{"id":"9186bf5b-dd01-40d0-8c11-ceb36a2b7761","arxiv_id":"2608.12870","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A gradient-based meta-learning optimizer that jointly tunes antenna positions, beamforming, and artificial noise improves secrecy rate in movable-antenna secure ISAC simulations compared with six baselines.","lead":"This paper combines movable antennas with secure integrated sensing and communication, and uses a gradient-based meta-learning optimizer to choose antenna positions, beamforming, and artificial noise. Simulations suggest this learned optimizer beats several classical and learning baselines in secrecy rate while keeping sensing beams intact.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central superiority claim rests on 10 channel realizations with no error bars or significance testing; reported margins may be within run-to-run variation.","rationale":"The reader identified Assumption 2 in Proposition 3 as the weakest assumption. That concern is legitimate, and the paper itself explicitly says Assumption 2 is a sufficient condition 'rather than to claim that the Adam-trained network globally guarantees such a property'. However, Proposition 3 is a conditional theoretical statement and does not carry the paper's central empirical claim; even if Assumption 2 is unverified, GML could still outperform baselines in practice. The more load-bearing weakness is the statistical basis of the superiority claim itself: Na = 10 channel realizations with no error bars or significance tests. In non-convex learned optimization, different random seeds and channel draws can produce large performance spreads, so small reported margins such as 4.83% and 6.67% may not be meaningful. The lack of released code and under-specified baselines compounds this by preventing independent verification of whether the comparison is fair. This concern supports, rather than overturns, the reader's CONDITIONAL verdict, so no verdict change is needed. The proposed concrete test—rerunning with more realizations and reporting paired confidence intervals—would settle whether the strongest claim actually lands.","tokens_in":24229,"tokens_out":3239,"duration_ms":36937,"concrete_test":"Obtain the authors' code or reproduce the comparison with Na = 100 or 1000 independent channel realizations for the key settings of Figs. 4 and 9 (PT sweep with M = 7, K = 2; theta_E sweep with PT = 10 dBm, M = 7). Report mean secrecy rate with 95% bootstrap confidence intervals and paired differences across identical channel realizations for GML versus PGA and GML versus ML. If the confidence interval for any paired difference includes zero at an operating point where the paper claims superiority over the 'entire angular range' or 'significantly outperform', the headline claim should be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim is empirical: GML 'achieves the highest secrecy rate over the entire angular range' and 'significantly outperform[s]' baselines across BS power, antenna count, CEE, and path loss. The evidence for this is curves averaged over Na = 10 independent channel realizations (Table II, Section IV), with no error bars, confidence intervals, or significance tests reported in Figs. 4-10. At a key operating point, the claimed advantage is only 4.83% over PGA and 6.67% over ML at PT = 20 dBm. In a non-convex, randomly initialized learned-optimizer setting, run-to-run variation across channel draws and initialization seeds can easily be of this size; 10 realizations are too few to establish uniform superiority. Additionally, the baselines are under-specified: ML, PGA, FPA, and RA reuse GML for W and N while AO and MVPSO use CVX-based subproblems, but no convergence criteria, particle counts, or implementation details are given, so part of the reported gap could reflect implementation choices rather than algorithmic advantage. This is not an internal inconsistency, but it means the central claim is not yet statistically evidenced.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a gradient-based meta-learning (GML) algorithm for a movable-antenna (MA) enabled secure integrated sensing and communication (ISAC) system, jointly optimizing antenna positions, transmit beamforming, and artificial noise to maximize the secrecy rate under a transmit power budget and a sensing beampattern constraint. The method uses a softmax-based continuous parameterization of antenna positions, a convex subproblem to generate a desired beampattern covariance matrix, and three neural networks that map gradients of the secrecy rate to variable update steps in a three-loop alternating scheme. The theoretical part proves that the Jacobian of the position mapping has full row rank, uses that to argue equivalence between optimization in the unconstrained and position domains, and provides a conditional convergence theorem under a sufficient descent assumption on the learned updates. Numerical results compare GML with six baselines across transmit power, antenna count, channel estimation error, eavesdropper angle, and path loss, and also show beampattern evolution.","tokens_in":24432,"tokens_out":6028,"duration_ms":60845,"significance":"If the empirical claims hold, the paper would introduce a viable model-driven, pre-training-free solver for a challenging non-convex MA-ISAC secure beamforming problem, with a differentiable surrogate for the min-distance antenna constraint that may be useful beyond this specific setting. Strengths include the clean mathematical derivation of the Jacobian rank and the self-contained first-order convergence analysis under explicit assumptions, as well as the falsifiable prediction that GML outperforms five MA baselines and a fixed-antenna baseline over a range of system parameters. However, the central superiority claim rests on statistical evidence that is currently too thin to support the strong wording used in the paper, and the convergence theory is not fully aligned with the implemented algorithm.","major_comments":[{"comment":"All simulation curves are averaged over Na = 10 independent channel realizations, and no error bars, confidence intervals, or significance tests are reported. At PT = 20 dBm, the claimed advantage over PGA and ML is only 4.83% and 6.67%, respectively; in a non-convex, randomly initialized learned-optimizer setting these margins can easily be within run-to-run variation. Ten realizations are insufficient to support the statement in Fig. 9 that GML 'achieves the highest secrecy rate over the entire angular range' or the repeated claim of 'significantly outperform[ing]' the baselines. Please report per-realization statistics, e.g., box plots or confidence bands, specify the number of random seeds for network initialization, and provide significance tests for the key comparisons.","section":"Section IV, Figs. 4-10, Table II"},{"comment":"The convergence proof of Proposition 3 analyzes a simultaneous recursion β_{i+1} = β_i + d_i with d_i = M_H(∇R_S(β_i)), where H is fixed, and imposes Assumption 2 on the descent of the full loss L(β) in Eq. (42). However, Algorithm 1 updates u, W, and N sequentially in three separate inner loops, the input to each network is the gradient of R_S (not of L), and the meta-parameters change across epochs. Thus the assumptions of Proposition 3 do not match the algorithm as implemented: the sufficient descent condition (49) is not verified for a network that takes ∇R_S as input, and convergence of the meta-parameter updates is only supported empirically by Fig. 3. The proof is internally consistent as a standalone statement, but it does not establish convergence of the proposed GML algorithm. Please either adapt the analysis to the alternating block structure, or explicitly state that Proposition 3 is a sufficient condition for a hypothetical fixed-update inner loop and not a convergence guarantee for Algorithm 1.","section":"Section III-C2, Eqs. (47)-(55), and Algorithm 1"},{"comment":"The sensing constraint (16e) is replaced by ||R_X - R_d||_F^2 ≤ ξ in Eq. (25), where R_d is obtained from the convex subproblem (24). No proof is given that satisfaction of (25) implies satisfaction of the original constraint (16e) for the chosen ξ and ε, and the relationship between ξ and ε is not characterized. The text at this point also incorrectly refers to (16d) instead of (16e), which obscures the fact that the minimum-distance constraint is unrelated to the sensing beampattern. Since the sensing quality constraint is one of the three primary constraints of P0, this missing link undermines the statement in Section IV-B that GML 'consistently achieves the goal of maximizing the system secrecy rate while ensuring sensing quality.' Please provide a quantitative relation between ξ and ε, or at least report the achieved MSEbp for the configurations in Figs. 11-15.","section":"Section III-A2, Eqs. (24)-(26)"},{"comment":"The softmax parameterization covers only the interior of the feasible position set; boundary configurations such as an antenna at the panel edge x_m = ±L or a tight active constraint |x_s - x_c| = d_min cannot be attained, only approached. The equivalence in (22) is stated 'for interior feasible points', and stationarity is characterized only for the interior. If the optimizer of P0 lies on the boundary, GML may be unable to represent it, so the claim of achieving the maximum secrecy rate is not justified for boundary optima. Please discuss whether the simulation settings produce interior optima, or provide a boundary-handling extension.","section":"Section III-A1, Propositions 1-2 and paragraph after Eq. (21)"},{"comment":"The baselines are under-specified: ML, PGA, FPA, and RA reuse GML for W and N, while AO and MVPSO use CVX-based subproblems, but no convergence criteria, particle counts, trust-region parameters, or stopping tolerances are reported. Since PGA's performance depends on its step-size schedule and MVPSO on the particle budget, part of the observed gap could reflect implementation choices rather than algorithmic advantage. Please provide full implementation details for all baselines, ideally with code release, so that the comparisons are reproducible and fair.","section":"Section IV, baseline descriptions"}],"minor_comments":[{"comment":"There are numerous typos and grammatical errors, including 'An movable antenna' in the abstract, 'efficiency', 'sufficient', 'parament', 'denots', and 'maximun'. These should be corrected and the manuscript should be carefully proofread.","section":"Throughout"},{"comment":"In the sentence 'which means (16d) can be transformed as follows', the constraint number should be (16e), not (16d).","section":"Section III-A2"},{"comment":"The figure shows 'std. band' for the meta-parameter variations, but the text does not describe how the standard deviation is computed or over how many realizations or seeds; please clarify.","section":"Fig. 3"},{"comment":"Table II lists the sensing threshold as ξ = 0.5, but Fig. 13 discusses ξ values of 0.8, 1.0, and 1.2; please clarify which ξ is used in each simulation and avoid the apparent inconsistency.","section":"Table II and Fig. 13"},{"comment":"The secrecy rate is defined with a [·]^+ operator, but the optimization and the loss function use R_S directly without the positive part; if R_S can be negative, this may produce misleading gradients. Please clarify how the [·]^+ is handled in the algorithm.","section":"Section II-C1, Eq. (13)"}],"recommendation":"major_revision","confidential_remarks":"The core modelling and the differentiable mapping for antenna positions are solid, and the paper addresses an interesting application. The main reservations are that the empirical evidence for the headline performance claim is statistically thin (Na = 10, no error bars), the convergence analysis is not aligned with the implemented algorithm, and the sensing constraint transformation is not rigorously justified. These are fixable with additional experiments, corrected theory, and clearer implementation details, so I recommend major revision rather than rejection. I would encourage the editor to require the authors to specify baseline implementations precisely and, if possible, to release code."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — quick read of arXiv:2608.12870. The paper is a competent application of gradient-based meta-learning to MA-enabled secure ISAC secrecy-rate maximization. What is actually new is the combination: softmax spacing mapping for the MA position constraints, penalty handling for power and sensing constraints, and three small MLPs that take gradients as inputs and output update steps, with meta-learning over epochs. The lemmas about the Jacobian of the softmax-to-position map are correct for interior feasible points, and the convergence proposition is honestly labeled as resting on an unverified sufficient condition. The paper does not oversell the theory. It does claim empirical superiority, and that is where the soft spots are.\n\nThe main weakness is the evidence for 'GML achieves the highest secrecy rate over the entire angular range.' All curves are averaged over 10 channel realizations, with no error bars and no significance tests. At PT = 20 dBm the margin over PGA is 4.83%, which is well within plausible run-to-run variation for nonconvex, randomly initialized learned optimizers. Baseline implementations are under-specified: ML/PGA/FPA/RA reuse GML for W and N, while AO/MVPSO use CVX subproblems, but no convergence criteria, particle counts, or CVX details are given. Penalty weights λ1 and λ2 are not reported. So the central performance claim is not yet statistically evidenced. That is addressable: release code with all hyperparameters, add error bars, and compare with stronger baselines from the cited literature.\n\nThe theoretical part is fine but thin: Proposition 3 is a standard descent lemma under Assumption 2, which the authors explicitly say is not guaranteed by Adam. So it is a conditional convergence result, not a guarantee. The complexity analysis is plausible.\n\nNo circular reasoning. Self-citations [24] and [30] are background only. The paper is honest about its own limitations, which counts for something.\n\nWho is this for? Researchers working on MA-enabled secure ISAC or on learned optimizers for nonconvex wireless problems. It deserves a serious referee — the idea is coherent, the math is internally consistent, and the weaknesses are fixable. If I were the editor, I would send it out, with the request that reviewers push for code/data and stronger statistics.\n\nRecommendation: engage with it, conditionally. Not a desk reject, not an accept as is.","headline":"Competent combination of gradient-based meta-learning with softmax spacing mapping for MA-enabled secure ISAC; the performance claims need stronger statistics before the superiority result can be trusted.","tokens_in":24987,"tokens_out":1713,"would_cite":true,"duration_ms":17136,"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 gradient-based meta learning algorithm can maximize the secrecy rate of a movable-antenna integrated sensing and communication system by jointly learning antenna positions, transmit beamforming, and artificial…","keywords":["movable antenna","secure ISAC","antenna positioning","transmit beamforming","artificial noise","gradient-based meta learning","secrecy rate","beampattern design"],"falsifier":"Run GML on a fixed channel and record, at each inner-loop iterate, the inner product $\\nabla L(\\beta_i)^T d_i$ between the loss gradient and the network-chosen step. If any non-terminal iterate has that inner product positive, or if $\\|d_i\\|$ exceeds any uniform $c_2\\|\\nabla L(\\beta_i)\\|$, then Assumption 2 fails for that run and Proposition 3's convergence proof does not apply, so the reported convergence would be an empirical observation rather than a guaranteed one. Alternatively, for a small system such as $M=3$, $K=2$, enumerate near-optimal antenna positions and beamformers on a fine grid and check whether GML's final secrecy rate falls below the best grid point; a persistent gap would indicate the method is not finding the true optimum.","tokens_in":1930,"feed_emoji":"📡","tokens_out":5245,"duration_ms":106113,"temperature":0.7,"pith_summary":"The paper addresses a security bottleneck of integrated sensing and communications (ISAC): the same waveform that senses can be intercepted. It proposes letting the base station's antennas physically move within a linear region, and claims that moving them, together with beamforming and artificial noise, can be jointly optimized to maximize the system secrecy rate while preserving a sensing beampattern. Because the resulting optimization is non-convex and the variables are tightly coupled, the paper introduces a gradient-based meta learning (GML) algorithm: one small neural network per variable, fed with the gradient of the objective, outputs that variable's update step, and the networks' parameters are themselves updated by a meta-learning loop. The paper argues this needs no pre-training phase on external data, and reports simulations in which GML consistently achieves the highest secrecy rate across transmit power, antenna count, channel-estimation error, path loss, and eavesdropper angle while satisfying the sensing constraint.","feed_headline":"Gradient-meta learning beats six baselines in secure MA-ISAC","feed_subtitle":"It tunes antenna positions, beamforming, and artificial noise from gradients, with no training-data distribution to fit.","key_machinery":"The working mechanism is a per-variable meta-network update rule: for each optimization variable (antenna position parameters u, beamforming matrix W, and noise matrix N), a small multilayer perceptron receives the gradient of the loss with respect to that variable and outputs the update increment. Feasibility is enforced structurally: antenna positions come from a softmax over $M+1$ gap ratios followed by cumulative summation, so the panel bounds and minimum antenna spacing hold automatically; the noise covariance is parameterized as $N N^H$, guaranteeing positive semidefiniteness; and the sensing beampattern constraint is converted into a convex subproblem for a desired covariance $R_d$, with the mismatch $\\|R_X - R_d\\|_F^2$ entering the global loss as a penalty. The meta-learning loop updates the network parameters with Adam so that the learned update rule adapts to the current channel realization rather than to a fixed training distribution.","core_discovery":"The central claim is that the formulated secrecy-rate maximization problem P0 is tractable by the proposed GML procedure and, under the tested conditions, yields the best secrecy-rate performance among the compared algorithms. Concretely, the paper claims that feeding the instantaneous gradient of the secrecy-rate loss into per-variable neural networks, and meta-updating those networks on the current scenario, lets the algorithm jointly position the movable antennas, design the beamforming matrix, and shape the artificial noise better than fixed-position arrays, random arrays, projected gradient ascent, plain meta learning, alternating convex approximations, or particle swarm search. It also claims that the softmax reparameterization of antenna positions preserves the feasible set in an optimization sense (Propositions 1 and 2), and that under a smoothness assumption plus a sufficient descent condition on the learned step, the inner loop converges to a first-order stationary point (Proposition 3).","pith_inferences":["An implicit consequence is that the same gradient-in, step-out recipe could be applied to other constrained non-convex wireless designs, such as RIS phase shifts, fluid-antenna port selection, or power allocation, whenever a differentiable reparameterization of the feasible set exists; the paper does not test this, but nothing in the mechanism is ISAC-specific.","The softmax gap-ratio mapping suggests a general way to handle order and spacing constraints: by allocating a one-dimensional panel's slack space through a simplex, the minimum-distance constraint is satisfied by construction, so the same construction could be reused for any sorted-placement problem.","Because the convergence guarantee is conditional on Assumption 2, which the paper itself does not claim to verify globally, a natural testable extension is to monitor $\\nabla L(\\beta_i)^T d_i$ during training and measure what fraction of inner-loop updates actually satisfy the descent inequality, then relate that fraction to the observed convergence rate.","Since the feasible region's boundary is never reached exactly by the softmax or $N N^H$ parameterization, deployment scenarios requiring antennas to sit at exact panel endpoints or exactly $d_{\\min}$ apart would need a post-processing rounding step, which the paper does not address."],"forward_implications":["If GML performs as reported, movable antennas give a measurable secrecy-rate gain over fixed or random arrays in secure ISAC, not only at nominal settings but also under channel estimation error and severe path loss.","The reparameterization result means that optimizing over unconstrained gap variables can stand in for optimizing antenna positions: every interior placement is reachable, and every descent direction in position space is mirrored by a descent direction in the gap space.","The convergence theorem gives a practical stopping rule: once the squared gradient norm sum is small, the inner loop is near a first-order stationary point of the penalized loss, so the reported secrecy rates correspond to stationary solutions rather than arbitrary iterates.","GML can balance sensing and secrecy through the threshold $\\xi$, because tightening $\\xi$ reshapes the transmit covariance toward the desired sensing beampattern while GML still retains the highest secrecy rate among baselines.","As a no-pre-training optimizer, GML can be applied to a new user or eavesdropper geometry or antenna-panel size without retraining on a dataset, as supported by the reported beampattern convergence across different user distributions."],"supporting_citations":[{"why":"Introduces movable antennas as an antenna paradigm and supplies the spatial degrees of freedom that the secure ISAC design exploits.","marker":"[14]"},{"why":"Supplies the model-driven meta-learning alternating-minimization idea that the GML algorithm adapts to this problem.","marker":"[37]"},{"why":"Provides the convex sensing-beampattern formulation used to construct the desired covariance R_d and the sensing penalty.","marker":"[41]"},{"why":"Establishes the secure radar-communication and jamming context that motivates artificial-noise design against malicious targets.","marker":"[5]"},{"why":"Represents the joint MA-ISAC optimization literature against which the paper positions its contribution.","marker":"[28]"},{"why":"Supplies the projected-gradient-ascent baseline for antenna position optimization.","marker":"[42]"},{"why":"Supplies the multi-velocity particle swarm optimization baseline for antenna-position search.","marker":"[43]"},{"why":"Supplies the channel estimation error model and a gradient-based meta learning application in communications.","marker":"[40]"}],"fun_headline_variants":["Gradient meta-learning beats six baselines in secure MA-ISAC","GML jointly tunes antennas, beamforming, noise for top secrecy","Meta-learned gradients optimize secure ISAC without pretraining","Gradient-driven meta-learning wins on MA-ISAC secrecy rate"],"cache_read_input_tokens":27136,"weakest_assumption_plain":"The convergence proof rests on the assumption that, at every inner-loop step, the neural network's update moves the loss downhill with a strength comparable to the gradient norm; the paper explicitly states that this is a sufficient condition rather than a property the trained network is known to guarantee.","fun_headline_variants_meta":{"raw":{"variants":["Gradient meta-learning beats six baselines in secure MA-ISAC","GML jointly tunes antennas, beamforming, noise for top secrecy","Meta-learned gradients optimize secure ISAC without pretraining","Gradient-driven meta-learning wins on MA-ISAC secrecy rate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000206,"raw_usage":{"total_tokens":1400,"prompt_tokens":951,"completion_tokens":449,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":375}},"tokens_in":567,"tokens_out":449,"duration_ms":5075,"temperature":1.0,"reasoning_tokens":375,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:34:08.189832+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GML on a fixed channel and record, at each inner-loop iterate, the inner product $\\nabla L(\\beta_i)^T d_i$ between the loss gradient and the network-chosen step. If any non-terminal iterate has that inner product positive, or if $\\|d_i\\|$ exceeds any uniform $c_2\\|\\nabla L(\\beta_i)\\|$, then Assumption 2 fails for that run and Proposition 3's convergence proof does not apply, so the reported convergence would be an empirical observation rather than a guaranteed one. Alternatively, for a small system such as $M=3$, $K=2$, enumerate near-optimal antenna positions and beamformers on a fine grid and check whether GML's final secrecy rate falls below the best grid point; a persistent gap would indicate the method is not finding the true optimum.","supporting_citations":[{"cited_title":"MU-MIMO communications with MIMO radar: From co-existence to joint transmission,","cited_arxiv_id":null,"evidence_quote":"Provides the convex sensing-beampattern formulation used to construct the desired covariance R_d and the sensing penalty."},{"cited_title":"Secure radar-communication systems with malicious targets: Integrating radar, communi- cations and jamming functionalities,","cited_arxiv_id":null,"evidence_quote":"Establishes the secure radar-communication and jamming context that motivates artificial-noise design against malicious targets."},{"cited_title":"Movable antenna-assisted integrated sensing and communica- tion systems,","cited_arxiv_id":null,"evidence_quote":"Represents the joint MA-ISAC optimization literature against which the paper positions its contribution."},{"cited_title":"Movable antenna-aided secure full-duplex multi-user communications,","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-velocity particle swarm optimization baseline for antenna-position search."},{"cited_title":"Robust beamforming for RIS-aided communications: Gradient-based manifold meta learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the channel estimation error model and a gradient-based meta learning application in communications."}],"review_version":1}