PaperlayerView exact revision on arXiv

Source-linked AI summary

MAPLE-Sec: Mobility-Aware Partitioned Language-Model Engine for Roaming IoT Security

Yuhang Qiu, Ziwei Li, Jinfeng Xu, Jialiang Xie, Fangmin Ren, Shirong Guo, Jing Yang, Sunil Prajapat, Por Lip Yee

arXiv:2610.37147v1cs.NI

TL;DR

Roaming IoT detectors face handover-driven loss of inference state, costly raw-data transmission, and latency limits on full-depth inference. MAPLE-Sec partitions a security-tuned model across device and edge, and reports improved F1 and lower latency than Cloud-only and EdgeShard.

  • Problem

    Roaming IoT detection must contend with handover state loss, costly raw-data transmission, and full-depth inference that exceeds mobile latency budgets.

    Source: §I · pp. 1–2

  • Method

    MAPLE-Sec partitions a security-tuned model across device and edge, combining trajectory-conditioned scheduling, token-budgeted early exit, and compressed inference-state migration.

    Source: §I · p. 2

  • Results

    MAPLE-Sec achieves a 9.3% relative F1 improvement over Cloud-only and a 2.9 percentage-point gain over EdgeShard, while reducing end-to-end latency by 2.5×.

    Source: §I · p. 2

  • Takeaways & Limitations

    MAPLE-Sec is designed to preserve detection continuity while controlling uplink, energy, and latency costs for roaming IoT endpoints.

    Source: §VIII · p. 13

    The mobility evaluation replays stationary traffic captures on generated vehicular traces, so traffic–mobility interdependence is approximated rather than measured.

    Source: §VIII · p. 13

Abstract

from arXiv · show

Mobile Internet-of-Things (IoT) endpoints such as connected vehicles, wearables, and drones continuously cross heterogeneous network domains, creating handover-driven attack surfaces that static, cloud-anchored detectors cannot follow. However, three obstacles block real-time detection on such roaming endpoints: detection state is lost at every inter-domain handover; transmitting raw telemetry to a fixed detector exhausts the uplink and the device battery; and full-depth language model inference cannot meet millisecond latency budgets on constrained hardware. To tackle these challenges, we propose MAPLE-Sec, a mobility-aware partitioning framework that splits a compact, security-tuned language model across a roaming device and its nearest edge node. First, a trajectory-conditioned partition scheduler recomputes the split point from predicted handover timing, residual battery, and uplink quality. Second, a token-budgeted early-exit mechanism emits a verdict at the shallowest layer that meets a confidence threshold. Finally, a learned activation-compression codec migrates the intermediate inference state across edge nodes during handover without retransmitting raw telemetry. We formalize the partition-and-migration decision as an on-device constrained optimization solved in sub-millisecond time. Extensive experiments across the CIC-IoT and Edge-IIoTset corpora under both IID and non-IID roaming conditions demonstrate that MAPLE-Sec sustains detection across mobility events, improving the converged F1-score by up to 9.3\% relative to Cloud-only and by 2.9 percentage points over the strongest edge-collaborative baseline (EdgeShard), while reducing end-to-end latency by up to 2.5$\times$. Ablation confirms that the activation-compression codec reduces migration energy by 2.4$\times$ versus the raw-activation migration variant.

I. INTRODUCTION

MAPLE-Sec addresses the loss of stateful detection context during handovers and the uplink, energy, and latency costs of roaming IoT detection. It combines device–edge model partitioning, adaptive early exit, and compressed state migration; the reported experiments show gains over Cloud-only and EdgeShard.

I. INTRODUCTION

  • Existing approaches do not preserve inference state across inter-domain handovers, while raw-data transmission strains uplink and battery resources and full-depth inference exceeds mobile latency budgets.In the evaluated mobility profile, handovers occur every 6–12 seconds; state loss can force a restart or a distant cloud fallback.Source: §I · pp. 1–2
  • MAPLE-Sec partitions a security-tuned model between device and nearest edge, combining trajectory-conditioned scheduling, token-budgeted early exit, and compressed state migration.The contributions also describe a joint optimization over latency, energy, and detection confidence.Source: §I · p. 2
  • MAPLE-Sec achieves a 9.3% relative F1 improvement over Cloud-only and a 2.9 percentage-point gain over EdgeShard, while reducing end-to-end latency by 2.5×.Experiments use a SUMO/LuST urban mobility trace on a physical Jetson Orin Nano.Source: §I · p. 2

II. BACKGROUND AND MOTIVATION

  • The motivating experiments replay CIC-IoT traffic over a SUMO-generated 5G NR mobility trace, with handovers every 6–12 seconds and a 20% attack rate.They use a 12-layer security-tuned encoder and the LuST Luxembourg scenario.Source: §II · p. 2

A. Threat Model

  • The threat model includes compromised endpoints and on-path attackers, but does not defend against a fully compromised edge node.ACC payloads must be integrity-protected against replay during handover.Source: A. Threat Model · p. 2

B. Handover-Induced Detection Discontinuity

  • At handover, Cloud-only detection collapses and recovers slowly, while edge-pinned detection plateaus near 0.62 F1; MAPLE-Sec stays above 0.90 F1 throughout.The figure caption reports a 54-point Cloud-only F1 loss and a 4-second recovery.Source: B. Handover-Induced Detection… · pp. 2–3, Figure 2 · p. 3

C. Uplink and Energy Cost Under Motion

Under motion, compressed activation offload lowers uplink and device energy costs compared with raw-activation offload. The latency profile also shows that most flows reach the confidence threshold at shallow depth, motivating budgeted early exit.

C. Uplink and Energy Cost Under Motion

D. Latency-Bounded Inference on Constrained Hardware

  • Most flows exceed the confidence threshold τ = 0.90 by layer 4, although cumulative latency grows with inference depth.The text characterizes this confidence saturation as structural for the encoder.Source: Figure 4 · p. 4, D. Latency-Bounded Inference on… · p. 4
  • The challenge-to-module mapping assigns state continuity to ACC, uplink and energy costs to TPS, and latency-bounded inference to TBEE.ACC migrates state, TPS minimizes transmitted volume under motion, and TBEE provides token-budgeted early exit.Source: D. Latency-Bounded Inference on… · p. 4

A. LLM and Transformer-Based Intrusion Detection

The related work spans language-model intrusion detection, edge inference, early exit, and mobility-aware offloading, but the paper identifies a gap in jointly preserving inference context and meeting roaming constraints. Its system model splits execution between an endpoint and edge node under latency, energy, and confidence considerations.

A. LLM and Transformer-Based Intrusion Detection

  • LLM-based intrusion detectors cover IoT, vehicular, and related settings, yet the cited approaches do not address handover-driven loss of stateful inference context.The cited work includes transformer-based IDS, federated or collaborative systems, and open-set detection.Source: A. LLM and Transformer-Based Intrusion… · p. 4

B. Edge Inference and Model Partitioning

  • The cited edge-inference and partitioning frameworks assume a stationary device-to-edge association and do not address state migration during handover.Examples include static layer sharding and hybrid model-parallel inference.Source: B. Edge Inference and Model Partitioning · p. 4

C. Early-Exit Neural Inference

  • Prior early-exit work includes class-aware exits and bandit-based policies, while TBEE adds an explicit token budget for worst-case exit under mobility constraints.The cited survey and hardware implementation broaden the related-work landscape but are not described as providing this budget feature.Source: C. Early-Exit Neural Inference · p. 4

D. Mobile Task Offloading Under Mobility

E. Research Gap and Positioning

  • Prior work addresses subsets of four requirements, but the paper identifies no framework combining mobility-adaptive partitioning, state migration, budgeted early exit, and realistic roaming evaluation.The cited categories include LLM-based IDS, model partitioning, early exit, task offloading, and newer anomaly-detection systems.Source: E. Research Gap and Positioning · p. 5

A. System Model

  • The system model places the model head and partition controller on the endpoint and the tail on the serving edge, with a cloud anchor store for shared behavioral anchors.Its global objective weighs latency and device energy subject to a detection-confidence constraint.Source: A. System Model · p. 5

B. Framework Overview

MAPLE-Sec divides inference between a roaming device and its serving edge, coordinating a mobility-aware split with confidence- and budget-based early exits. The scheduler adjusts the split to handover timing, battery, and uplink conditions.

B. Framework Overview

  • The framework runs the device head and edge tail, with TPS setting the split and ACC migrating the compressed activation and anchor scores at handover.TBEE checks for an admissible early verdict during device execution; otherwise, the activation is sent to the serving edge node.Source: B. Framework Overview · p. 5

C. Trajectory-Conditioned Partition Scheduler (TPS)

  • TPS chooses the split point from handover prediction, residual battery, and uplink quality, running more layers locally when handover is imminent or uplink is poor.It selects the split minimizing expected device cost by enumerating the discrete layer choices; its scheduler can react to a new handover prediction within one inference interval.Source: C. Trajectory-Conditioned Partition… · p. 5

D. Token-Budgeted Early Exit (TBEE)

  • TBEE exits at the shallowest layer meeting the confidence threshold or when the remaining token budget is exhausted, bounding worst-case tail latency.The threshold is calibrated on a validation split to meet the detection-error constraint; each layer adds one lightweight classifier evaluation.Source: D. Token-Budgeted Early Exit (TBEE) · p. 5

E. Activation-Compression Codec (ACC)

ACC compresses split activations for migration across edge nodes, preserving downstream verdict information while trading coded payload size against fidelity.

E. Activation-Compression Codec (ACC)

  • ACC is a learned encoder-decoder codec that migrates inference state across edge nodes during handover without retransmitting raw telemetry.The next edge node decodes the compressed activation and resumes the tail; ACC also forwards running anchor scores.Source: E. Activation-Compression Codec (ACC) · pp. 5–6
  • The codec is trained to preserve downstream-verdict information in the split activation while penalizing coded rate, with ρ trading fidelity against size.Its loss compares the tail verdict on reconstructed and original activations and adds a coded-rate penalty.Source: E. Activation-Compression Codec (ACC) · p. 6
  • For a 12-layer encoder with hidden size 256, ACC compresses a 1,024 B activation to a 160 B bottleneck, a 6.4× ratio at ρ = 0.05.The two-layer MLP encoder and mirrored decoder each take under 0.3 ms on a Jetson Orin Nano; transfer over a 4 Mb/s cell-edge uplink takes 0.32 ms.Source: E. Activation-Compression Codec (ACC) · p. 6
  • At ρ = 0.05, ACC achieves 84% payload reduction with verdict fidelity of 0.968, its operating point for minimizing latency subject to the fidelity constraint.The payload is 160 B, compared with 1,024 B for the raw split activation.Source: E. Activation-Compression Codec (ACC) · p. 7

V. PROBLEM FORMULATION AND SOLUTION

The formulation makes the partition and upcoming-handover migration choices subject to detection-confidence, transfer-time, battery, and split-depth constraints.

A. Problem Formulation

  • The decision variables include the device-edge partition point k and a binary flag μ indicating whether to migrate state at the upcoming handover.The formulation also specifies end-to-end latency and device energy quantities, though their expressions are not present in these passages.Source: A. Problem Formulation · p. 7
  • Endpoint compute and transmit power are denoted separately as πc and πtx in the device-energy formulation.The optimization also includes latency and energy costs for each flow.Source: A. Problem Formulation · p. 7
  • The constraints enforce detection confidence, transfer before predicted handover, residual-battery limits, and a split point no deeper than the exit layer.The problem is a mixed-integer program over a small discrete domain.Source: A. Problem Formulation · p. 7
  • The handover penalty is zero when the payload can transfer before handover and grows linearly with any timing deficit.The penalty function is part of the formulation's handover-aware cost.Source: A. Problem Formulation · p. 7
  • After each verdict, the behavioral anchor score is updated by an exponential moving average of the exit-layer class probability.The smoothing coefficient αs lies between 0 and 1, and ACC migrates these anchor scores alongside the compressed activation.Source: A. Problem Formulation · p. 7
  • Exact enumeration over the feasible partition and exit-layer choices with binary migration flag takes O(L^2) time, which is microsecond-scale for a few tens of layers.The feasible domain is small enough for the exact optimum to be found on the endpoint.Source: A. Problem Formulation · p. 7

B. Complexity Analysis

MAPLE-Sec's per-flow cost advantage over EdgeShard comes from early exits and reduced uplink payload, while codec migration adds little latency.

B. Complexity Analysis

  • Fig. 6 shows that at 4 Mb/s, TPS shifts the split from k = 6 to k = 10, reducing migrated payload from 1,024 B to 160 B.The figure also compares TBEE's confidence and budget exits and ACC's payload-fidelity operating point with naïve counterparts.Source: Figure 6 · p. 8
  • TBEE exits at layer 4.3 on average out of 12, cutting executed layers by 64%; TPS reduces uplink payload from 1,024 B to 160 B.ACC encoding and migration add less than 0.6 ms, compared with 35 ms saved by early exit.Source: Table III · p. 8, B. Complexity Analysis · p. 8

C. Solution Approach

MAPLE-Sec decouples partition selection from per-flow early exit, using a lookup policy and an online controller. Its exit rule is linear in model depth, while ACC migration bounds handover blind spots under the stated constraint.

C. Solution Approach

  • TPS precomputes an objective-minimizing split for each mobility, battery, and uplink grid point; TBEE selects each flow’s exit online, and migration is triggered when handover is predicted within the tail’s remaining time.The lookup runs in O(1), the combined controller in O(L) per flow, and it meets the sub-millisecond budget without an iterative solver.Source: C. Solution Approach · p. 8
  • The exit rule terminates in O(L) steps and returns a feasible solution when the confidence constraint holds; its cost gap vanishes as calibration error approaches zero and is empirically below 3%.The below-3% gap is reported for the temperature-scaled classifiers used in Section VI.Source: C. Solution Approach · p. 8
  • With ACC active and the migration constraint satisfied, the probability of a handover blind spot longer than the payload-transfer time is 0.038, versus 1.0 for an edge-pinned detector without migration.The bound approaches zero as handover prediction accuracy improves.Source: C. Solution Approach · p. 8

VI. IMPLEMENTATION AND EXPERIMENTAL SETUP

On a Jetson Orin Nano and under simulated roaming across two IoT corpora, MAPLE-Sec converges to higher F1 than the compared baselines and improves latency and mobility resilience. The evaluation reports detection, system-efficiency, and mobility metrics.

VI. IMPLEMENTATION AND EXPERIMENTAL SETUP

  • Experiments replay CIC-IoT and Edge-IIoTset flows over 1,200 SUMO/LuST vehicle traces, spanning 14,000 handovers; the physical Jetson device communicates with an RTX 4090 edge server over emulated 5G conditions.Both corpora use 80/10/10 splits and are evaluated under IID and non-IID partitions.Source: §VI · pp. 8–9

A. Evaluation Metrics

  • Detection quality uses macroaveraged F1 after 60 seconds and held-out test data; system measures include latency, energy, uplink volume, handover recovery, blind-spot rate, and throughput.Latency is measured from packet arrival to verdict emission, and the blind-spot threshold is 200 ms.Source: A. Evaluation Metrics · p. 9

B. Superiority of MAPLE-Sec

  • MAPLE-Sec’s non-IID advantage over EdgeShard widens relative to IID, and the authors attribute the distribution-shift benefit to ACC’s migration of behavioral context across handovers.The reported F1 values are 0.941 versus 0.892 under non-IID conditions.Source: B. Superiority of MAPLE-Sec · p. 10, Figure 8 · p. 10
  • MAPLE-Sec achieves the lowest mean latency (23 ms), lowest tail latency (38 ms at P95), and zero deadline violations, with a 3.8% blind-spot rate versus 100% for edge-pinned and 61% for Cloud-only.Source: B. Superiority of MAPLE-Sec · p. 10
  • Under IID, MAPLE-Sec reaches F1 0.962 versus EdgeShard 0.933 and Cloud-only 0.880; under non-IID, it reaches 0.941 versus EdgeShard 0.892, with a 2.5× latency reduction over EdgeShard.The IID convergence is roughly 1.6× faster; the non-IID condition uses α = 0.3.Source: Figure 7 · p. 10, B. Superiority of MAPLE-Sec · p. 10, Figure 8 · p. 10
  • MAPLE-Sec achieves the highest F1, precision, recall, and AUC-ROC on both corpora under IID conditions; its CIC-IoT precision is 0.965 versus EdgeShard’s 0.941.The reported Edge-IIoTset recall is 0.937 versus 0.905 for EdgeShard.Source: B. Superiority of MAPLE-Sec · p. 10, Table V · p. 11

C. Impact of Residual Battery Budget

TPS adapts the device-edge split to residual battery, reducing latency relative to a fixed split across the reported battery range. At lower battery, it offloads earlier to respect the energy cap.

C. Impact of Residual Battery Budget

  • At 20% battery, TPS offloads earlier and latency rises to 41 ms while respecting the energy cap; at high battery it reaches 23 ms, versus a fixed-split baseline’s constant 55 ms.Even at 80% battery, TPS has a 12 ms latency advantage, with payload reduced from 160 B to 110 B.Source: C. Impact of Residual Battery Budget · p. 11, Figure 10 · p. 11

D. Impact of Edge-Node Scale

MAPLE-Sec scales throughput across cooperating edge nodes, while ablations show differing roles for ACC, TBEE, and TPS. Sensitivity results report performance across confidence, codec-rate, and objective-weight settings.

D. Impact of Edge-Node Scale

E. Ablation Study: Design Priority and Tradeoffs

  • Without ACC, raw activation transfer fails to complete within the handover window in 38% of events at 4 Mb/s cell-edge uplink, and migration energy rises 2.4×.With a fixed split, TPS’s alternative is to move the partition from k = 6 to k = 9 below 6 Mb/s, reducing payload from 160 B to 80 B while raising device-side compute.Source: E. Ablation Study: Design Priority and… · pp. 11–12
  • Forcing full-depth inference raises latency from 23 ms to 47 ms and causes 23% of flows to violate the 40 ms deadline; TBEE’s B0 = 8 bounds worst-case latency at 38 ms.The reported confidence plateau at layer 4 motivates early exit for many flows.Source: E. Ablation Study: Design Priority and… · pp. 11–12
  • Removing ACC causes the largest F1 drop, 7.3 pp, and a 38.2% blind-spot rate; removing TBEE leaves F1 unchanged but yields 23% deadline violations, while removing TPS raises energy 1.7×.The study ranks module impact as ACC > TBEE > TPS.Source: E. Ablation Study: Design Priority and… · p. 12

F. Hyperparameter Sensitivity

  • Across confidence thresholds, IID F1 peaks at 0.962 for τ = 0.90; varying the latency-energy weights leaves MAPLE-Sec ahead of EdgeShard in F1 at every tested setting.As codec weight ρ varies from 10^-4 to 10^-1, payload size falls from 310 B to 60 B while F1 remains above 0.950.Source: F. Hyperparameter Sensitivity · p. 12
  • MAPLE-Sec’s non-IID F1 advantage over EdgeShard grows from 2.9 to 4.9 pp on CIC-IoT and from 3.0 to 5.3 pp on Edge-IIoTset.The reported evaluation uses α = 0.3 Dirichlet attack-class skew.Source: F. Hyperparameter Sensitivity · p. 12

G. Comparative Analysis with Recent Approaches

MAPLE-Sec is distinguished by combining mobility-adaptive device-to-edge partitioning, handover state migration, and latency-bounded inference. The conclusion reports benchmark gains while identifying evaluation boundaries and future extensions.

G. Comparative Analysis with Recent Approaches

  • Across three recent approaches, MAPLE-Sec uniquely combines mobility-adaptive device-to-edge partitioning and handover state migration with latency-bounded inference.The first two properties are absent from all three approaches.Source: G. Comparative Analysis with Recent… · p. 12

VIII. CONCLUSION

  • Ablation ranks C1 > C3 > C2: removing ACC drops F1 to 0.889 and raises migration energy 2.4×; removing TBEE breaches 40 ms for 23% of flows.Removing TPS raises energy 1.7× without reducing F1.Source: Figure 12 · p. 13
  • MAPLE-Sec improves F1 by 9.3% relative to Cloud-only and 2.9 percentage points over EdgeShard, with 2.5× lower latency.The codec reduces migration energy by 2.4× versus raw activation transfer.Source: §VIII · p. 13
  • Future work proposes jointly learning the codec and anchor representation across mobility traces and extending MAPLE-Sec to cooperative detection among roaming devices.Source: §VIII · p. 13
  • Evaluation replays stationary captures onto SUMO mobility traces, runs the edge tail and ACC decoder on a desktop GPU, and compares ACC only with raw migration—not quantization or learned image codecs.Source: §VIII · pp. 13–14
Loading 2610.37147v1…