{"id":"67342133-fcbc-482f-acec-b4d3dafe991e","arxiv_id":"2504.16146","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"An aerial active STAR-RIS with a generative diffusion-model DRL controller improves simulated covert satellite-to-ground capacity and fairness, but with limited reproducibility evidence.","lead":"A research team proposes mounting an active smart reflecting surface on a drone to relay covert signals from a GEO satellite to ground users in dense cities, and uses a diffusion-model-based deep reinforcement learning controller to steer the drone and tune the surface. Their simulations show large gains in sum rate, fairness, and covertness over standard baselines, but the paper provides no code, no field test, and no statistical error bars on the results.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (18f) omits p_a g_a in the active-RIS power budget; with Table 2 parameters this may make the simulated beamforming physically infeasible and the headline capacity gains unverified.","rationale":"The central claim depends on solving ASCCOP, whose active-power constraint (18f) is dimensionally inconsistent with the signal model. In Eq. (6), the signal incident on the active RIS is sqrt(p_a g_a) h_ar s_i + z_r, so after Phi the output signal power is p_a g_a||Phi h_ar||^2 + sigma_r^2||Phi||^2. Eq. (18f) drops the p_a g_a factor in the first term. With Table 2's p_a=59 dBW/MHz and g_a=51 dBi, this is a roughly 10^11 factor, so the feasible set for beta and phi used in training is far larger than the physical one. Unless the code happened to implement the corrected constraint, the reported capacities are not shown to be achievable within the AASTAR-RIS power budget. The paper also never reports P_active_max, so the reader cannot infer the intended scale. This is a load-bearing problem-formulation error, not merely a missing reproducibility detail. The concrete re-derivation and re-run test would settle whether the simulation used the corrected constraint and whether the headline gains survive. I therefore keep the reader's conditional verdict: acceptance requires correction and re-evaluation. The reader's stated weakest assumption, sim-to-real generality, is real but secondary; the power-budget inconsistency is more immediate because it affects the validity of the simulation results themselves.","tokens_in":24554,"tokens_out":14031,"duration_ms":138164,"concrete_test":"Independently re-derive Eq. (18f) from Eq. (6) and rerun the GDPG training/evaluation with the corrected constraint p_a g_a||Phi[n]h_ar[n]||^2 + sigma_r^2||Phi[n]||^2 < P_active_max for a stated P_active_max (e.g., 1 W). If the mean sum capacity, fairness, or covert-violation curves change materially, or no feasible beta exists with the reported p_a, g_a, L0, and alpha, then the original results relied on an under-constrained power budget. Also release the code or at least the final trained policy's beta_m values to verify feasibility.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.2 formulates the AASTAR-RIS power budget as ||Phi[n]h_ar[n]||^2 + sigma_r^2||Phi[n]||^2 < P_active_max, but Eq. (6) shows the actual output power is p_a g_a||Phi[n]h_ar[n]||^2 + sigma_r^2||Phi[n]||^2: the incident satellite signal sqrt(p_a g_a) h_ar s_i[n] is amplified by Phi. The p_a g_a factor (roughly 10^11 with pa=59 dBW/MHz and ga=51 dBi) is omitted from (18f). Hence the feasible set for the beamforming gains beta_m and phases phi_m used in training is far larger than the physical one; the optimized Phi may violate the true power budget, and the simulated 10-20 bps/Hz capacities and covert-violation counts are not established as achievable. The value of P_active_max is never reported, so the scale of feasible beta_m cannot be checked. This is a load-bearing problem-formulation error: the central claim of jointly optimizing active beamforming under a strict power constraint is not supported by the optimization problem as written.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes an aerial active STAR-RIS (AASTAR-RIS) mounted on a low-altitude platform to relay covert GEO satellite signals to multiple ground users in a dense urban environment. The authors derive the minimal detection error probability under Warden noise uncertainty with perfect Warden CSI, convert the covert requirement into a per-slot constraint, and formulate an optimization problem (ASCCOP) that maximizes the sum of Jain-fair channel capacities by jointly optimizing the platform trajectory and the RIS transmission coefficients. Because the problem is non-convex and long-term, the authors reformulate it as an MDP and propose GDPG, a diffusion-model-based deterministic policy gradient algorithm with an action-gradient improvement mechanism. Simulations compare GDPG with DRL benchmarks and ablation baselines, and a Raspberry Pi test reports inference latency.","tokens_in":24799,"tokens_out":9521,"duration_ms":83165,"significance":"The covert-constraint derivation is careful and parameter-free: the minimal DEP expression in Eq. (15) and the resulting inequality in Eq. (16) follow from standard noise-uncertainty modeling, and the capacity expression in Eq. (7) correctly accounts for active-RIS amplification noise. If the optimization problem were correctly constrained, the AASTAR-RIS concept combined with GDPG would be a useful design for covert satellite-terrestrial downlinks, and the use of diffusion policies for constrained continuous control is of interest to the DRL-for-wireless community. The practicality test on Raspberry Pi is a useful addition. However, the central performance claims rest entirely on single-run simulator results, and one load-bearing constraint in the problem formulation is dimensionally inconsistent, so the reported capacity and covertness gains are not yet established.","major_comments":[{"comment":"The active-RIS power budget omits the incident satellite signal power. With the received signal at the RIS equal to sqrt(p_a g_a) h_ar[n] s_i[n] + z_r[n] as in Eq. (6), the expected output power of the RIS is p_a g_a ||Phi[n] h_ar[n]||^2 + sigma_r^2 ||Phi[n]||^2, not ||Phi[n] h_ar[n]||^2 + sigma_r^2 ||Phi[n]||^2. Table 2 sets p_a = 59 dBW/MHz and g_a = 51 dBi, so the omitted factor is about 10^11 in linear units; the constraint as written is also dimensionally inconsistent because ||Phi h_ar||^2 is dimensionless while sigma_r^2 ||Phi||^2 is a power. Since P_active_max is never reported, the feasible set used in training cannot be checked, and the optimized beamforming may violate the true power budget. Correcting this requires reformulating Eq. (18f) and the corresponding penalty term in Eq. (21), and re-running the simulations; the performance claims in Section 5 are not established as achievable under the current formulation.","section":"Section 3.2, Eq. (18f)"},{"comment":"All performance comparisons are single-curve training results with no averaging over random seeds and no error bars or confidence intervals. DRL training is stochastic, and the reported gaps between GDPG and TD3/SAC/VAE-DPG may be within run-to-run variance; reporting at least five seeds with shaded interquartile ranges or confidence bands is necessary to support the claim that GDPG \"significantly outperforms\" the benchmarks. In addition, the penalty coefficients r_pc, r_pr, and r_pp in Eq. (21) are never given in Table 2, and P_active_max is not specified; these parameters determine how strictly the covert and power constraints are enforced during training, so the current results are not reproducible or verifiable.","section":"Section 5.2, Figs. 5-10"}],"minor_comments":[{"comment":"The acronym is inconsistent: the abstract of the provided manuscript uses \"AAT-RIS\" and \"aerial active transmissive reconfigurable intelligent surface,\" while the body and contributions consistently use \"AASTAR-RIS\" and \"aerial active simultaneously transmitting and reflecting reconfigurable intelligent surface.\" Please unify the terminology.","section":"Abstract and Section 1"},{"comment":"The stated state-space dimension 2M(K+1)+3K+9 does not match the listed state components. Counting the 2D positions (2), user coordinates (2K), Warden coordinates (2), the real/imaginary decompositions of the Mx1 complex channels h_ar and h_r1,...,h_rK (2M(K+1)), the previous action of dimension 2+2M, the previous K capacities, and the previous reward gives 2MK+4M+3K+7; this mismatch propagates to the space-complexity expressions in Section 4.4.1 and should be corrected.","section":"Section 4.1.2, Eq. (20)"},{"comment":"The training-complexity expressions are internally inconsistent: the opening sentence gives O(GN(B+C+(4+T)|theta|+6|phi|)), while the bullet summary concludes O(GN(B+2|theta|+4|phi|)) and drops the T|theta| term from the diffusion updates. These formulas should be reconciled.","section":"Section 4.4.1"},{"comment":"There are several minor typos and grammatical issues, including \"fails to exceed\" in Eq. (18f), \"AASATR-RIS\" in Section 3.3, and the ungrammatical sentence beginning \"Let h_ar[n] and h_rk[n] stand for the channel gains from ...\" in Section 2.3. A careful proofread is recommended.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The power-budget error in Eq. (18f) is the main gate for this paper. I would not accept the current version without correcting that constraint, reporting P_active_max and the reward penalty coefficients, and re-running the simulations with multi-seed statistics. The covert-constraint derivation is sound, and the paper is within scope for the journal; the issues are fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on arXiv:2504.16146. This is a competent systems-engineering paper, not a breakthrough. The genuinely new part is the combination: an airborne active STAR-RIS (really a transmissive RIS, since all elements are in full transmission mode) relaying GEO satellite signals to multiple ground users under a Warden-based covert constraint, solved with a diffusion-policy DRL algorithm called GDPG. The covert DEP derivation (Eqs. 10-16) is a correct restatement of standard noise-uncertainty results. The MDP formulation and complexity analysis are sensible, and the algorithm is a credible adaptation of the authors' earlier GDM+DPG work. The Raspberry Pi inference test is a useful practical addition.\n\nThe soft spots match the reader's report. All performance claims come from a single simulator run: no random seeds, no error bars. Several load-bearing parameters are not reported—P_active_max and the reward penalties r_pc, r_pr, r_pp. The state dimension formula does not match the listed state vector. The abstract and PDF describe different system names (AAT-RIS vs AASTAR-RIS), and calling it STAR-RIS while putting every element in full transmission mode is a terminology problem. The citation pattern is mostly fine, but refs [9] and [16] have the same title at the same venue with different pages and author lists, which looks like a broken reference.\n\nOn the stress-test: Eq. (18f) is indeed missing p_a g_a on the incident satellite term; the true output power of an active RIS includes p_a g_a||Phi h_ar||^2. But the stress-test's arithmetic overstates the impact. With Table 2 values and pa converted from dBW/MHz to per Hz, p_a g_a is about 1e5, not 1e11, and the path losses are huge. The covert constraint binds first: for rho=3 dB and epsilon=0.1 it caps beta around 2-4, keeping the true output power below roughly a milliwatt. So unless P_active_max is set to an implausibly small value, the missing factor is not what inflates the capacity gains. Still, it is a real formulation error and must be fixed, with P_active_max reported.\n\nThe central claim—that jointly optimizing trajectory and active beamforming with GDPG beats the listed benchmarks—is plausible but not established. The engineering idea is worth taking seriously, and the paper deserves proper peer review. I would not cite it until the simulations are cleaned up and the parameters disclosed.","headline":"A competent systems-engineering paper with a real new combination and a standard covert-constraint derivation, but the headline gains rest on simulations that lack seeds, error bars, and a couple of load-bearing parameters.","tokens_in":25315,"tokens_out":11487,"would_cite":false,"duration_ms":101074,"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 mobile aerial platform carrying an active STAR-RIS, whose trajectory and beamforming are jointly optimized by a generative diffusion-model deep reinforcement learning algorithm, can provide fair covert…","keywords":["Active STAR-RIS","Covert communications","Satellite-terrestrial networks","Fairness","Deep reinforcement learning","Generative diffusion model","Trajectory optimization"],"falsifier":"Run the trained GDPG policy in an outdoor testbed with a UAV-mounted active STAR-RIS, a GEO-like source, and a passive Warden radiometer; measure the empirical detection error probability and per-user capacity. The central claim fails if the empirical DEP drops below $1-\\varepsilon$ at the claimed transmit settings or if the fairness index falls substantially below the simulated value. A cheaper simulation check: set the Warden's noise uncertainty $\\rho$ to 0 dB and test whether any trajectory and beamforming satisfying the Eq. (16) constraint admits positive capacity; if none exists, the covert-feasible operating region rests entirely on the assumed noise uncertainty.","tokens_in":24368,"feed_emoji":"🛰️","tokens_out":8289,"duration_ms":71175,"temperature":0.7,"pith_summary":"The paper tries to establish that a mobile aerial relay equipped with an active simultaneous transmitting and reflecting reconfigurable intelligent surface (STAR-RIS) can make satellite-to-ground communication both faster and undetectable in dense urban environments where direct satellite links fail. It derives a closed-form covert constraint from the Warden's minimal detection error probability under perfect CSI and environmental noise uncertainty, then formulates an optimization problem (ASCCOP) that maximizes the sum of Jain's-index-weighted channel capacities subject to that constraint. Because the problem is non-convex, long-horizon, and high-dimensional, the paper proposes the GDPG algorithm, which uses a generative diffusion model as the policy and an action-gradient step for improvement. Simulations show the approach outperforms trajectory-only, beamforming-only, passive-RIS, and standard DRL baselines in sum capacity, fairness, and covertness, and runs in tens of milliseconds on a Raspberry Pi.","feed_headline":"Flying RIS relay lifts covert satellite capacity and fairness","feed_subtitle":"A generative deep-RL controller steers an amplifying RIS drone so urban users get fair, hidden GEO links.","key_machinery":"The load-bearing object is the AASTAR-RIS: an aerial platform whose active STAR-RIS transmission coefficient matrix is $\\Phi[n] = \\mathrm{diag}(\\beta_m[n] e^{j\\phi_m[n]})$, with amplification gains $\\beta_m[n] > 1$ and phase shifts $\\phi_m[n] \\in [0, 2\\pi]$, letting the relay amplify and steer the GEO satellite's signal. The covert constraint comes from the Warden's minimal detection error probability $\\xi^*[n]$ under perfect CSI and bounded environmental noise uncertainty $\\rho$; it reduces to an upper bound on the power $\\iota[n]$ reaching the Warden. The algorithm that carries the optimization is GDPG: a generative diffusion model acts as the policy, producing the action $a[n]$ by $T$ denoising steps conditioned on the state, while an action-gradient step $a \\leftarrow a + \\eta_a \\nabla_a Q(s, a)$ refines state-action pairs for policy improvement, stabilized by double critics and target networks. The diffusion policy is what makes exploration of the high-dimensional, frequently-penalized action space effective compared with unimodal Gaussian policies.","core_discovery":"The central claim is that covertness in a satellite-terrestrial downlink can be actively engineered rather than merely tolerated. By mounting an active STAR-RIS on a moving aerial platform and jointly optimizing its trajectory, per-element amplification gains, and phase shifts, the system satisfies a strict covert constraint derived from the Warden's minimum detection error probability while maximizing the sum of fair channel capacities across ground users. The paper proves the resulting optimization problem is non-convex and long-term, then supplies the generative deterministic policy gradient (GDPG) algorithm, in which a diffusion model produces actions through iterative denoising and a gradient-ascent refinement of state-action pairs performs policy improvement. In the reported simulations, this approach attains higher mean sum channel capacity, a higher fairness index, and fewer covert-constraint violations than DDPG, TD3, SAC, a VAE-enabled DPG, and ablations optimizing only trajectory or only beamforming.","pith_inferences":["Extension: The covertness guarantee hinges on the assumed Warden noise uncertainty ($\\rho = 3$ dB in the simulations); a Warden with a calibrated low-uncertainty radiometer would shrink the feasible region, so the practical margin of the scheme is set by how conservative that uncertainty budget is.","Extension: The same GDPG machinery could be transferred to LEO satellite handover or terrestrial UAV relay problems with eavesdroppers, replacing the STAR-RIS phase/amplitude model with the relevant channel model and retaining the diffusion-policy exploration benefit.","Extension: Since the reward penalty coefficients for covert, power, and position violations are not numerically specified, a direct ablation sweeping those coefficients would clarify whether the covert constraint is enforced by the penalties or by the action-gradient refinement, a test the paper does not report.","Extension: A field experiment with a real Warden radiometer measuring empirical detection error probability under the trained policy would be the direct validation of the covertness claim and would expose any sim-to-real gap."],"forward_implications":["If the central claim holds, a single aerial platform with an active STAR-RIS can act as a covert relay for GEO satellite links in dense urban environments without any dedicated jamming device, since covertness is achieved through the relay's own power and phase control under environmental noise uncertainty.","The closed-form covert constraint gives a per-slot, checkable bound on the power arriving at the Warden, so any candidate trajectory and beamforming policy can be tested for covertness before deployment.","The GDPG algorithm's design—diffusion-model policy plus action gradient—provides a general template for other high-dimensional resource-allocation problems where strict constraints make most actions illegal and rewards are sparse.","The reported inference times (tens of milliseconds, below the GEO propagation delay) indicate the controller could run online on an embedded board, making the scheme deployable in real time if the simulation environment is representative."],"supporting_citations":[{"why":"Supplies the UAV-mounted RIS covert relay architecture and the covert communication model that the AASTAR-RIS system extends.","marker":"[8]"},{"why":"Defines the STAR-RIS transmission coefficient model used for the active beamforming design.","marker":"[24]"},{"why":"Provides the environmental noise uncertainty model and the minimal detection error probability derivation that yields the covert constraint.","marker":"[29]"},{"why":"Basis of the deterministic policy gradient framework that GDPG extends with a diffusion policy.","marker":"[34]"},{"why":"Introduces the diffusion-model policy representation used to generate actions via denoising in the GDPG algorithm.","marker":"[36]"},{"why":"Supplies the action-gradient policy improvement mechanism and the diffusion-model DRL update procedure adopted in GDPG.","marker":"[38]"},{"why":"Defines Jain's fairness index used in the objective to weight user capacities fairly.","marker":"[30]"},{"why":"Justifies the worst-case assumption that the Warden has perfect CSI for all links, used to derive the strict covert constraint.","marker":"[22]"}],"fun_headline_variants":["Aerial active RIS boosts covert satellite fairness","Diffusion-driven RIS drone secures covert satellite links","Active STAR-RIS on drone maximizes covert capacity","Covert satellite links get fairness via flying RIS","GDPG steers aerial RIS for covert fairness gains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the simulation environment—Rician factors, path-loss exponents, noise levels, Warden noise uncertainty, and user geometry—matches a real dense-urban deployment closely enough that the policy trained in simulation delivers the reported covert capacity, fairness, and detection-error performance in practice; a large sim-to-real gap would invalidate the claimed gains.","fun_headline_variants_meta":{"raw":{"variants":["Aerial active RIS boosts covert satellite fairness","Diffusion-driven RIS drone secures covert satellite links","Active STAR-RIS on drone maximizes covert capacity","Covert satellite links get fairness via flying RIS","GDPG steers aerial RIS for covert fairness gains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000509,"raw_usage":{"total_tokens":2511,"prompt_tokens":1013,"completion_tokens":1498,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":629,"completion_tokens_details":{"reasoning_tokens":1436}},"tokens_in":629,"tokens_out":1498,"duration_ms":10004,"temperature":1.0,"reasoning_tokens":1436,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:15:45.774079+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained GDPG policy in an outdoor testbed with a UAV-mounted active STAR-RIS, a GEO-like source, and a passive Warden radiometer; measure the empirical detection error probability and per-user capacity. The central claim fails if the empirical DEP drops below $1-\\varepsilon$ at the claimed transmit settings or if the fairness index falls substantially below the simulated value. A cheaper simulation check: set the Warden's noise uncertainty $\\rho$ to 0 dB and test whether any trajectory and beamforming satisfying the Eq. (16) constraint admits positive capacity; if none exists, the covert-feasible operating region rests entirely on the assumed noise uncertainty.","supporting_citations":[{"cited_title":"Covert communication assisted by UAV-IRS,","cited_arxiv_id":null,"evidence_quote":"Supplies the UAV-mounted RIS covert relay architecture and the covert communication model that the AASTAR-RIS system extends."},{"cited_title":"Simultaneously transmitting and reflecting (STAR) RIS aided wireless communi- cations,","cited_arxiv_id":null,"evidence_quote":"Defines the STAR-RIS transmission coefficient model used for the active beamforming design."},{"cited_title":"Robust covert multicasting aided by STAR-RIS with hardware impairment,","cited_arxiv_id":null,"evidence_quote":"Provides the environmental noise uncertainty model and the minimal detection error probability derivation that yields the covert constraint."},{"cited_title":"Deterministic policy gradient algorithms,","cited_arxiv_id":null,"evidence_quote":"Basis of the deterministic policy gradient framework that GDPG extends with a diffusion policy."},{"cited_title":"Multi-objective aerial collaborative secure communication opti- mization via generative diffusion model-enabled deep reinforce- ment learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the action-gradient policy improvement mechanism and the diffusion-model DRL update procedure adopted in GDPG."},{"cited_title":"Optimal tradeoff between sum-rate efficiency and Jain’s fairness index in resource allocation,","cited_arxiv_id":null,"evidence_quote":"Defines Jain's fairness index used in the objective to weight user capacities fairly."},{"cited_title":"Achieving covert wireless communications using a full-duplex receiver,","cited_arxiv_id":null,"evidence_quote":"Justifies the worst-case assumption that the Warden has perfect CSI for all links, used to derive the strict covert constraint."}],"review_version":1}