{"id":"a3bdc31f-2a2d-4f1d-bc61-abd3156ec449","arxiv_id":"2511.22910","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A partitioned RIS that steers the data signal to the legitimate receiver and artificial noise to an eavesdropper, configured by iterative or DFT-based phase search, improves secrecy capacity in SDR experiments.","lead":"This paper tests a wireless security setup where a smart reflecting surface splits its elements: half boosts the signal for the intended receiver, half aims artificial noise at an eavesdropper. It compares two practical ways to set the surface's phases and shows, in simulations and a radio testbed, that the scheme raises secrecy capacity while limiting the eavesdropper's data rate.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DFT codebook is mathematically impossible: 256 mutually orthogonal binary vectors of length 128 cannot exist, undermining the iterative-vs-DFT comparison.","rationale":"The central claim has two parts: (i) the partitioned-RIS AN system achieves notable SC improvements, and (ii) the iterative phase-search method outperforms the DFT-based method. The reader's weakest assumption concerned no-LOS, which limits transferability but is explicitly stated and does not invalidate the experiments under that assumption. I instead identified a concrete mathematical impossibility in the DFT codebook description: 256 mutually orthogonal binary vectors of length 128 cannot exist. This is not a matter of assumption or scope; it is an internal inconsistency. If the DFT implementation uses fewer than 256 orthogonal codewords, the comparison is unfair because the iterative method is allowed more trials; if it uses 256 non-orthogonal vectors, the method is not a DFT-based orthogonal sweep. In both cases, the paper's claim that the iterative method outperforms the DFT method is not reliably established. This flaw is addressable by correcting the codebook description and re-running the comparison with equal trial counts, so the paper remains conditionally acceptable rather than rejected. The reader did not flag this issue, hence our disagreement on the weakest assumption.","tokens_in":8405,"tokens_out":19249,"duration_ms":186011,"concrete_test":"Generate the 256 codewords described in Section III-C.2 as binary 128-dimensional vectors. Compute all pairwise inner products; verify that at most 128 can be mutually orthogonal. If the codebook is not orthogonal or contains 256 vectors, re-run the DFT-based experiment with a valid 128-codeword Hadamard/DFT codebook and equalize the number of trials to 256 (e.g., two full sweeps or additional random codewords). If the iterative method still yields higher SC under equal trial counts and a well-defined codebook, the comparative claim survives; otherwise, it is an artifact of trial-count mismatch or codebook design.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The DFT-based algorithm (Section III-C.2) is described as using N=256 mutually orthogonal codewords, each of dimension N/2=128. Over the binary phase alphabet {0°,180°}, the maximum number of mutually orthogonal length-128 vectors is 128 (Hadamard bound). Therefore, the described codebook cannot exist. If only 128 orthogonal codewords are used, the DFT sweep has half the trials of the iterative method (256), making the comparison in Fig. 2 and the contribution claim ('iterative method... converges to a higher final value compared to the DFT-based method') unfair. If 256 non-orthogonal codewords are used, the method is no longer a DFT-based orthogonal sweep as claimed. Either way, the experimental comparison between iterative and DFT-based optimization is not well-defined, directly undercutting a central claimed result of the paper.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an artificial-noise (AN) driven RIS-assisted physical-layer security system. A 256-element RIS is partitioned into two equal halves: one half is phase-configured to steer the communication signal (CS) to a legitimate receiver (Bob), and the other half is configured to steer AN toward an eavesdropper (Eve). The authors derive capacity expressions, formulate a power-allocation problem between CS and AN subject to a secrecy-capacity / leakage constraint, and propose two practical phase-optimization methods: a sequential iterative method and a DFT-codebook sweep. The system is evaluated both in MATLAB simulations and in an SDR-based testbed. The central claims are that the proposed system achieves notable secrecy-capacity improvements, that the iterative method outperforms the DFT-based method, and that a power-allocation choice can limit Eve's capacity to a small fraction of Bob's capacity.","tokens_in":8648,"tokens_out":7469,"duration_ms":69364,"significance":"If the claims are sustained, the paper would be a useful experimental contribution to RIS-assisted physical-layer security: the 256-element SDR testbed and the measurement campaign are valuable, and the explicit power-vs-secrecy trade-off under a leakage constraint is of practical interest. The paper also ships concrete algorithms for binary-phase RIS optimization and compares them on a real setup, which is uncommon. However, the strength of the claims is currently limited by the ill-defined DFT codebook, the ad hoc determination of the leakage-constrained power-split parameters, and the lack of a clear derivation of the CVX-based optimization. These issues are load-bearing because they concern the main comparison and the claimed optimality of the power allocation.","major_comments":[{"comment":"The DFT-based routine is described as sweeping N=256 mutually orthogonal codewords, each of dimension N/2=128. Over the binary phase alphabet {0°,180°} specified in Section III-C, the maximum number of mutually orthogonal length-128 vectors is 128 (Hadamard bound). Therefore the described 256-codeword orthogonal sweep cannot exist. If only 128 codewords were used, the comparison in Fig. 2a is not fair (256 iterative trials vs. 128 DFT trials). If 256 non-orthogonal codewords were used, the description is inaccurate. Please provide the exact codebook and either rerun the comparison with equal trial counts or qualify the iterative-vs-DFT conclusion.","section":"Section III-C.2 and Fig. 2a"},{"comment":"The values α1=0.46 and 0.58 are not obtained by solving the optimization problem P in Eq. (10). The text states they are 'determined' using Fig. 2b so that Eve's capacity is lower than 1% of Bob's capacity. This is a post hoc selection from the same measured curves that are later used to validate the approach. The paper's contribution statement claims the power allocation is optimized to suppress Eve; as written, the experiments only show that a manually chosen α1 can satisfy the 1% constraint. Please either solve P and report the resulting α1 values, or clearly frame the experiment as a feasibility test at manually selected operating points.","section":"Section IV, Fig. 3"},{"comment":"The system model assumes no line-of-sight between the transmitter and receivers, so all received energy comes from the RIS. The testbed geometry in Table I shows Alice, Bob, and Eve placed within roughly one to two meters of each other, and the paper does not describe any shielding or antenna orientation that would suppress direct paths. If direct Alice-to-Bob or Alice-to-Eve components exist in the measurements, Eqs. (7)-(8) omit them, which would change the capacity calculations and the validity of the 1% leakage constraint. Please describe how direct paths were eliminated in the experiment, or extend the model to include them.","section":"Section II, Eqs. (5)-(8)"},{"comment":"The text states that 'the optimum values of α1 are calculated using the Matlab CVX tool,' but no convexity proof or precise formulation is provided. The SINR constraints in Eqs. (11)-(12) and the log-capacity objective in P are not generally convex in α1/α2; a one-dimensional search over α1 would be a different, defensible procedure. Please specify the exact optimization problem given to CVX, or state that an exhaustive or golden-section search was used. Without this, the label 'optimum' for the CVX-based results is unsupported.","section":"Section IV, Fig. 4"},{"comment":"In Eq. (7), the numerator contains β1^2+β2^2, where β1 and β2 are the contributions of the same CS symbol from the two RIS partitions. These two components arrive at Bob with a determinate relative phase and should combine coherently, so the instantaneous signal power is |β1 + β2 e^{jφ}|^2, not β1^2+β2^2, unless β1 and β2 are defined as independent random processes and the indicated expression is an expectation. This distinction matters because Eqs. (7)-(8) are used to compute the capacity values in Figs. 2-3. Please specify the exact definitions and assumptions, and either revise the capacity expressions or state the incoherent-combining approximation explicitly.","section":"Section III-A, Eq. (7)"}],"minor_comments":[{"comment":"The notation for the RIS partitions is inconsistent: the text says Bob's partition has indices n=1,...,N/2, while Eq. (3) assigns i=b to the summation over n=N/2+1,...,N. Please align the notation.","section":"Eqs. (3)-(4)"},{"comment":"Please clarify what one 'iteration' means for each algorithm: for the iterative method, a single element toggle; for the DFT method, one codebook evaluation. The x-axis in Fig. 2a should be labeled accordingly so that the 256-trial comparison is unambiguous.","section":"Section III-C.1 / Fig. 2a"},{"comment":"The 'zero case' in Fig. 2c is defined only as assigning zero phase to all elements; it would help to state that this is the benchmark mirror-like reflection and whether it corresponds to α1=1 or another operating point.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The DFT-codebook issue is the most serious: it directly affects the paper's main comparative claim. If the authors can provide a valid codebook and rerun the comparison with matched trial counts, the iterative-vs-DFT conclusion may survive. The other major concerns (ad hoc α1 selection, missing CVX details, direct-path assumption) are fixable with clearer modeling and explicit statements about what is optimized and what is measured. I do not see grounds for rejection, but the current manuscript overstates the optimality of its power allocation and the fairness of its algorithm comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What's actually new: an SDR testbed with a 256-element RIS partitioned to steer CS to Bob and AN to Eve, with capacity measurements for two practical phase-search algorithms. That experimental core is the paper's real value. The setup is thoughtfully described, simulation and measurement curves track each other reasonably, and the power-allocation trade-off, where holding Eve to 1% of Bob's capacity costs α1 dropping from ~0.99 to 0.46–0.58, is a concrete design guide for this niche. The paper earns credit for that.\n\nThe soft spot you need to know about is not the lack of error bars or the no-LOS assumption, though those are real limitations. It is the DFT codebook. The text says the DFT-based algorithm uses N=256 mutually orthogonal codewords, each of dimension N/2=128, while all RIS phases are binary {0°,180°}. That is impossible: at most 128 mutually orthogonal binary vectors of length 128 exist. So either the codebook has only 128 orthogonal codewords, making the 256-trial comparison against the iterative method unfair, or it has 256 non-orthogonal patterns and is mislabeled as an orthogonal DFT sweep. Both readings undercut the central \"iterative outperforms DFT\" result. This is a load-bearing flaw, not a typo. The authors need to specify the exact codebook (rows, columns, orthogonality definition) and rerun the comparison with equal trial counts under a valid codebook, e.g., a 128-codeword Hadamard set.\n\nOther issues are less severe: the α1 values are read off measured curves rather than solved from the stated optimization, so they should be called empirical heuristics, not optimal allocations. The CVX analysis in Fig. 4 lacks a derivation and convexity justification, so it reads as an unexplained black box. The no-LOS assumption is clearly stated, but it means the gains won't transfer to deployments with a direct Alice–receiver path. No code or data is provided, which limits the reproducibility of an experimental study.\n\nFor whom: researchers working on measured RIS-PLS or on practical phase-search codebooks. The testbed data and the partitioned-RIS AN concept are useful anchors. I would send it to peer review because the experimental contribution is real and the flaws are fixable, but I would not pass it as-is. Ask for a corrected DFT implementation, a clear derivation for the CVX part, and at least the measurement data. That level of revision would make this a solid reference.","headline":"Worth sending to review: an experimental RIS-PLS testbed with a real contribution, but the iterative-vs-DFT comparison rests on an impossible codebook and needs correction before any of its conclusions are used.","tokens_in":9135,"tokens_out":3595,"would_cite":false,"duration_ms":34949,"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":"A RIS split between signal and artificial-noise beams improves secrecy capacity, and an iterative phase-search method beats a DFT-based sweep.","keywords":["reconfigurable intelligent surface","physical layer security","artificial noise","secrecy capacity","RIS partitioning","phase shift optimization","software defined radio","power allocation"],"falsifier":"In the same testbed, remove the obstruction that blocks the direct Alice–Bob and Alice–Eve paths and rerun the experiment with the same α1 values; if Eve's capacity rises above the 1% threshold or the secrecy-capacity gap between iterative and DFT methods shrinks, the claim that the partitioned RIS alone controls Eve's channel is refuted.","tokens_in":8300,"feed_emoji":"📡","tokens_out":6040,"duration_ms":48520,"temperature":0.7,"pith_summary":"This paper proposes and experimentally validates a physical-layer security scheme in which a reconfigurable intelligent surface (RIS) is split into two halves: one half is tuned to strengthen the communication signal at the legitimate receiver (Bob), and the other half is tuned to steer artificial noise at the eavesdropper (Eve). The authors derive secrecy capacity from the received-signal expressions, then formulate a power-allocation problem that maximizes secrecy capacity while keeping Eve's capacity below a specified fraction of Bob's. Two practical phase-tuning methods are compared: an iterative element-by-element toggle between 0° and 180°, and a sweep over orthogonal DFT codebooks. Both simulation and software-defined-radio experiments show that the iterative method reaches higher secrecy capacity, and that imposing a 1% leakage limit forces roughly half the transmit power into the artificial-noise beam. The paper concludes that the AN-driven partitioned-RIS approach is feasible and effective for practical deployments.","feed_headline":"Half the RIS beams noise at eavesdroppers to raise secrecy","feed_subtitle":"Testbed measurements show a split 256-element surface with iterative tuning keeps Eve near zero while lifting Bob's rate","key_machinery":"The central mechanism is the partitioned RIS, where N/2 elements are configured to maximize the received communication-signal power at Bob and the other N/2 to maximize the received artificial-noise power at Eve. The phase configurations are restricted to binary shifts {0°, 180°}, and are found either by a sequential iterative algorithm that toggles each element and keeps the phase that raises the target received power, or by a DFT-based algorithm that sweeps N mutually orthogonal codewords. The secrecy-capacity expression Cs = [Cb − Ce]+ and the SINR-ratio constraint η = γ̄e/γ̄b carry the power-allocation trade-off: tightening η from 10% to 1% lowers the optimal data-power fraction α1 from","core_discovery":"On its own terms, the paper establishes that a single RIS partitioned into two equal segments—one acting as a beamformer for the data signal to Bob, the other as a jammer directing artificial noise at Eve—yields measurable secrecy-capacity gains. The derivation starts from the signal model in which only reflected paths reach the receivers; Bob's and Eve's capacities are expressed in terms of aligned and non-aligned signal components, and secrecy capacity is defined as the non-negative difference Cb − Ce. The optimization problem maximizes secrecy capacity over the power split α1, α2 subject to a minimum capacity for Bob and an explicit cap on Eve's capacity. Using binary phase shifts and onl","pith_inferences":["Because the entire analysis assumes no line-of-sight between Alice and the receivers, the measured gains should be understood as upper bounds for scenarios with a direct path; with a direct component, Eve can bypass the RIS and the 1% leakage calibration would need re-tuning.","The two RIS segments are optimized independently; a joint optimization that coordinates the segments, or adapts the partition ratio, could push secrecy capacity further or handle multiple eavesdroppers, which the paper lists as future work.","The results are tied to a specific free-space geometry and a single 256-element surface; testing in richer multipath environments would show whether the iterative method's advantage persists when the channel is not purely line-of-sight to the RIS.","The capacity expressions treat all non-aligned components as interference, but in a real receiver, some of those components might be exploitable; the measured bit/s/Hz figures therefore depend on the assumption that Eve's decoder treats everything else as noise."],"forward_implications":["If the claims hold, a practical RIS-based secure link can be built with a fixed 256-element surface and two single-antenna transmitters, requiring only received-power measurements for configuration, no full channel state information.","The iterative phase-search method should be preferred when the configuration budget is limited to a fixed number of trials, since it dominates the DFT sweep in final secrecy capacity in both simulation and measurement.","Deploying the scheme with a strict 1% leakage cap means budgeting roughly half the transmit power to artificial noise, which cuts secrecy capacity compared to the unconstrained optimum near α1 ≈ 0.99.","The measured trade-off between the leakage ratio η and the achievable secrecy capacity provides a concrete calibration point for securing links with heterogeneous eavesdropper threats."],"fun_headline_variants":["Split RIS: signal to Bob, noise to Eve, secrecy up","One RIS, two jobs: boost Bob, jam Eve","RIS halves: data to user, noise to eavesdropper","Tested: split RIS lifts secrecy via artificial noise","RIS partition: beam signal, jam eavesdropper"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The system assumes there is no line-of-sight between the transmitter and receivers, so all received power comes from RIS reflections; if a direct path exists, the eavesdropper's channel is not controlled by the RIS and the reported secrecy-capacity gains and leakage constraints would not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Split RIS: signal to Bob, noise to Eve, secrecy up","One RIS, two jobs: boost Bob, jam Eve","RIS halves: data to user, noise to eavesdropper","Tested: split RIS lifts secrecy via artificial noise","RIS partition: beam signal, jam eavesdropper"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000219,"raw_usage":{"total_tokens":1255,"prompt_tokens":698,"completion_tokens":557,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":442,"completion_tokens_details":{"reasoning_tokens":475}},"tokens_in":442,"tokens_out":557,"duration_ms":6149,"temperature":1.0,"reasoning_tokens":475,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T19:36:52.829873+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the same testbed, remove the obstruction that blocks the direct Alice–Bob and Alice–Eve paths and rerun the experiment with the same α1 values; if Eve's capacity rises above the 1% threshold or the secrecy-capacity gap between iterative and DFT methods shrinks, the claim that the partitioned RIS alone controls Eve's channel is refuted.","supporting_citations":[],"review_version":1}