RIS-Aided mmWave Localization Under Cross-Link Interference via Beam-Domain ML Fingerprinting

RIS-Aided mmWave Localization Under Cross-Link Interference via Beam-Domain ML Fingerprinting Md Tarek Hassan

, Dmitry Zelenchuk

, and Muhammad Ali Babar Abbasi

arXiv:2608.07444v1 [eess.SP] 7 Aug 2026

Centre for Wireless Innovation (CWI), School of Electronics, Electrical Engineering and Computer Science (EEECS), Queen’s University Belfast, U.K. Emails: {mhassan15, d.zelenchuk, m.abbasi}@qub.ac.uk.

Abstract—Accurate user equipment (UE) localization is critical for beam management in reconfigurable intelligent surface (RIS)assisted millimeter-wave (mmWave) based sixth-generation (6G) networks, especially if the direct base-station–UE links are unavailable. This paper proposes a beam-domain fingerprint framework that maps the received signal-to-noise ratio (SNR) across a small set of predefined RIS reflection states to the UE azimuth angle and range, without requiring channel state information (CSI). Crucially, we extend the framework to a realistic interference-impaired scenario in which a nearby crosslink interferer (CLI) corrupts the clean SNR fingerprint, yielding a signal-to-interference-plus-noise ratio (SINR) fingerprint; an interference-to-noise ratio (INR)-constrained calibration strategy keeps the interference level physically interpretable. Four machine-learning (ML) regressors are evaluated under both conditions. Simulation results at 28 GHz with a 20×20 RIS show that k-nearest neighbors (KNN) achieves the lowest angle MAE of 0.37 degrees and range MAE of 4 cm under clean conditions, rising to 1.4 degrees and 7.6 cm under interference. A key finding is that interference degrades angle estimation substantially more than range estimation across all models, a consequence of the asymmetric encoding of location information in the beam-domain fingerprint. Index Terms—6G, Beam Management, Machine Learning, Reconfigurable Intelligent Surface, UE Localization

I. I NTRODUCTION Sixth-generation (6G) wireless networks are expected to support applications such as indoor navigation, autonomous robotics, and extended reality that demand sub-meter positioning accuracy as a native service rather than an add-on [1]. Millimeter-wave (mmWave) bands are naturally aligned with this requirement: their short wavelengths produce fine angular resolution, and their multi-hundred-megahertz bandwidths support centimeter-level ranging in principle [2]. However, mmWave propagation is inherently fragile, as buildings, vehicles, and even human bodies can obstruct the direct path entirely, rendering the link unusable [3]. Reconfigurable intelligent surfaces (RISs) have emerged as a practical remedy by acting as programmable phase-shifting mirrors that create This work was supported in part by the Department for the Economy of Northern Ireland under the U.S. Ireland Research and Development Partnership under Grant USI 269, in part by the Horizon Europe Project MULTIPLY-6G and in part by the Project REASON (Research on Enabling Architectures and Solutions for Open Networks) sponsored by the Department of Science, Innovation and Technology (DSIT). M. T. Hassan is supported by EPSRC and is also with Rajshahi University of Engineering & Technology (RUET), Bangladesh.

virtual line-of-sight paths around obstacles, restoring both communication and sensing coverage in otherwise blocked environments [4]. Beyond improving coverage, RISs offer a structural advantage for localization that is absent in conventional deployments, enabling users to be within the array’s near-field region. Hence, the received wavefront exhibits measurable curvature that varies distinctly from one spatial point to another [5], [6], encoding range information directly into the received signal and enabling joint azimuth and distance estimation from a single RIS with no additional anchor [7]. For mmWave frequencies, the Rayleigh distance of even a moderately sized RIS falls within the typical indoor operating range, placing users firmly in the near field and making this effect practically relevant [8]. Model-based estimators can exploit near-field wavefront curvature through maximum-likelihood or subspace methods, but they rely on a precisely known channel model [9]. In practice, however, 1-bit phase quantization adopted in established low-cost, low-complexity RIS architectures combined with a discrete beam codebook and the strongly nonlinear mapping that arises when the RIS steers into a limited number of fixed directions, substantially degrades the accuracy of analytic estimators [10]. Moreover, such methods require explicit channel state information (CSI) estimation, whose overhead scales with the array size and becomes prohibitive for large apertures [7]. These limitations motivate a data-driven alternative. Specifically, rather than deriving position analytically, a machinelearning (ML) regression model learns to map the received measurements to location directly from labeled training data samples [11], [12]. A compact and attractive measurement vector for this purpose is a beam-domain SNR fingerprint, formed by recording the received signal-to-noise ratio (SNR) under each of the predefined RIS reflection states. Each state illuminates the environment from a distinct angular direction, so the resulting dimensional vector encodes the user’s spatial signature without any explicit CSI estimation [12], [13]. In a fully blocked base station (BS)-user equipment (UE) scenario, where the RIS provides the only communication path as illustrated in Fig. 1, this cascaded fingerprint is the only available localization cue, yet the probing overhead with small probing beam states remains modest. Despite this promise, one practical complication has received very little attention from a localization perspective. In a multi-user mmWave deployment under flexible time-division

duplex (TDD) scheduling, a UE transmitting uplink in one cell can cause cross-link interference (CLI) in a neighboring cell operating on the same mmWave carrier frequency [14]. In RIS-assisted deployments, CLI is further compounded because the passive RIS reflects all impinging signals without discrimination, scattering the CLI signal toward the target UE in a beam-state-dependent manner and converting the clean SNR fingerprint into a signal-to-interference-plus-noiseratio (SINR) fingerprint at inference time. Since the ML model is trained on interference-free data, this mismatch between training and inference directly degrades localization accuracy. While prior work has addressed RIS-assisted localization [12], joint beamforming and positioning [13], and multi-user RIS systems [15], the impact of CLI-induced SINR fingerprint corruption on ML-based near-field joint angle and range estimation is largely unexplored, an important gap this paper fills. The main contributions of this paper are as follows. • A near-field focused BS–RIS–UE localization framework is proposed using a compact B-dimensional beamdomain SNR fingerprint as the ML input for joint azimuth-angle and range estimation. • A CLI model is introduced where a nearby UE in an adjacent cell induces beam-state-dependent leakage via RIS sidelobe scattering, corrupting the clean SNR fingerprint into a SINR fingerprint at inference time; an INRconstrained calibration ensures physical interpretability. • SINR corruption degrades angle estimation far more severely than range under CLI among the four ML regressor models evaluated. The remainder of the paper is organized as follows. Section II describes the system and channel models. Section III presents the ML-based localization framework and the interference model. Section IV evaluates performance under clean and interfered conditions, and Section V concludes. II. RIS-A SSISTED S YSTEM M ODEL A. System Geometry We consider an RIS-assisted mmWave localization system operating at a carrier frequency of f = 28 GHz, where the direct BS–UE link is blocked, as illustrated in Fig. 1. The study is based on a down-scaled experimental scenario using a 20 × 20 1-bit RIS developed in our laboratory at Queen’s University Belfast [16]. For practical pico-cell or macro-cell deployments, both the RIS aperture and the localization region can be scaled accordingly. The RIS is deployed on the yzplane and consists of L = My Nz passive elements arranged as a uniform planar array, with inter-element spacings dy and dz . The BS is fixed at pBS = (xt , yt , zt ) and the target UE at pu = (xu , yu , zu ). With a fixed UE elevation angle θr = 90◦ , the position reduces to two unknowns: azimuth φu from RIS to UE and range du from the RIS center, with xu = du cos φu and yu = du sin φu . All BS and UE distances satisfy the nearfield condition d < 2D2 /λ (where D is the effective aperture dimension) for the 20 × 20 RIS at 28 GHz as illustrated in Fig. 2(a), so the wavefront is spherical and the received signal carries both azimuth and range information.

Fig. 1: RIS-aided localization system model. B. Element-Wise Pattern Model pLet rt,l = ∥pBS − pl ∥, ru,l = ∥pu − pl ∥, and ρl = yl2 + zl2 be the BS-to-element, element-to-UE, and elementto-RIS-center distances, respectively. Each element is modeled as a cosine-pattern radiator. Applying the law of cosines to the triangle formed by the BS, the RIS centre, and element l, the BS-side off-boresight angle satisfies cos αt,l =

2 d21 + rt,l − ρ2l , 2 d1 rt,l

(1)

where d1 = ∥pBS ∥ is the BS–RIS distance. The incidence projection onto the RIS surface is xt inc cos θt,l = . (2) rt,l Combining both, the BS-side power pattern factor is
q t inc Ft,l = cos αt,l cos θt,l ,

(3)

where qt = Gt /2−1 is the pattern exponent that maps transmit antenna gain Gt (linear) with respect to a cosine radiation model. The analogous UE-side quantities are [4] 2 − ρ2l d2u + ru,l xu ref , cos θu,l = , 2 du ru,l ru,l
qr ref Fu,l = cos αu,l cos θu,l ,

cos αu,l =

(4) (5)

where qr = Gr /2 − 1 and Gr is the receive antenna gain. C. Cascaded BS–RIS–UE Channel Model The RIS applies B predefined reflection states sequentially; index b ∈ {1, . . . , B} identifies the active state. Under state (b) b, element l is assigned the 1-bit phase shift ϕl ∈ {0, π} determined by the codebook steering direction ϕb : ( 0, ϕpp,l (ϕb ) mod 2π ∈ [0, π), (b) ϕl = (6) π, otherwise, where ϕpp,l (ϕb ) = −k0 yl sin ϕb is the ideal progressive phase at element l for azimuth ϕb , and k0 = 2π/λ is the free-space (b) (b) wavenumber. The reflection coefficient is Γl = β ejϕl ,

(a) Near-field RIS geometry.

(b) Two-cell deployment scenario showing CLI.

Fig. 2: System model. (a) Near-field geometry of the RIS-aided BS–UE link. (b) A cross-link interferer UE(j) in an adjacent cell transmits uplink on the same carrier frequency; its signal impinges on the passive RIS of BS-A and scatters toward the target UE(k), corrupting the SNR fingerprint into a SINR fingerprint. where β ∈ (0, 1] is the element reflection amplitude. The effective cascaded channel from the BS to user u is L p   Ae X Ft,l Fu,l (b) (b) hu = Γl exp −jk0 (rt,l + ru,l ) , (7) 4π rt,l ru,l l=1

where Ae = dy dz is the element area. The power-domain factors Ft,l and Fu,l appear under square roots in (7) because they are power patterns, whereas the channel coefficient is a field quantity. The received power at user u is 2

Pu(b) = Pt Gt Gr h(b) , u

(8)

(b)

where Pt is the BS transmit power.

(b)

ωj

D. Beam-Domain SNR Fingerprint The deterministic thermal noise power in this scenario is σ 2 = kB T0 Bw Fn ≈ −80.95 dBm, where kB is Boltzmann’s constant, T0 = 290 K, Bw = 400 MHz, and Fn = 7 dB is the receiver noise figure. The per-state SNR at user u is (b)

Pu . (9) σ2 Activating all B states yields the beam-domain fingerprint  T xu = γu(1) , γu(2) , . . . , γu(B) , (10) γu(b) =

which encodes the user’s spatial signature without requiring CSI. This fingerprint is passed to a trained regression model, [φ̂u , dˆu ] = fML (xu ),

BSs operate on the same mmWave carrier frequency, UE(j)’s uplink signal propagates through the shared medium and impinges on the passive RIS surface of BS-A. The RIS cannot discriminate between the intended downlink pilot from BSA and the cross-link uplink signal from UE(j): both are reflected according to the active phase profile Φb . A fraction of UE(j)’s uplink energy therefore scatters toward UE(k) as beam-state-dependent interference, whose strength is governed (b) by the sidelobe coupling ωj of the active beam profile toward UE(j)’s angular direction. The normalized leakage weight is expressed as

(11)

where fML (·) maps the B-dimensional observation to the two localization parameters. III. P ROPOSED I NTERFERENCE - BASED L OCALIZATION A PPROACH A. Interference-Aware Inference Model At inference time, a cross-link interferer UE(j) from an adjacent BS-B transmits uplink on the same carrier frequency as the localization pilot of BS-A in Fig. 2(b). Because both

=

|hj |2 (m) 2

maxm |hj

| +ε

(b)

,

ωj

∈ [0, 1],

(12)

which captures the relative RIS coupling gain toward user j under beam state b; since beam b is optimized for user k, (b) coupling to user j arises through sidelobe leakage and ωj ≪ 1 in general. The beam-dependent interference coefficient is Pj (b) ξb = ηeff ρ ω , (13) Pk j where Pj /Pk is the cross-link power ratio, ρ ∈ (0, 1] is the temporal overlap factor, which is fraction of the probing interval during which user j transmits simultaneously with user k’s pilot reception, and ηeff ∈ (0, 1] is the effective coupling factor calibrated below. The per-beam INR-based SINR at user k is then (b)

(b)

INRk = ξb γk ,

(14)

(b)

(b)

γ̃k =

γk

(b)

.

(15)

1 + ξb γ k

To prevent unrealistically severe degradation, ηeff is chosen so that the maximum INR over all test samples i ∈ S and beam states does not exceed Γlin max = 10 dB. The pre-calibration coupling term is defined as Pj (b) (b) ω γ , i ∈ S, (16) ζi,b = ρ Pk j i

Table I: Comparison of SNR-based and Interference-Aware SINR-based localization performance. Model KNN RF DT SVR

SNR Angle

SNR Range

SINR Angle

SINR Range

Performance Change

MAE (◦ )

R2

MAE (m)

R2

MAE (◦ )

R2

MAE (m)

R2

∆ϕr (◦ )

∆R2

∆r (m)

∆R2

0.370 3.223 3.992 8.125

0.997 0.909 0.742 0.626

0.040 0.054 0.060 0.092

0.958 0.942 0.910 0.872

1.392 4.607 5.833 9.837

0.967 0.871 0.661 0.581

0.076 0.083 0.105 0.116

0.900 0.877 0.751 0.804

+1.022 +1.384 +1.841 +1.712

−0.030 −0.038 −0.081 −0.045

+0.036 +0.029 +0.045 +0.024

−0.058 −0.065 −0.159 −0.068

Algorithm 1 Beam-Domain Fingerprint Generation and Interference-Aware SINR Localization Require: RIS geometry, BS/interferer positions, azimuth angle set Aφ , K distances, B reflection states, system parameters, test set S, regressors {fm } Ensure: SNR fingerprint dataset D; SINR-based localization 1: — Phase 1: Training Fingerprint Generation — 2: Pre-compute BS-side distances and pattern factors 3: for all φu ∈ Aφ do (k) 4: Draw K distances {du }K k=1 5: for k = 1 to K do 6: for b = 1 to B do (b) (b) (b) 7: Compute hu , Pu , γu via (7)–(9) 8: end for (k) 9: Form xu via (10); store (xu , φu , du ) in D 10: end for 11: end for 12: Train regressors {fm } on D 13: — Phase 2: Interference-Aware Inference — (b) (b) 14: Compute {hj }, {ωj } via (7), (12) 15: for all i ∈ S do 16: for b = 1 to B do (b) (b) 17: Compute INRi , γ̃i via (14)–(15) 18: end for 19: Form x̃i ; predict via (11) and (18) 20: end for 21: return D; φ̂SINR , dˆSINR k

k

Based on that, the coupling factor is set to   Γlin max ηeff = min η, , maxi,b ζi,b

(17)

(b)

guaranteeing maxi,b INRk ≤ Γlin max . The corrupted finger(1) (B) T print x̃k = [γ̃k , . . . , γ̃k ] is passed to the same trained model as in (11) [φ̂SINR , dˆSINR ] = fML (x̃k ) . k k

(18)

Since fML is trained on clean fingerprints only, any accuracy loss directly quantifies each regressor’s robustness to fingerprint corruption without retraining. B. Dataset, Regression Models and RIS Phase Generation The training dataset D comprises N = 600 labeled examples, built by sweeping UE azimuth over 30 uniformly-spaced angles with K = 20 distances per angle. For each (φu , du ),

four 1-bit beam states—sufficient for angular discrimination across 0◦ –90◦ —are probed and the SNR fingerprint formed via (10) in algorithm 1. Classical regressors, such as k-nearest neighbors (KNN), random forests (RF), decision trees (DT), and support vector regressors (SVR) are chosen over deep learning to isolate CLI effects without confounding factors such as architecture or data volume. All are trained on an 80/20 split of clean fingerprints and tested under both SNR and SINR conditions with the same feature scaler. Given φ̂u , the RIS controller selects the nearest codebook beam and steers toward the unknown UE. IV. S IMULATION R ESULTS AND D ISCUSSION A. Simulation Settings and ML Performance Parameters We consider a 20 × 20 RIS with dy = dz = 4.6 mm and β = 0.7. The BS is placed at d1 = 0.8 m with (θt , φt ) = (90◦ , −30◦ ), while the UE elevation is fixed at θr = 90◦ and the azimuth varies within 0◦ ≤ φr ≤ 90◦ . The transmit power is Pt = 20 dBm, and the antenna gains are Gt = 15 dBi and Gr = 5 dBi. Four beam SNR measurements at each UE point for four swept beams ϕb ∈ {11◦ , 33◦ , 56◦ , 78◦ } are used as the ML inputs, whereas the outputs are the UE azimuth φr and range d2 . For the interference case, a nearby user at φj = 45◦ and d2,j = 0.48 m is introduced only at inference, with Pj /Pk = 0.25, η = 0.20, ρ = 1.00, and an INR cap of 10 dB. Mean absolute error (MAE) measures the average absolute difference between predicted and true values; lower is better. The coefficient of determination R2 measures how well predictions track the ground-truth trend, with R2 = 1 indicating a perfect fit. Table I compares prediction accuracy under the SNR baseline and the SINR interference scenario. Under clean conditions, KNN achieves the best angle MAE of 0.37◦ and range MAE of 4 cm, outperforming RF, DT, and SVR by a large margin. Under CLI, all models degrade, yet KNN remains the most robust, with angle and range MAE increases of only +1.022◦ and 3.6 cm, respectively. A central observation is that ∆ϕr consistently exceeds ∆r across all models: interference impairs angle estimation far more than range estimation, a finding elaborated in Section IV-C. B. Spatial Localization Validation Fig. 3(a) shows polar-domain predictions for the best and worst localization cases. Each subplot overlays the GT arc (red), predicted-range arcs, and predicted-angle spokes, where the annular band visualizes range error ∆r and spoke offset shows angle error ∆ϕ. In the best case (i, iii), all models

r = 1.130m

r = 1.230m

-30° -15°

0

0°

0.5

15°

GT (90.0°, 1.130 m) KNN: =0.0°, r = 0.00m r RF: DT:

=0.0°, r

r = 0.00m

=0.0°,

r = 0.00m

r

SVR:

90.0

0

0°

4.0

15°

0.5 =0.0°, r = 0.02m r

30°

75°

1.5

75°

90°

=44.6°,

r = 0.08m

=83.0°,

r = 0.03m

r

=75.6°,

r = 0.30m

GT 19.0 KNN: 19.0 RF: 20.3 DT: 19.0 SVR: 45.5

GT (19.0 deg, 0.770 m)

60°

(i) SNR Case

r = 1.230m

0

0° 15°

0.5

GT (90.0°, 1.130 m) =0.0°, r = 0.00m KNN: r RF: DT:

=0.0°, r

r = 0.00m

=0.0°,

r = 0.00m

r

SVR:

90.0

r

=0.0°,

RIS

15°

0.5

GT (4.0°, 1.230 m) =0.0°, r = 0.07m KNN: r RF: DT:

r r

=44.9°,

r = 0.09m

=83.0°,

r = 0.03m

SVR:

r = 0.02m

30°

1

0°

4.0

r

=74.9°,

-40 -50 -60 GT (19.0 deg, 0.770 m)

r = 0.28m

-70

30°

1

GT 19.0 KNN: 19.0 RF: 30.4 DT: 19.0 SVR: 47.0

-30

-30° -15°

0

-50

-70

BS

-15°

-40

-60

(ii) SNR Baseline - Worst Case

BS RIS

15

45°

45° 75°

r

SVR:

r = 1.130m -30°

90°

r

DT:

45°

60°

(i) SNR Baseline - Best Case

1.5

RF:

30°

1

45° 90°

GT (4.0°, 1.230 m) KNN: =0.0°, r = 0.06m r

Pr (dBm)

1

-15°

RIS

Pr (dBm)

RIS

1.5

-30

-30°

BS

BS

60°

1.5

75°

90°

(iii) SINR Baseline - Best Case

30

45

UE Scan Angle

60°

60 r

75

(deg)

(ii) SINR Case

(iv) SINR Baseline - Worst Case

(a)

(b)

Fig. 3: (a) Polar localization results: SNR and SINR best-worst case. (b) RIS radiation pattern for GT (19◦ , 0.770 m) under SNR and SINR scenarios. 1

1

0.9

0.9

0.8

0.8 0.7 Model KNN RF DT SVR

0.6 0.5

SNR90 0.90 deg 5.54 deg 5.00 deg 22.06 deg

SINR90 3.45 deg 9.18 deg 14.00 deg 22.54 deg

Delta +2.55 deg +3.64 deg +9.00 deg +0.47 deg

CDF

CDF

0.7

0.4

0.6

Model KNN RF DT SVR

0.5 0.4

0.3

KNN SNR RF SNR DT SNR SVR SNR KNN SINR

0.2 0.1 0 0

5

10

15

RF SINR DT SINR SVR SINR 90th pct. 20

0.3

SINR90 0.1637m 0.1819m 0.2413m 0.2231m

KNN SNR RF SNR DT SNR SVR SNR KNN SINR

0.2 0.1 25

SNR90 0.1034m 0.1377m 0.1491m 0.1805m

+0.0603m +0.0442m +0.0921m +0.0426m

RF SINR DT SINR SVR SINR 90th pct.

0 0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

Angle Prediction Error (deg)

Range Prediction Error (m)

Fig. 4: CDF of angle error (dashed = SNR, solid = SINR).

Fig. 5: CDF of range error (dashed = SNR, solid = SINR).

cluster tightly around the GT under SNR with only marginal drift under SINR. In the worst case (ii, iv), the GT lies at 4◦ near the beam-coverage edge, where a single beam provides little discriminative margin, causing DT and SVR to exhibit large angular confusion (∆ϕ > 75◦ ) even without interference. Across both cases, SINR conditions push all markers further from the GT, with angle errors growing disproportionately relative to range errors, corroborating the CDF analysis. Fig. 3(b) plots received power versus scan angle for GT (19◦ , 0.770 m). Under SNR, KNN and DT steer correctly to 19◦ , while SVR mispredicts by 26.5◦ , incurring a substantial power loss. Under SINR, RF degrades to 30.4◦ and SVR to 47◦ , confirming that even a moderate angle MAE causes the beam to miss the UE entirely.

C. Angle and Range Prediction Accuracy Fig. 4 shows the CDF of absolute angle error |∆ϕr |. Under the SNR baseline (dashed), KNN reaches the 90th percentile at only 0.90◦ ; under SINR (solid), this grows to 3.45◦ (∆ = +2.55◦ ), while DT reaches 14.00◦ , the largest degradation among the four models. Fig. 5 presents the range CDF, where the SNR–SINR gap is substantially narrower than in Fig. 4. This contrast reveals a fundamental asymmetry: range is encoded in the absolute power level of the fingerprint, which shifts uniformly under interference and is therefore relatively robust to angular beam leakage; angle, by contrast, is encoded in the relative power ratios between beams, which are directly perturbed when an interferer lies at a beam-dominance boundary. KNN’s 90th -percentile range error grows by only

Table II: (A) Mean quantization loss ∆Pr | (dB) for 1-bit vs. ideal phase (positive = magnitude; Worst row = signed minimum). (B) Computational complexity (Ntr = 480, 5-run median). (A) Quantization Loss ∆Pr | (dB)

(B) Computational Complexity

d2 (m)

Beam

Model

0.2

0.4

0.6

0.8

1.0

1.2

5.99 1.42 2.45 1.78

6.30 1.50 3.06 2.13

6.89 1.60 2.94 2.17

7.19 1.58 2.78 2.19

7.37 1.54 2.65 2.20

7.50 1.48 2.58 2.19

DT

Worst −13.55 −18.31 −15.71 −13.06 −12.26 −13.00

SVR

ϕ◦b 1 ϕ◦b 2 ϕ◦b 3 ϕ◦b 4

5.4 cm under interference, compared with a 2.55◦ angle increase. D. Quantization Loss and Computational Complexity Table II(A) reports mean 1-bit quantization loss per beam; Beam ϕb1 incurs the highest average penalty (5.99–7.50 dB) with worst-case −18.31 dB at d2 = 0.4 m, while Beams ϕb2 – ϕb4 remain below 3.1 dB. Despite this, binary phase control is sufficient for fingerprint-based localization since the regressors exploit relative inter-beam SNR ratios, which are preserved under a non-uniform quantization penalty. Table II(B) confirms all four models satisfy real-time constraints (< 1 ms inference): DT is lightest at 0.035 ms, while KNN achieves the best accuracy at only 0.199 ms, making it the recommended choice for online deployment. V. C ONCLUSION This paper investigated ML-based near-field UE localization in a blocked RIS-assisted mmWave system under cross-link interference. A four-beam swept fingerprint serves as the compact ML input, and an INR-constrained calibration strategy ensures a physically meaningful interference scenario. Among the regressors, KNN consistently achieved the lowest MAE of 0.37◦ (angle) and 4 cm (range) under clean conditions, rising to 1.39◦ and 7.6 cm under interference, alongside an inference latency of 0.19 ms per sample that satisfies realtime requirements. A fundamental asymmetry was identified, i.e., interference degraded the angle MAE by up to 4.4× while the range MAE increased by at most 1.3×. This disparity arises because range is encoded in absolute beam power, which remains robust to angular leakage, whereas angle relies on inter-beam power ratios that are sensitive to user interference at a beam-dominance boundary. Future work will extend the framework to distributed RIS-assisted deployments, to higherbit codebooks such as 2-bit or 3-bit, to online adaptation of the interference coupling model, and to experimental validation of the proposed approach. R EFERENCES [1] H. Tataria, M. Shafi, A. F. Molisch, M. Dohler, H. Sjöland, and F. Tufvesson, “6G wireless systems: Vision, requirements, challenges, insights, and opportunities,” Proceedings of the IEEE, vol. 109, no. 7, pp. 1166–1199, 2021.

KNN RF

Complexity Train: O(1) Infer: O(Ntr B) Train: O(T Ntr B log Ntr ) Infer: O(T log Ntr ) Train: O(Ntr B log Ntr ) Infer: O(log Ntr ) 2–3 ) Train: O(Ntr Infer: O(NSV B)

Train (ms) Infer (ms) 0.2

0.1986

64.2

2.3125

0.7

0.0353

1.6

0.2355

[2] A. Shahmansoori, G. E. Garcia, G. Destino, G. Seco-Granados, and H. Wymeersch, “Position and orientation estimation through millimeterwave MIMO in 5G systems,” IEEE Transactions on Wireless Communications, vol. 17, no. 3, pp. 1822–1835, 2018. [3] M. A. ElMossallamy, H. Zhang, L. Song, K. G. Seddik, Z. Han, and G. Y. Li, “Reconfigurable intelligent surfaces for wireless communications: Principles, challenges, and opportunities,” IEEE Transactions on Cognitive Communications and Networking, vol. 6, no. 3, pp. 990–1002, 2020. [4] W. Tang et al., “Path Loss Modeling and Measurements for Reconfigurable Intelligent Surfaces in the Millimeter-Wave Frequency Band,” IEEE Trans. on Commun., vol. 70, no. 9, pp. 6259–6276, 2022. [5] D. Dardari, N. Decarli, A. Guerra, and F. Guidi, “LOS/NLOS nearfield localization with a large reconfigurable intelligent surface,” IEEE Transactions on Wireless Communications, 2022. [6] D.-R. Emenonye, H. S. Dhillon, and R. M. Buehrer, “On the limits of single anchor localization: Near-field versus far-field,” IEEE Wireless Communications Letters, vol. 13, no. 2, pp. 540–544, 2024. [7] H. Lu et al., “A tutorial on near-field XL-MIMO communications towards 6G,” IEEE Communications Surveys & Tutorials, vol. 26, no. 4, pp. 2213–2257, 2024. [8] Y. Liu, C. Ouyang, Z. Wang, J. Xu, X. Mu, and A. L. Swindlehurst, “Near-field communications: A comprehensive survey,” IEEE Communications Surveys & Tutorials, vol. 27, no. 3, pp. 1687–1728, 2025. [9] A. Elzanaty, A. Guerra, F. Guidi, and M.-S. Alouini, “Reconfigurable intelligent surfaces for localization: Position and orientation error bounds,” IEEE Transactions on Signal Processing, vol. 69, pp. 5386–5402, 2021. [10] Y. Lin, S. Jin, M. Matthaiou, and X. You, “Channel estimation and user localization for IRS-assisted MIMO-OFDM systems,” IEEE Transactions on Wireless Communications, vol. 21, no. 4, pp. 2320–2335, 2022. [11] S. Katla, M. El-Hajjar, A. Mourad, and L. Hanzo, “Deep learning aided fingerprint-based beam alignment for mmwave vehicular communication,” IEEE Transactions on Vehicular Technology, vol. 68, no. 11, pp. 10 858–10 871, 2019. [12] W. Wang and W. Zhang, “Joint beam training and positioning for intelligent reflecting surfaces assisted millimeter wave communications,” IEEE Transactions on Wireless Communications, vol. 20, no. 10, pp. 6282–6297, 2021. [13] K. Li, M. El-Hajjar, and L.-L. Yang, “Reconfigurable intelligent surface aided position and orientation estimation based on joint beamforming with limited feedback,” IEEE Open Journal of the Communications Society, vol. 4, pp. 748–767, 2023. [14] 3GPP, “Study on Cross Link Interference (CLI) Handling and Remote Interference Management (RIM) for NR (Release 16),” 3rd Generation Partnership Project (3GPP), Technical Report TR 38.828, 2019. [15] H. Zhang, H. Zhang, B. Di, K. Bian, Z. Han, and L. Song, “MetaLocalization: Reconfigurable intelligent surface aided multi-user wireless indoor localization,” IEEE Transactions on Wireless Communications, vol. 20, no. 12, pp. 7743–7757, 2021. [16] G. Machado et al., “Design and Experimental Characterization of a Wideband Dual-Polarized mmWave Reconfigurable Intelligent Surface for B5G/6G Communications,” IEEE Transactions on Antennas and Propagation, 2026, (Early Access).

Related documents

• LatentWave: JEPA Pretraining for Wireless Foundation Models#259516
• Localization-induced coherent backscattering effect in wave dynamics.#1039602
• Selective Interference Suppression of Siamese-Net in Heterogeneous Interference Channels#780937
• Selective Interference Suppression of Siamese-Net in Heterogeneous Interference Channels#779562
• Towards energy-efficient massive MIMO-NOMA systems with sigma-delta ADCs and group SIC detection.#151476