence and greater stability than SARSA). The plotted quantity is the average reward defined by the hand-crafted functions in Section III.B.3, Eqs. (12)-(14), i.e., the same objective the DQN is trained to maximize. No independent outcome metric (detection probability, post-suppression SINR, monopulse RMSE, or false-target rejection rate) is reported for the learned policies; Fig. 15 shows one angle-estimation scenario but not a DQN-vs-SARSA comparison. Consequently, the convergence/stability result is a statement about a proxy, and the constants 30, -100, and the hop-cost c in Eqs. (12)-(13) are never calibrated against a real or high-fidelity radar model. In addition, the paper does not state whether SARSA was given the same state vector; if SARSA lacks the FDJR jamming-type output and passive-radar parameters, the comparison in Fig. 12 confounds algorithm advantage with privileged-information access. The central claim would fail if either (a) SARSA with identical state/action/reward closes the gap, or (b) policies that maximize Eqs. (12)-(14) do not improve an external anti-jamming metric.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a unified radar anti-jamming framework that combines jamming-type recognition with DQN-based anti-jamming waveform decision-making. Recognition uses STFT and SPWVD time-frequency representations fused through an H-ResNet and an attention-based cross-modal fusion module with time-domain features. The recognized jamming type, together with passive-radar parameter estimates, is fed into a DQN that selects frequency-hopping or cover-pulse waveforms. The authors evaluate recognition accuracy against SVM, VGG-16, and 2D-CNN baselines and compare the decision network against SARSA in terms of reward convergence and stability.","tokens_in":14966,"tokens_out":3411,"duration_ms":41733,"significance":"If the claims are established, the proposed closed-loop recognition-plus-decision framework is a useful step toward practical cognitive radar electronic warfare, and the combination of cross-modal feature fusion with active-passive radar information is a plausible architecture. The recognition experiment reports a large margin in overall accuracy (Table VII), and the paper addresses a relevant gap by linking recognition output to downstream waveform selection. However, the current evidence is incomplete: the recognition comparisons lack statistical robustness, the reported improvement numbers are internally inconsistent, and the decision-making comparison relies entirely on the hand-crafted reward that the DQN is trained to maximize, with no external anti-jamming performance metric. The paper does not provide code, trained models, or a reproducible experimental protocol, which limits verification.","major_comments":[{"comment":"The dataset splitting is stated inconsistently: Section IV.B says the proportion of datasets used for training, testing, and validation is 0.3:0.6:0.1, while Section IV.D says Table VII reports results with 80% training samples. Because all recognition claims depend on the exact training/evaluation protocol, the authors must specify the precise split, the number of samples per class, and whether the 0.3:0.6:0.1 split or an 80% split was used for the reported metrics.","section":"Section IV.B and Section IV.D"},{"comment":"The abstract reports improvements of 1.41%, 2.5%, and 14.51% in overall accuracy compared to SVM, VGG-16, and 2D-CNN, but Table VII and Section IV.D report improvements of 3.75%, 1.76%, and 2.32%, respectively. The two sets of numbers are inconsistent and the abstract's ordering and magnitudes appear to be erroneous. This must be corrected, because the abstract is the primary statement of the paper's contribution.","section":"Abstract and Table VII"},{"comment":"The central decision-making claim of faster convergence and greater stability than SARSA is supported only by the average-reward curves in Fig. 12, where the reward is computed from the hand-crafted functions in Eqs. (12)-(14). Since the DQN is trained to maximize exactly these rewards, convergence on this metric is not an independent measure of anti-jamming success. The paper should report external metrics such as post-suppression SINR, detection probability, false-target rejection rate, or monopulse angle-estimation error for the learned policies, and should calibrate the thresholds and costs (30, -100, and c) against a physical or high-fidelity radar model. In addition, the comparison with SARSA does not state whether SARSA was given the same state vector (including the FDJR jamming-type output and passive-radar parameters) and the same reward function; without this parity, Fig. 12 may reflect a difference in privileged information rather than an algorithmic advantage.","section":"Section III.B.3, Eqs. (12)-(14), and Fig. 12"},{"comment":"The recognition results are reported as a single confusion matrix and single-point metrics with no error bars, no multiple random seeds, and no statistical significance testing. Given the class distribution visible in Fig. 10 (e.g., RDFT has far more test samples than AJ or RFT), the claimed margins of 1.76-3.75% over baselines could be within run-to-run variability. The authors should provide mean and standard deviation over repeated runs or k-fold cross-validation, and report per-class support counts.","section":"Section IV.D and Fig. 10"}],"minor_comments":[{"comment":"The text refers to 'Table 4' when it should refer to Table VII, and contains the typo 'butter than' instead of 'better than'.","section":"Section IV.D"},{"comment":"Tables IV and V are presented after the text says 'the backbone network structure ... is shown in Table II and Table III'; the table numbering should be aligned.","section":"Section IV.B and Tables IV-V"},{"comment":"The abbreviation 'DRFTJ' is used inconsistently; the paper elsewhere uses RDFTJ. Please standardize all jamming-type acronyms.","section":"Section I, Contribution 2"},{"comment":"The subfigure captions duplicate the labels (a)-(c) for both SARSA and the proposed algorithm, which makes it hard to identify which curve corresponds to which method; use distinct labels.","section":"Fig. 12"},{"comment":"The displayed definition of the ReLU activation is garbled; please rewrite it in standard mathematical notation.","section":"Section III.A, Eq. (9)"},{"comment":"The H-ResNet module is cited to DenseNet [36], but the text does not explain the relationship; either cite a direct source or clarify how the module is derived from DenseNet.","section":"Section III.A and Reference [36]"}],"recommendation":"major_revision","confidential_remarks":"The paper has a reasonable scope for eess.SP, but the internal inconsistencies between the abstract and the tables, the ambiguous dataset split, and the lack of an external metric for the DQN comparison should be resolved before publication. The manuscript would benefit from a thorough proofreading pass and a clearer experimental protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. This is an integration paper with a coherent architecture, but the evidence for the headline DQN claims is not there yet. It deserves a serious referee, not a desk reject, but it needs major revision.\n\nThe genuinely new piece is the closed loop: jamming recognition feeding an anti-jamming waveform selector that also gets passive-radar parameter estimates. Recognition via STFT/SPWVD and attention-based cross-modal fusion is assembled from known parts, and the per-jamming-type reward shaping is a reasonable idea. The instinct to include an external outcome measure (monopulse angle estimation in Fig. 15) is right, even though it covers only one scenario.\n\nThe soft spots are substantial and mostly fixable. The abstract's accuracy improvements (1.41%, 2.5%, 14.51%) don't match Table VII (3.75%, 1.76%, 2.32%). The dataset split is 0.3:0.6:0.1 in Section IV.B and 80% training in IV.D. No error bars or significance tests anywhere. The DQN vs SARSA comparison in Fig. 12 is uncontrolled: we never learn whether SARSA received the same state vector, particularly the recognition output and passive-radar parameters, and the reported metric is the hand-crafted reward the DQN is trained to maximize, not detection probability, post-suppression SINR, or false-target rate. The stress-test note on this is accurate. The reward constants (30, -100, hop cost c) are uncalibrated; the paper claims 'complex environments' but tests only a small set of simulated conditions. No code, data, or hyperparameters were released.\n\nIf I were advising the authors, I'd ask for corrected numbers, a matched SARSA baseline with identical state/action/reward, an external anti-jamming metric, and released code/data. The recognition results are plausible and the integration idea is useful, so I'd engage with it.\n\nWho's it for: radar EW and cognitive radar folks who want a reference architecture for recognition-to-countermeasure coupling. I'd cite it if I wrote about unified anti-jamming designs, but only after the fixes.\n\nRecommendation: send it to peer review. It's not a reject on sight, but a strong reviewer should not wave it through.","headline":"A plausible recognition-to-countermeasure integration whose DQN half is over-claimed and under-tested.","tokens_in":15613,"tokens_out":3028,"would_cite":false,"duration_ms":35575,"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 argues that radar jamming type recognition and anti-jamming waveform selection should be designed as one closed-loop framework, and reports that its fused recognizer reaches 95.45% overall accuracy while its DQN converges…","keywords":["radar anti-jamming","jamming recognition","cross-modal fusion","Deep Q-Network (DQN)","time-frequency analysis","active-passive radar fusion","DRFM jamming","frequency-agile radar"],"falsifier":"Run the trained DQN against a hardware-in-the-loop or high-fidelity simulated jammer and measure post-suppression detection probability, SINR, and angle-estimation error across episodes; if the policies that earn high simulated reward do not systematically improve those measured outcomes over fixed or random frequency-hopping policies, the reported convergence is an artifact of the hand-coded rewards rather than a real anti-jamming capability.","tokens_in":14476,"feed_emoji":"📡","tokens_out":7278,"duration_ms":73368,"temperature":0.7,"pith_summary":"This paper argues that radar jamming type recognition and anti-jamming waveform selection are two halves of one problem and should be built as a single closed loop, not as separate modules. It proposes a framework in which STFT and SPWVD time-frequency images are fused by a hierarchical ResNet, combined with time-domain features through an attention-based cross-modal module, and classified into three jamming types: aiming jamming, range false target jamming, and range dense false target jamming. The recognition result then feeds a Deep Q-Network that picks suppression waveforms, with hand-crafted reward functions tailored to each jamming type. In the paper's simulations the unified recognizer reaches 95.45% overall accuracy, beating SVM, VGG-16, and 2D-CNN, and the DQN converges around episode 600 with smaller fluctuations than SARSA. A reader should care because this is the integration pattern a practical cognitive radar would need under digital radio frequency memory (DRFM) jamming.","feed_headline":"Fused jamming recognition hits 95.45% accuracy in anti-jam radar","feed_subtitle":"Unified DQN framework also converges faster and steadier than SARSA across three DRFM jamming types.","key_machinery":"The two load-bearing mechanisms are the cross-modal fusion module and the DQN strategy network with its reward functions. The fusion module lets two time-frequency representations, STFT and SPWVD, processed by a hierarchical ResNet and concatenated, be reweighted against time-domain features through a self-attention block: query, key, and value vectors are computed, and a global-average-pooling plus Sigmoid step produces channel weights that rescale the original features. That reweighted representation is what carries the recognition accuracy. On the decision side, the DQN selects discrete frequency-agile waveforms, uses an online and target Q-network with experience replay, and is driven by the paper's three hand-coded rewards: an SINR-threshold reward with values 30 and -100 for AJ, a frequency-hopping reward with a hopping cost for RFTJ, and a cover-pulse timing reward for RDFTJ. Coupling happens because the jamming type from the recognizer and the jamming parameters from passive radar enter the DQN state vector.","core_discovery":"The paper's central claim is that combining multi-modal jamming recognition with deep-reinforcement-learning waveform selection in one framework improves both stages at once. The recognition branch extracts two time-frequency representations, concatenates deep features from a hierarchical ResNet, and fuses them with time-domain features using a multi-head attention mechanism before a three-class Softmax classifier. The decision branch models anti-jamming as a DQN over discrete waveform actions, with an online network, a target network, and experience replay; its state includes the recognition result plus passive-radar parameter estimates. According to the reported results, this integrated design raises overall recognition accuracy to 95.45% compared with 91.70%, 93.69%, and 93.13% for SVM, VGG-16, and 2D-CNN, and its reward curves converge near episode 600 with visibly less oscillation than SARSA across AJ, RFTJ, and RDFTJ scenarios.","pith_inferences":["An extension the paper leaves implicit is feeding softmax probabilities rather than the hard jamming label into the DQN state; that would let the decision network gracefully handle recognition uncertainty instead of committing to a possibly wrong class.","The reward-calibration question is the natural next experiment: the hand-set thresholds, the hopping cost, and the cover-pulse window should be checked against detection probability or receiver-operating-characteristic curves in a hardware-in-the-loop test, since maximizing simulated rewards may not track physical detection performance.","Ablating the attention weighting in the cross-modal module would isolate how much of the 95.45% overall accuracy comes from cross-modal fusion rather than from the larger network capacity of the hierarchical ResNet.","Because all simulations use a fixed carrier frequency, bandwidth, SNR, and JNR range, testing with parameter drift across episodes would reveal whether the learned DQN policy generalizes or is tuned to the simulated grid."],"forward_implications":["A radar that adopts this framework can identify AJ, RFTJ, or RDFTJ and switch to the matching suppression waveform in the same operational loop, instead of treating recognition and countermeasure selection as separate steps.","The reported convergence around episode 600 with small reward fluctuation suggests the DQN policy can reach a stable anti-jamming behavior within a few hundred simulated jamming episodes.","Because passive-radar broadband estimates supply the carrier frequency and bandwidth that narrowband active radar measures poorly, the framework's decisions are conditioned on more precise jamming parameters than active-only systems.","Across training-set sizes, the multi-modal recognizer reportedly keeps higher overall accuracy, recall, precision, and F1 than single-modal baselines, which is the regime where real collected jamming data are scarce.","In the constructed application scenario with a target at 5 degrees and a jammer at 2 degrees, the selected suppression waveforms restore usable monopulse angle estimation, indicating the anti-jamming choices preserve tracking information."],"supporting_citations":[{"why":"This reference supplies the time-frequency image representation that the recognition branch starts from.","marker":"[13]"},{"why":"This reference supplies the SPWVD method that produces the second time-frequency modality in the fusion.","marker":"[35]"},{"why":"This reference supplies the hierarchical ResNet backbone used to extract deep features from the time-frequency images.","marker":"[36]"},{"why":"This reference supplies the AJ reward-function form that the paper adapts in Eq. (12).","marker":"[41]"},{"why":"This reference supplies the DQN and epsilon-greedy exploration mechanisms the decision network builds on.","marker":"[22]"},{"why":"This reference supplies the DQN-based cognitive-radar anti-jamming method the paper extends.","marker":"[26]"},{"why":"This reference is the SVM baseline whose accuracy the unified recognizer must beat.","marker":"[43]"},{"why":"This reference is the VGG16 baseline used in the recognition comparison.","marker":"[44]"},{"why":"This reference is the 2D-CNN baseline used in the recognition comparison.","marker":"[45]"}],"fun_headline_variants":["Unified jamming defense: cross-modal fusion plus DQN hits 95.45%","Cross-modal fusion boosts radar jamming recognition to 95.45%","DQN with cross-modal fusion improves anti-jam radar accuracy","Unified radar anti-jamming: fusion + DQN beats SVM and CNN","95.45% jamming recognition via cross-modal attention fusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the hand-coded reward functions in Eqs. (12)-(14) correctly encode what it means for a real radar to defeat AJ, RFTJ, and RDFTJ, so that maximizing them in simulation transfers to physical suppression performance.","fun_headline_variants_meta":{"raw":{"variants":["Unified jamming defense: cross-modal fusion plus DQN hits 95.45%","Cross-modal fusion boosts radar jamming recognition to 95.45%","DQN with cross-modal fusion improves anti-jam radar accuracy","Unified radar anti-jamming: fusion + DQN beats SVM and CNN","95.45% jamming recognition via cross-modal attention fusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000709,"raw_usage":{"total_tokens":3246,"prompt_tokens":1051,"completion_tokens":2195,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":667,"completion_tokens_details":{"reasoning_tokens":2096}},"tokens_in":667,"tokens_out":2195,"duration_ms":15142,"temperature":1.0,"reasoning_tokens":2096,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:31:24.865435+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained DQN against a hardware-in-the-loop or high-fidelity simulated jammer and measure post-suppression detection probability, SINR, and angle-estimation error across episodes; if the policies that earn high simulated reward do not systematically improve those measured outcomes over fixed or random frequency-hopping policies, the reported convergence is an artifact of the hand-coded rewards rather than a real anti-jamming capability.","supporting_citations":[{"cited_title":"Radar Signal Recognition Based on Deep Feature Fusion of Multiple Time -frequency Images[C]//2023 International Conference on Wireless Communications and Signal Processing (WCSP)","cited_arxiv_id":null,"evidence_quote":"This reference supplies the time-frequency image representation that the recognition branch starts from."},{"cited_title":"Time -frequency DOA estimate algorithm based on SPWVD[C]//2005 IEEE International Symposium on Microwave, Antenna, Propagation and EMC Technologies for Wireless Communications","cited_arxiv_id":null,"evidence_quote":"This reference supplies the SPWVD method that produces the second time-frequency modality in the fusion."},{"cited_title":"Reinforcement learning for compressed - sensing based frequency agile radar in the presence of active interference[J]","cited_arxiv_id":null,"evidence_quote":"This reference supplies the AJ reward-function form that the paper adapts in Eq. (12)."},{"cited_title":"Cooperative Frequency Scheduling for Netted Agile Radars via Decentralized Q -Learning[J]","cited_arxiv_id":null,"evidence_quote":"This reference supplies the DQN and epsilon-greedy exploration mechanisms the decision network builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This reference supplies the DQN-based cognitive-radar anti-jamming method the paper extends."},{"cited_title":"Distorted SAR target recognition with virtual SVM and AP -HOG feature[C]//IET International Radar Conference (IET IRC 2020)","cited_arxiv_id":null,"evidence_quote":"This reference is the SVM baseline whose accuracy the unified recognizer must beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This reference is the VGG16 baseline used in the recognition comparison."},{"cited_title":"Deep fusion for radar jamming signal classification based on CNN[J]","cited_arxiv_id":null,"evidence_quote":"This reference is the 2D-CNN baseline used in the recognition comparison."}],"review_version":1}