PRB Starvation at Slot Time via Rank Indicator Inflation: Attacking Proportional Fair Memory in the Scheduler

PRB Starvation at Slot Time via Rank Indicator Inflation: Attacking Proportional Fair Memory in the Scheduler Ryan Barker∗ , Julia Boone∗ , Tolunay Seyfi∗ , Alireza Ebrahimi∗ , Ethan Farquhar† , Brian Davis† , Joseph Boccuzzi‡ , Fatemeh Afghah∗

arXiv:2611.07225v1 [cs.NI] 4 Nov 2026

∗ Holcombe Department of Electrical and Computer Engineering, Clemson University

Clemson, SC 29634, USA rcbarke@clemson.edu, jcboone@clemson.edu, tseyfi@clemson.edu, alireze@clemson.edu, fafghah@clemson.edu † Savannah River National Laboratory, Jackson, SC 29831, USA ethan.farquhar@srnl.gov, brian.davis@srnl.gov ‡ NVIDIA Corporation, Santa Clara, CA 95051, USA jboccuzzi@nvidia.com

Abstract—Proportional fair (PF) scheduling in fifth-generation (5G) New Radio (NR) ranks user equipment (UE) by instantaneous rate over an exponentially weighted moving average (EWMA) of delivered service. We identify a served-but-undelivered denialof-service (DoS) attack in which rank indicator (RI) inflation induces persistent physical resource block (PRB) starvation. Unsupported spatial layers cause first-attempt delivery failure, keeping the attacker’s service EWMA low; the evaluated PF allocator repeatedly grants it the full PRB pool, denying honest UEs transmission opportunities. We formalize this recurrence and evaluate it in 3,300 paired, 200 slot multi-user shards spanning rank 1 and rank 2 Sionna Ray Tracing geometries, three link adaptation controllers, PF and round robin (RR), four target identities, and five seeds. On the rank 2 geometry, selective RI inflation yields 94–95% target capture, reduces PRB Jain fairness from 0.997 to 0.28, and produces mean honest zerogrant runs of 41–56 ms; RR preserves equal allocation. Exact offline replay and a Monte Carlo recurrence isolate the deliveredservice denominator as the causal path. Cross-field and controller ablations show a weaker, controller-dependent channel-quality trigger, while activity-gated EWMA defenses remain structurally blind. The results motivate RI-consistency checks before feedback acquires scheduling authority and slot-time PRB evidence inside the O-RAN Distributed Unit. Index Terms—Open Radio Access Network, 5G New Radio, Proportional Fair scheduling, PRB starvation, slot time telemetry, scheduler native metrics, Sionna.

I. I NTRODUCTION Mission-critical 5G/6G services, including industrial control and Ultra Reliable Low Latency Communications (URLLC), depend on continuous service at scheduler timescale [1], [2]. A cell can nevertheless maintain acceptable second-scale throughput or fairness while an honest user equipment (UE) receives no transmission grant for tens of milliseconds. In an Research supported by the NVIDIA Academic Grant Program using DGX Spark, Sionna, and Sionna-RT. It is also supported in part by the State of South Carolina through the Battelle Savannah River Alliance Workforce Development Program and the National Science Foundation under Grants CNS-2202972, CNS-2318726, and CNS-2232048.

Open Radio Access Network (O-RAN), interoperable telemetry and control remain largely above the sub-10 ms decisions made inside the O-RAN Distributed Unit (O-DU) [3], [4]. We therefore ask whether falsified channel state information (CSI) can cause a scheduler to repeatedly reward a UE that consumes resources but delivers no service, and whether the resulting denial-of-service condition can be isolated at slot time. The attack surface is proportional fair (PF) memory. PF divides an instantaneous-rate estimate by an exponentially weighted moving average (EWMA) of delivered service. Earlier starvation studies reached this edge through online traffic shaping: an idle-then-bursty UE lets its average forget prior service and resumes with elevated priority [5], [6]. We expose a second path that requires no payload transmission and originates in CSI authority. Downlink link adaptation uses the rank indicator (RI) to select spatial layers [7], [8]. In the evaluated model, unsupported spatial layers cause failed delivery that consumes physical resource blocks (PRBs) without raising the target’s service average. PF repeatedly favors the target, denying honest UEs transmission opportunities: denial of service (DoS) and network brown-out conditions through scheduler-amplified PRB starvation. The execution regime is part of this claim. AtlasRAN shows that host-operating-system radio emulation can support protocol tests and controlled functional ablations, but real-time-factor (RTF) dilation, input/output backpressure, and instrumentation loss can invalidate wall-clock deadline or deployment-capacity inference [9]. We therefore use modeled slot progression in a paired Sionna physical layer and ray tracing harness. The claim is deliberately causal and policy-specific: it concerns scheduler behavior under the stated PF recurrence, not over-the-air attack realization, fronthaul timing, or deployed cell capacity. This paper makes four contributions. First, it formalizes a served-but-undelivered PF edge and derives a closed-form starvation horizon, extending classical idle/burst analysis to a continuously scheduled UE. Second, it performs a 3,300 shard

paired ablation across two multi-user ray tracing geometries, three link adaptation controllers, PF and round robin (RR), four target identities, five seeds, and RI, channelquality, Hybrid Automatic Repeat Request (HARQ), and precoder controls. Under selective rank 2 RI inflation, PF assigns 94–95% of downlink slots to the target and reduces PRB Jain fairness from 0.997 to 0.28. Third, exact schedule replay and a targetwise Monte Carlo recurrence verify that the channel-grounded numerator is unchanged and that delivery failure reaches allocation through the EWMA denominator. Fourth, the analysis shows why activity-gated EWMA defenses do not engage and identifies the required defense boundary: RI-consistency verification before allocation, paired with a slot-time PRB ledger inside the O-DU. II. PF M EMORY AND THE F EEDBACK ATTACK S URFACE

Fig. 1. Paired effect-injection harness. Sionna PHY and ray tracing sources generate channel-grounded state. Honest and attacked legs share channel, traffic, controller, target, and random state; the injector modifies only the selected report field. The allocator and delivered-bit ledger close the PF-memory loop. Multi-user results use modeled-time ray-tracing geometries, not an over-the-air attack implementation.

PF remains common in cellular scheduling because it balances multi-user diversity and long-term service by normalizing instantaneous opportunity with a smoothed service history [10]. The Third Generation Partnership Project (3GPP) standardizes the same-slot scheduling decision. The target observes its honest procedures and measurements but does not prescribe a single, report and the codebook cap r max , but not the other UEs, universal base-station scheduler [7]. Silence-aware proportional allocator state, random seed, or future channel. scheduling and the starvation-resistant proportional allocation The primary channel supports rank ri = 2 for each UE with quantiles (PAQ) method reduce burst-entry priority by while the codebook allows up to four scheduled layers. The freezing or shadowing the average-rate state during inactivity target reports [5], [6]. These methods address an idle attacker. They do not rbi [t] = min(rmax , ri [t] + ∆r ) , ∆r ∈ {0, 1, 2, 3}. (1) address a UE that remains backlogged, receives every grant, and repeatedly delivers zero bits. The rank 1 ray tracing control uses one transmit dimension CSI integrity work shows that syntactically valid feedback and four receive dimensions per UE, so its codebook cap is can still be semantically false. RI tampering can induce one and RI inflation is clipped before scheduling. We retain spatial misconfiguration [11], compressed-feedback poisoning this lockstep control to separate an ineffective report from the can disrupt MU-MIMO beamforming [12], and low-power rank 2 geometry, which has spatial headroom to over-claim. signaling can alter receiver or scheduler behavior without broadband jamming [13]. We do not implement these waveform B. PF Memory and the Served-but-Undelivered Edge paths. We inject their scheduler-visible RI or channel-quality Algorithm 1 summarizes the capture loop. The evaluated PF consequences before the same-slot decision and measure the allocator scores UE i at slot t by allocation response. SEi [t] The observability problem is architectural. Near-real-time , Ri [t + 1] = (1 − α)Ri [t] + αqi [t]ϕi [t], si [t] = RAN Intelligent Controller (RIC) loops operate above the Ri [t] + ε (2) O-DU scheduler cadence, whereas distributed applications (dApps) colocated with the O-DU are proposed for faster where SEi [t] = log2 (1+γi [t]) uses channel-grounded signal-toinference and control [14]. A one-second KPI can average over interference-plus-noise ratio (SINR) γi [t], ϕi [t] is the granted twenty 50 ms deprivation episodes. A slot-time PRB ledger transport-block bit count, and qi [t] ∈ {0, 1} is first-attempt retains their continuity and exposes whether a fairness loss delivery. The guard ε = 1 bounds the first-slot score. The came from channel opportunity, scheduling policy, or delivery update factor is α = 0.05, giving time constant τ = 1/α = 20 modeled slots, or 10 ms. Because the numerator never consumes failure. the falsified report, any allocation change enters through III. T HREAT M ODEL AND S LOT-T IME A NALYSIS delivered service in the denominator. A. System and Adversary If a UE receives no delivered bits for T updates, whether idle or unsuccessfully served, The harness in Fig. 1 serves four UEs in a single cell on a virtual 273 PRB reference carrier with 30 kHz subcarrier Ri [t + T ] = (1 − α)T Ri [t] ≈ e−T /τ Ri [t]. (3) spacing and numerology µ = 1, which gives a 0.5 ms slot. A 40 slot time-division duplex (TDD) period contains 28 downlink For ∆r ≥ 1 in the evaluated rank 2 geometry, unsupported slots. In each downlink slot, the allocator grants the full PRB layers dominate the modeled block error probability. The pool to one UE. The primary attack is selective: three UEs measured target therefore has qi [t] = 0 on each served report honestly and one target modifies only its own RI before slot, so Ri [t] → 0 while it consumes the full grant and

Algorithm 1 rank inflation capture at 0.5 ms slot time Require: supported rank ri , cap rmax , inflation ∆r ≥ 1 1: Target reports r bi ← min(rmax , ri + ∆r ); other CSI is unchanged. 2: PF selects arg maxk SEk /(Rk + ε); the winner receives all 273 PRBs. 3: Modeled delivery fails for the target; set qi = 0 and update Ri using (2). 4: The target score approaches SEi /ε; honest averages must decay before readmission. 5: Record per-slot winner, PRBs, decode, and EWMA state for replay and run-length metrics.

si [t] → SEi [t]/ε. This is the same geometric memory edge as an idle interval, but it is persistent because there is no traffic burst completion or backlog drain to end it. After an honest UE j wins once and delivers ϕj bits, its EWMA is approximately αϕj . Following n unserved updates, it can strictly outrank the pinned target only if SEj > SEi and   ln(αϕj /[ε(SEj /SEi − 1)]) , (4) nstarve (j) = − ln(1 − α) + where ⌈x⌉+ = max(0, ⌈x⌉). If SEj ≤ SEi , the strict horizon is unbounded absent a target decode, a tie-break, or another policy intervention. The measured bit counts place this horizon between 150 and 219 modeled updates across target identities, explaining why recovery depends on rare target outcomes rather than PF’s nominal fairness memory. C. Scheduler-Native Metrics Winner-take-all allocation makes instantaneous four UE Jain fairness equal to 1/4 in every downlink slot, independent of policy. We therefore sum each UE’s PRBs over a window W and compute P X ( i Gi (W ))2 . (5) Gi (W ) = PRBi [t], JW = P 2 4 i Gi (W ) t∈W

establish multi-user PRB starvation. The allocation claims use two stationary urban-canyon RT geometries with four UEs: a 1 × 4 per-UE rank 1 control and a 2 × 4 per-UE rank 2 spatial surface, with mean channel-grounded SINR from 9.6 to 11.7 dB. Every channel is evaluated with outer loop link adaptation (OLLA), self-adaptive link adaptation (SALAD), or SALAD with reported channel quality information in its modulation and coding scheme (MCS) path (SALAD-CQI) [8], [19]. A type I codebook selector chooses spatial rank. Multi-user RT cells then apply PF or RR to a virtual 273 PRB, 100 MHz, 30 kHz reference carrier. The Sionna grid has 14 orthogonal frequencydivision multiplexing (OFDM) symbols and 64 tones and yields scalar effective SINR. PRB accounting is frequency-flat and does not alter SINR; inter-user interference is not modeled. The multi-user audit contains 3,300 paired 200 slot shards. The rank 2 RT surface evaluates selective RI inflation at ∆r = 0–3 and PMI randomization off/on. The rank 1 RT surface evaluates selective CQI/SINR bias at 0, 3, 6, 9, 12, 15–21 in 1 dB steps, 24, 27, and 30 dB; lockstep HARQ flips at p ∈ {0, .05, .10, .15, .25, .50}; and clipped lockstep RI controls. Matched null legs are retained. Every cell spans three link adapters, PF/RR, and five seeds; selective cells also sweep target UE 0–3. Selective rank 2 RI is the primary hypothesis, CQI/SINR bias tests cross-field generality, and HARQ/PMI are specificity controls that are not pooled with the headline result. Honest and attacked legs share channel, traffic, mobility, controller, target, and random state; the injector changes only the selected report field. Recorded fields include channelgrounded SINR, spectral efficiency, granted bits, first-attempt decode, winner, PRB count, scheduled rank, and PF state. Offline replay recomputes every PF decision, and a Monte Carlo recurrence draws Bernoulli decodes at each UE’s measured rate. The complete 200 slot shard is the replication unit. Fig. 2 reports 95% percentile-bootstrap intervals over 60 shards per point, while Table II retains target identities separately. Slots are not treated as independent samples.

We report target win share, J200 over the full shard, the V. R ESULTS AND D ISCUSSION minimum J20 over 20 consecutive downlink decisions, and A. Capture, Fairness, and Saturation the longest contiguous zero-grant run for each honest UE. Fig. 2(a) shows the direct scheduling effect. Under honest PF, Each shard has 200 modeled slots and 140 downlink slots, so run lengths are censored at 70 ms. Throughput Jain fairness each target wins approximately one quarter of downlink slots. is computed separately from delivered bits to show when At ∆r = 1, pooled target share rises to 0.946 and remains identical at strengths two and three. Across target identities, the aggregate service equality masks allocation capture [15]. measured range is 0.940–0.953, while the controller means are IV. E XPERIMENTAL M ETHOD 0.955 for OLLA, 0.923 for SALAD, and 0.960 for SALADThe analyzed sweep separates single UE physical layer CQI. One unsupported layer is therefore sufficient to zero (PHY) controls from multi-user Sionna Ray Tracing (RT) first-attempt delivery and saturate the PF denominator path surfaces [16], [17]. The PHY controls are additive white across all three controllers. Higher strengths change scheduled Gaussian noise (AWGN) with a 1 × 1 rank 1 link, Third rank from three to four but leave scheduling, Jain, and runGeneration Partnership Project (3GPP) Tapped Delay Line C length metrics unchanged. and D (TDL-C/D) non-line-of-sight (nLoS) and line-of-sight Allocation fairness collapses much more sharply than (LoS) channels with 2 × 2 rank 2 links, and a Clustered Delay throughput fairness (Table I). PF PRB J200 falls from 0.997 Line A (CDL-A) mixed channel with a 2 × 4 rank 2 link to a pooled mean of 0.279, and the minimum J20 reaches [18]. They isolate controller and feedback response but cannot its winner-take-all floor of 0.25. The pooled mean longest

Fig. 2. Selective RI inflation on the rank 2 urban-canyon geometry. Each point pools 60 shards across four target identities, three link adapters, and five seeds; error bars are 95% shard-bootstrap intervals. Target capture and both fairness metrics saturate at ∆r = 1. PF reaches the four UE fairness floor and produces long honest-user blackouts, while RR remains report-blind in its PRB rotation.

TABLE I S ELECTIVE RANK 2 RI I NFLATION . VALUES A RE R ANGES ACROSS F OUR TARGET I DENTITIES ; E ACH C ELL C ONTAINS 15 S HARDS . RUN L ENGTH I S M EAN (M AXIMUM ) IN M ILLISECONDS . Policy ∆r PF PF RR RR

Target win

PRB J200

Tput. J200

Min. J20

Run ms

Decode

4.7–5.3 .88–.89 (8.0) 41.1–55.5 1–3 .940–.953 .275–.283 .715–.729 .250 .00 (69.5) 1.5 0 .250 1.000 .997 1.000 .88–.89 (1.5) 1.5 1–3 .250 1.000 .748–.749 1.000 .00 (1.5) 0 .241–.259

.997

.999

.863

Its scheduled rank remains one, target decode remains approximately 0.92, PF PRB J200 remains 0.995, and win shares remain near 0.24, 0.24, 0.26, and 0.26. The exploit therefore requires spatial headroom between the true channel support and the codebook rank cap. The other feedback controls do not reproduce the collapse: nonzero lockstep HARQ flipping retains mean PF PRB J200 = 0.994, while selective PMI randomization on the rank 2 surface yields mean target share 0.315 and PRB J200 = 0.973. These specificity checks and the rank 1 null are not pooled with the selective RI claim. B. Mechanism Verification

Replaying Eq. (2) on each downlink decision while advancing the EWMA on every modeled slot reproduces 100.0% of recorded PF winners in every selective rank 2 cell. Advancing honest zero-grant run is 50.5 ms; target-specific means span the EWMA only on downlink slots reduces match to 0.981– 41.1–55.5 ms, with observed maxima from 44.5 to 69.5 ms. 0.983. The schedule is therefore fully explained by the stated Between 68.9% and 83.5% of honest runs exceed the 10-ms PF recurrence and its all-slot update cadence. Because SEi PF time constant. RR holds target share at 0.25, PRB J200 at is channel-grounded, the falsified RI reaches allocation only 1.0, and longest runs at three downlink slots, or 1.5 ms. Under through qi ϕi and the denominator Ri . RR, the target’s zero decode rate lowers delivered-throughput The Monte Carlo recurrence predicts target capture of Jain fairness to about 0.748, but the target harms itself rather 0.943–0.995 against 0.940–0.953 measured, with per-cell winthan capturing honest PRBs. share error at most 0.022. Predicted longest runs are 115.8– Throughput fairness masks the allocation mechanism. PF 140 downlink slots versus 82.2–111 measured; independent throughput J200 = 0.715–0.729 is near RR’s 0.748 because Bernoulli outcomes overstate runs by omitting residual temporal either policy contains one zero-delivery UE. The PRB ledger structure that can readmit an honest UE. The closed-form separates them: RR wastes the target’s fixed quarter, whereas horizon is 150–219 updates, consistent with long or censored PF redirects about 70% of additional opportunities away from runs. Together, replay and recurrence isolate delivered-service honest UEs. Delivered throughput alone cannot distinguish memory as the dominant policy mechanism. self-inflicted failure from scheduler-amplified starvation. Table II rules out one favorable target placement: all RI inflation is inert on the rank 1 control at every strength. four targets capture at least 94% of downlink slots despite

TABLE II PF M ODEL C HECK AT ∆r = 1. W IN S HARE AND L ONGEST RUN A RE M EASURED /M ONTE C ARLO M EANS OVER 15 S HARDS PER TARGET. H ORIZON I S THE C LOSED -F ORM N UMBER OF M ODELED U PDATES . Target

Win share

UE0 UE1 UE2 UE3

.950/.995 .953/.988 .940/.943 .940/.951

Mean Absolute Longest Run Horizon Error (MAE) (Slots) .022 .017 .007 .007

108.0/140.0 82.2/140.0 111.0/115.9 103.0/120.6

219.3 211.8 149.5 154.0

different channel-grounded spectral efficiencies and horizons. The agreement supports the analytical mechanism without treating the closed form as an exact stochastic predictor. The numerator choice also makes the test conservative with respect to report authority. Equation (2) computes SEi from channel-grounded SINR, so neither RI nor CQI manipulation can directly raise the PF numerator. Fig. 4 shows a weaker denominator trigger under selective CQI/SINR over-reporting on the rank 1 geometry. The pooled PF target share reaches 0.50–0.52 at 24–30 dB, but the controller ablation is decisive: OLLA reaches 0.924–0.958, SALAD-CQI reaches 0.326–0.345, and vanilla SALAD remains at 0.25 because its MCS path is CQI-agnostic. RR remains at 0.25 with 1.5 ms runs. Failed delivery is therefore the common scheduler mechanism, while exposure depends on which feedback each controller consumes. C. Per-Slot Mechanism Trace

Fig. 3. Representative selective RI shard (∆r = 1, OLLA, target UE0, seed 0). The target wins 95.7% of downlink slots with zero delivered bits; isolated honest grants reset their EWMAs before geometric decay. Gray denotes uplink or special slots; the longest strict blackout is 34.0 ms.

Fig. 3 resolves the RI mechanism at individual-slot resolution. The target receives almost every complete PRB grant, while honest UEs appear as sparse impulses. Its zero delivered bits keep Ri below ε; each honest success raises that UE’s EWMA by approximately αϕj , which then decays until readmission. This is the readmit-and-reset behavior predicted by Eqs. (2)– (4). The shard is not a worst case: its target share is 0.957 and longest blackout is 34.0 ms, below the pooled maxima. It illustrates the mechanism without replacing the 60 shard aggregate and shows how isolated honest grants can improve long-window throughput while contiguous deprivation persists. D. Observability, Defenses, and Limitations The gated-EWMA variants used against bursty traffic can be written with a binary activity gate gi [t] as Ri [t + 1] = (1−αgi [t])Ri [t]+αgi [t]qi [t]ϕi [t]. Under RI inflation, the target is continuously backlogged and receives the full grant, so payload, PRB-threshold, and consecutive-small-grant gates all evaluate to gi [t] = 1. There is no idle interval to freeze. Activity gating cannot detect an attacker whose scheduled activity is the attack. A one-second KPI can average across many 41–56-ms blackouts, whereas the O-DU ledger exposes their continuity. The appropriate control point is therefore upstream of the allocator: compare reported RI with base-station spatial evidence before the report acquires scheduling authority. This follows Zero Trust’s principle that authenticated origin does

Fig. 4. Selective CQI/SINR over-reporting on the rank 1 geometry, pooled over 60 shards per point with 95% bootstrap intervals. At 24–30 dB, PF share reaches 0.50–0.52 and mean honest runs reach 19–22 ms; controller exposure ranges from no change under CQI-agnostic SALAD to 0.924–0.958 capture under OLLA. RR remains report-blind.

not imply semantic trust and aligns with O-DU-resident dApps and robust MU-MIMO scheduling research [14], [20], [21]. We do not design or evaluate that verifier here. RR is a causal control, not a proposed replacement for PF. Its report-blind rotation breaks the feedback loop but abandons channel-opportunistic scheduling and still serves an undecodable target. A practical defense should preserve PF for plausible feedback, constrain inconsistent RI, and use a bounded fallback when spatial evidence is uncertain. We identify this control point and its slot-time evidence, not a threshold or actuator. The scope is intentionally narrow. The allocator is winnertake-all with one virtual PRB pool, ε = 1, and α = 0.05; deployed schedulers can split PRBs and use different memory constants. The harness uses table-based block error rate (BLER), scalar grants, and per-layer BLER aggregation. It excludes coded forward-error correction execution, HARQ process timing, transport-block refitting, radio-frequency injection, fronthaul, and over-the-air realization. The AWGN, TDL-C/D, and CDL-A PHY cells test single UE link adaptation only; the starvation claim comes from modeled-time multi-user RT cells. Four UEs, five seeds per controller, 200 slot shards, and a 70 ms run-length ceiling limit tail resolution. No learning model is trained. These limits make the result a controlled scheduler ablation, not a deployment claim. Future work should repeat it on a timing-faithful inline-accelerated PHY such as Aerial CUDA-Accelerated RAN (ACAR), where graphics processing unit (GPU) decoding headroom and slot deadlines can be measured jointly, and explicit multi-user PHY interference modeling can test whether adversarial CSI induces inter-layer, inter-user, or intra-cell co-channel interference (CCI) in addition to scheduler-mediated starvation [22]–[24]. VI. C ONCLUSION A rank indicator is one integer in a CSI report, but it carries significant scheduling authority. Across all three evaluated linkadaptation controllers and all four target identities, inflating it by one on a channel with spatial headroom drives PF toward 94–95% target capture and leaves honest UEs without grants for tens of milliseconds. RR prevents PRB capture by construction, although it continues wasting the dishonest UE’s allocation. Exact schedule replay and the recurrence model show that the effect is not a report-inflated numerator: failed delivery decays PF’s service denominator while the target remains scheduled. The weaker, controller-dependent CQI ablation and the HARQ/PMI specificity controls reinforce that boundary. The evaluated winner-take-all PF allocator thus amplifies RI-induced delivery failure into DoS through honest-user PRB starvation. The next step is causal RI-consistency verification and slottime enforcement inside the O-DU, followed by validation on a coded, timing-faithful PHY. R EFERENCES [1] 3GPP SA1, “Service Requirements for the 5G System (Stage 1),” 3rd Generation Partnership Project, Technical Specification TS 22.261 V20.2.0, 2025.

[2] M. Keshavarz, A. Shamsoshoara, F. Afghah, and J. Ashdown, “A realtime framework for trust monitoring in a network of unmanned aerial vehicles,” in IEEE INFOCOM 2020 - IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS), 2020, pp. 677–682. [3] O-RAN Alliance, “E2 Service Model for Key Performance Measurements,” O-RAN Alliance, Technical Specification, 2025. [4] A. Garcia-Saavedra, X. Costa-Perez, M. Gramaglia et al., “O-RAN: Disrupting the virtualized RAN ecosystem,” IEEE Commun. Standards Mag., vol. 5, no. 4, pp. 96–103, 2021. [5] J. Yang, Y. Zhang, Y. Wang, and P. Zhang, “Average rate updating mechanism in proportional fair scheduler for HDR,” in Proc. IEEE GLOBECOM, 2004, pp. 3464–3466. [6] S. Bali, S. Machiraju, and H. Zang, “PAQ: A starvation-resistant alternative to proportional fair,” in Proc. IEEE ICC, 2008, pp. 3012– 3016. [7] 3GPP, “NR; Physical Layer Procedures for Data,” 3rd Generation Partnership Project, Technical Specification TS 38.214, Release 18, 2025. [8] K. I. Pedersen et al., “An overview of downlink radio resource management for UTRAN long-term evolution,” IEEE Commun. Mag., vol. 47, no. 7, pp. 86–93, 2009. [9] R. Barker et al., “AtlasRAN: Timing-aware evaluation of open-source 5G platforms for integrated wireless testbeds,” in Proc. IEEE Military Commun. Conf. (MILCOM), 5G/6G Military Communications Workshop (WS10), 2026, arXiv:2603.14661. [10] S.-B. Lee, I. Pefkianakis, A. Meyerson, S. Xu, and S. Lu, “Proportional fair frequency-domain packet scheduling for 3GPP LTE uplink,” in Proc. IEEE INFOCOM, 2009, pp. 2611–2615. [11] S. Timilsina and M. Krunz, “Tampering attacks on CSI feedback reports in 5G new radio,” in Proc. IEEE MILCOM, 2025, pp. 1005–1010. [12] F. Meneghello, F. Restuccia, and M. Rossi, “WHACK: Adversarial beamforming in MU-MIMO through compressed feedback poisoning,” IEEE Trans. Wireless Commun., vol. 23, no. 11, pp. 17 252–17 265, 2024. [13] N. Ludant and G. Noubir, “SigUnder: A stealthy 5G low power attack and defenses,” in Proc. ACM WiSec, 2021, pp. 250–260. [14] S. D’Oro et al., “Next-generation RAN architecture: Distributed applications for real-time control,” O-RAN Alliance, Contributed Research Report RR-2024-10, 2024. [15] R. Jain, D.-M. W. Chiu, and W. R. Hawe, “A quantitative measure of fairness and discrimination for resource allocation in shared computer systems,” Tech. Rep. DEC-TR-301, 1984. [16] J. Hoydis et al., “Sionna: An open-source library for next-generation physical layer research,” 2022, arXiv:2203.11854. [17] ——, “Sionna RT: Differentiable ray tracing for radio propagation modeling,” 2023, arXiv:2303.11103. [18] 3GPP, “Study on Channel Model for Frequencies from 0.5 to 100 GHz,” 3rd Generation Partnership Project, Technical Report TR 38.901 V18.1.0, Release 18, 2026. [19] R. Wiesmayr et al., “SALAD: Self-adaptive link adaptation,” 2025, arXiv:2510.05784. [20] S. Rose, O. Borchert, S. Mitchell, and S. Connelly, “Zero trust architecture,” National Institute of Standards and Technology, Tech. Rep. SP 800-207, 2020. [21] Y. Wu et al., “R3 : A real-time robust MU-MIMO scheduler for ORAN,” IEEE Trans. Wireless Commun., vol. 23, no. 11, pp. 17 727– 17 743, 2024. [22] NVIDIA Corporation, “NVIDIA aerial CUDA-accelerated RAN,” 2026. [Online]. Available: https://github.com/NVIDIA/aerial- cudaaccelerated-ran [23] R. Barker et al., “Six times to spare: Characterizing GPU-accelerated 5G LDPC decoding for edge-RSU communications,” in Proc. IEEE Veh. Technol. Conf. Fall, 2026, arXiv:2602.04652. [24] R. Gadiyar, L. Kundu, and J. Boccuzzi, “Building softwaredefined, high-performance, and efficient vRAN requires programmable inline acceleration,” NVIDIA Technical Blog, Oct. 2023. [Online]. Available: https://developer.nvidia.com/blog/building-software-definedhigh - performance - and - efficient - vran - requires - programmable - inline acceleration/

Related documents

• DiffusionHijack: Supply-Chain PRNG Backdoor Attack on Diffusion Models and Quantum Random Number Defense#180590
• Fast Privacy Loss Accounting for Random Allocation via Multiscale FFT#1535423
• Fast Privacy Loss Accounting for Random Allocation via Multiscale FFT#1514901
• TIME-WAIT Assassination Hazards in TCP#136681
• DAXFS: A Lock-Free Shared Filesystem for CXL Disaggregated Memory#2710