MilCivVeh: Military & Civilian Vehicles

• MilCivVeh is a cross-domain research nexus that integrates military and civilian vehicular challenges through shared models of mobility, perception, and communications.
• It leverages common optimization structures for off-road route planning, DSRC coexistence, and mmWave V2X communications to address terrain and risk factors.
• The framework also includes an object-detection dataset for evaluating adversarial poisoning and ISAC techniques, highlighting practical system limitations and future directions.

MilCivVeh denotes a research nexus centered on military and civilian vehicles under shared methodological concerns in mobility, perception, communications, and sensing. In the current literature, the term appears in two closely related senses. In mobility and communications work, it functions as a shorthand for military/civilian vehicular problems in which common abstractions, constraints, and optimization structures are reused across defense and civilian domains. In computer vision, it also names a public object-detection dataset, “MilCivVeh,” short for Military Civilian Vehicles, created to study poisoning attacks and defenses on military-relevant imagery (Betz et al., 18 May 2026, Liezenga et al., 3 Sep 2025).

1. Terminological scope and research domains

MilCivVeh research is organized around the observation that many vehicle-centric problems are structurally shared across military and civilian settings. The route-planning literature treats off-road movement in open or weakly structured terrain as a network optimization problem whose cost is driven by soil trafficability and spatial risk fields; although the motivating application is military, the same framework is explicitly extended to wildfire response, disaster relief, agricultural operations, and off-road recreation (Betz et al., 18 May 2026). Communications work makes a parallel claim: V2V and V2X abstractions built for civilian C-ITS, mmWave sidelink, or DSRC coexistence are directly reusable for convoy communications, ad hoc formations, and other military vehicular scenarios, provided that additional security or mission-specific layers are added (Drago et al., 2020, Coll-Perales et al., 2018).

A second, narrower usage is the MilCivVeh dataset for object detection. That dataset contains three classes—civilian car, military truck, and military tank—and was introduced to evaluate whether poisoning attacks demonstrated on COCO and VOC transfer to military object-detection pipelines, and whether data-level patch detection can identify such attacks (Liezenga et al., 3 Sep 2025). This dual usage suggests that MilCivVeh is less a single formalism than a recurring cross-domain design pattern: shared vehicle physics, shared communication constraints, and shared perception pipelines, combined with domain-specific threat models.

2. Off-road mobility, terrain risk, and route optimization

In off-road mobility planning, MilCivVeh is instantiated as a minimum-cost mixed-integer linear program over a discretized terrain grid. The operational area is represented as nodes connected by arcs to adjacent nodes, with a binary decision variable $X_{i,j} \in \{0,1\}$ indicating whether arc $(i,j)$ belongs to the selected route. The objective is a single weighted-sum cost,

$c = p + h + r + \mu,$

where $p$ is soil-strength or immobilization risk, $h$ is proximity risk relative to recent friendly convoys, $r$ is proximity risk relative to known enemy activity, and $\mu$ is a small uniform positive penalty used to prevent cost-free meandering (Betz et al., 18 May 2026).

The terrain and mobility stack is explicitly physics-informed. Soil moisture is estimated at 10 m resolution via EMT+VS using coarse soil-moisture input, fine-resolution topography, and vegetation data; soil strength is then estimated with STRESS, which uses ROSETTA-derived hydraulic properties, texture-based friction angle and apparent cohesion, and an unsaturated soil mechanics model to compute moisture-variable cohesion; those outputs are mapped to an equivalent empirical Rating Cone Index (RCI) (Betz et al., 18 May 2026). Vehicle capability enters through the Vehicle Cone Index for 50 passes, $VCI_{50}$, and the soil penalty is

$$p = \begin{cases} P(VCI_{50} - RCI) & \text{if } VCI_{50} \ge RCI \ 0 & \text{otherwise.} \end{cases}$$

This formulation makes heavier vehicles more constrained because more nodes satisfy $RCI < VCI_{50}$.

The tactical terms are purely spatial exponentials. Historical-route risk is modeled as

$(i,j)$0

with $(i,j)$1 the distance to the closest node used by any previous convoy, and enemy-activity risk is

$(i,j)$2

with $(i,j)$3 the distance to the closest known enemy activity location. The baseline penalty is fixed at $(i,j)$4 (Betz et al., 18 May 2026).

The case study at Maxwell Ranch in northern Colorado uses a 10 m grid, June 10, 2022 environmental conditions, and vehicle scenarios ranging from $(i,j)$5 psi to $(i,j)$6 psi. Scenario analysis shows that objective values vary by five orders of magnitude depending on coefficient settings and vehicle properties, and that route lengths for a 9 km diagonal separation are typically 9.6–14.6 km. Lighter vehicles exhibit routing in which tactical terms and soil risk are of similar scale, whereas heavier trucks are constrained to narrow high-RCI corridors and show comparatively little flexibility in avoiding enemy or historical-route risk (Betz et al., 18 May 2026). The model does not include explicit speed or fuel functions, no epsilon-constraint or lexicographic multi-objective formulation is used, and the environment is static. Those omissions are important because they delimit the current MilCivVeh routing formulation as optimization-centric rather than fully dynamic mission planning.

3. MilCivVeh as an object-detection dataset and security benchmark

As a dataset, MilCivVeh is a small, custom object-detection corpus constructed from publicly available online resources for adversarial-AI research in the military domain. It contains 1,456 images, 2,438 annotated objects, and 3 classes: civilian car, military truck, and military tank. Images are resized to $(i,j)$7 pixels for experiments, and the dataset is used with bounding-box-based detection models including YOLOv3 and YOLOv10m (Liezenga et al., 3 Sep 2025).

The dataset’s role is methodological rather than merely taxonomic. It is used to study patch-based poisoning attacks derived from BadDet and Global Misclassification Attack (GMA), with military truck designated as the target class. Poisoned training images contain a blended adversarial patch, using the interpolation form $(i,j)$8 with $(i,j)$9, and labels are modified so that patched objects are learned as the target class. Training-time poisoning rates range from 20% to 40% depending on model, and patch sizes vary from $c = p + h + r + \mu,$0 to $c = p + h + r + \mu,$1 pixels in attack experiments (Liezenga et al., 3 Sep 2025).

The empirical finding is that poisoning attacks transfer only partially. For YOLOv3, the clean model reaches mean F1 of 94.2% on the clean test set; with 30% poisoning, ASR ranges from about 17.3% to 28.9%, and with 40% poisoning the maximum ASR reaches about 52.2%, but with substantial degradation in clean F1. For YOLOv10m, the clean model reaches mean F1 of 94.9%; 20% poisoning yields low ASR, while 30% and 40% poisoning can raise ASR to about 39.4% and 49.5%, again with substantial degradation in clean performance (Liezenga et al., 3 Sep 2025). The authors therefore conclude that MilCivVeh detectors are in principle vulnerable to patch-based poisoning, but that the attack, as implemented, requires high poisoning fractions and is not especially stealthy.

The same dataset is also used to evaluate patch detection. AutoDetect, an autoencoder-based detector trained on MS COCO and calibrated on a small clean MilCivVeh validation subset of 99 images, computes per-pixel reconstruction error, aggregates it over overlapping slices, fits a Gaussian clean-error model, and uses the maximum slice error percentile as a poisoning score. On the MilCivVeh detection set, AutoDetect achieves AUROC 0.941, compared with 0.902 for PAD and 0.538 for PaDiM (Liezenga et al., 3 Sep 2025). The paper notes, however, that MilCivVeh is limited in size, scene diversity, object scale realism, and viewpoint diversity; it is mostly ground-to-ground and does not represent air-to-ground imagery or richer military target taxonomies.

4. Shared-medium coexistence and military prioritization in DSRC

A distinct MilCivVeh problem is coexistence between civilian and military users on the same DSRC channel. In the DSRC urban-operations model, military and civilian vehicles are represented as two independent homogeneous Poisson point processes, $c = p + h + r + \mu,$2 and $c = p + h + r + \mu,$3, with $c = p + h + r + \mu,$4 and $c = p + h + r + \mu,$5 in the reported simulations. Both broadcast Basic Safety Messages at 10 Hz using IEEE 802.11 DCF on a 10 MHz DSRC channel, and the analytical model decomposes per-packet outcome into successful delivery, expiration, synchronized collision, and hidden-node collision (Kim et al., 2020).

The paper derives busy-slot probability, transmission-start probability, expiration probability, SYNC collision probability, hidden-node collision probability, and then expresses successful delivery as

$c = p + h + r + \mu,$6

which is also taken as PDR. Average delay and throughput are then defined from the probabilities of success, collision, and expiration using standard DCF timing parameters (Kim et al., 2020).

The proposed mitigation is a distance-based suppression protocol. A civilian node computes its distance to the nearest military vehicle,

$c = p + h + r + \mu,$7

and suppresses its own DSRC transmission whenever $c = p + h + r + \mu,$8. Military nodes themselves are unchanged; there is no modification to contention windows, EDCA parameters, or PHY power. The mechanism instead reduces the number of active civilian contenders in a protected zone around military nodes (Kim et al., 2020).

Simulation results show monotonic improvement in military PDR and throughput, and reduced delay, as $c = p + h + r + \mu,$9 increases. The reported average delay ranges roughly from 1.0 ms down to 0.9 ms, well below the nominal 100 ms allowable latency, while higher threshold distances correspond to larger civilian “quiet bubbles” and stronger military prioritization (Kim et al., 2020). The paper is explicit about the cost: civilians within the threshold region effectively lose DSRC service. This is technically a hard prioritization scheme, and the trade-off between mission-critical military QoS and civilian safety messaging becomes a central MilCivVeh coexistence issue rather than a purely MAC-layer optimization problem.

5. mmWave V2V, NR sidelink, and hybrid control/data-plane design

At higher data rates, MilCivVeh communications research shifts from DSRC coexistence to mmWave V2V and NR sidelink. MilliCar extends ns-3 with 3GPP NR V2X sidelink at mmWave for vehicle-to-vehicle communications. It implements MmWaveSidelinkSpectrumPhy, MmWaveSidelinkPhy, and MmWaveSidelinkMac, together with 3GPP TR 37.885 V2V highway and urban propagation models, LOS/NLOS/NLOSv states, DFT-based beamforming, configurable numerologies 2 and 3, and PSSCH-based data transmission integrated with LTE RLC/PDCP upper layers (Drago et al., 2020). Example scenarios include fixed-separation highway links, interfering vehicle groups, and platoon-like formations. With 400 MHz bandwidth and short distances, end-to-end delay can be around 1 ms or less at low offered load, whereas high offered load or longer distance increases queueing delay, particularly because the first version uses fixed slot allocations and does not implement HARQ (Drago et al., 2020).

The same line of work identifies beam alignment, deafness, and control overhead as core barriers for mmWave vehicular links. A sub-6 GHz assisted mmWave MAC addresses those barriers by assigning sub-6 GHz V2X to the control plane and mmWave to the data plane. Neighbor discovery, LOS identification, beam selection, and scheduling are piggybacked onto periodic 802.11p beacons through RTS-like and CTS-like fields, while mmWave is reserved for payload transmission (Coll-Perales et al., 2018). In the reported highway simulations, this design reduces per-neighbor control overhead from about 5800 bytes to about 100 bytes, eliminates the 12% mmWave time previously spent on control, schedules 100% of LOS neighbors for all tested transmitter ratios, and raises sub-6 GHz channel busy ratio only modestly, by about 0.77% at 15% mmWave transmitters and 2.04% at 40% (Coll-Perales et al., 2018).

The contrast between the two works is instructive. MilliCar provides a full-stack, standards-oriented simulator for mmWave NR sidelink behavior, including queueing and channel dynamics; the sub-6 assisted MAC paper shows that control/data-plane decoupling can materially improve spatial reuse and reduce access delay in dense vehicular settings (Drago et al., 2020, Coll-Perales et al., 2018). A plausible implication is that MilCivVeh systems requiring both convoy robustness and sensor-grade throughput will often need hybrid multi-band designs rather than a single-RAT solution.

6. Integrated sensing and communication for military and civilian vehicles

MilCivVeh also encompasses integrated sensing and communication (ISAC), especially at mmWave where radar and V2X functions compete for spectrum and hardware. In the reported 60 GHz ISAC study, three waveform families are compared under a common discrete-time model: FMCW, PMCW, and IEEE 802.11ad preamble-based radar using Golay complementary sequences (Sneh et al., 10 Feb 2026). The radar scene includes both single scatterers and extended targets: a pedestrian modeled by 27 point scatterers and a car modeled by 6905 metallic plates.

The principal sensing result is a ranking of waveform robustness under motion. FMCW exhibits poor autocorrelation properties with high sidelobes and a peak-to-sidelobe ratio of about 8 dB for a moving single scatterer, while PMCW improves to about 13 dB but still produces strong sidelobes that obscure extended-target structure. Standard 802.11ad Golay radar reaches about 43 dB PSLR, and Doppler-resilient Golay sequences based on Prouhet–Thue–Morse ordering maintain near-zero sidelobes even for moving targets, making the spatial extent of pedestrians and vehicles clearly visible in range–Doppler maps (Sneh et al., 10 Feb 2026).

The communication implication is that 802.11ad offers a communication-centric ISAC architecture in which the channel-estimation field in the PHY preamble is reused as a radar probe, while payload carries V2V or V2I data. The paper does not report BER or throughput curves, but it argues that high radar performance can be achieved without fundamentally compromising communication robustness when preamble overhead and hardware resources are managed appropriately (Sneh et al., 10 Feb 2026).

Implementation results reinforce that claim. The radar processing chain—FFT-based matched filtering, Doppler steering, and IFFT-based range recovery—is mapped to an AMD Xilinx Zynq MPSoC. A 24-bit fixed-point implementation reduces BRAM, DSP, LUT, and FF usage relative to single-precision floating point, lowers dynamic power from 3.454 W to 3.146 W, and increases acceleration over the ARM processing system from 9× to 10.6×, while preserving matched-filter output quality (Sneh et al., 10 Feb 2026). For MilCivVeh design, this establishes that ISAC is not only a waveform question but also an embedded-systems question shaped by SWaP constraints.

7. Recurring limitations and research directions

Across the cited works, several limitations recur. Off-road route planning is static, risk fields are purely spatial, and the current MILP omits dynamic re-planning, explicit slope constraints, multi-vehicle coordination, and uncertainty-aware optimization, although those are identified as extensions such as replanning loops, multi-commodity flow, and robust optimization (Betz et al., 18 May 2026). The MilCivVeh dataset is intentionally a proof-of-concept resource: it is small, limited to three classes, dominated by relatively large ground-view objects, and does not cover air-to-ground sensing or physical adversarial patches in operational conditions (Liezenga et al., 3 Sep 2025).

Vehicular communications work similarly exposes modeling gaps. MilliCar assumes perfect beam alignment, pre-configured mode-2-like slot patterns, no HARQ, and generic 3GPP propagation rather than mission-specific terrain, jamming, or security models (Drago et al., 2020). The sub-6 assisted mmWave MAC is evaluated on highways with fixed 50 ms mmWave transmission durations and does not provide a fully interference-aware spatial-reuse algorithm (Coll-Perales et al., 2018). The DSRC prioritization protocol quantitatively optimizes military performance but leaves authentication, spoofing resistance, and explicit civilian QoS accounting unresolved (Kim et al., 2020). ISAC evaluation, finally, is limited to one radar node, line-of-sight assumptions, a small target set, and no explicit communication-link performance analysis (Sneh et al., 10 Feb 2026).

Taken together, these limitations indicate that MilCivVeh is converging toward a broader systems problem rather than a set of isolated subfields. Terrain-aware routing, adversarially robust perception, multi-band vehicular networking, spectrum coexistence, and radar-communication integration are all being developed with abstractions that transfer between civilian and military vehicles, but each currently abstracts away parts of the operational stack. A plausible implication is that future MilCivVeh research will increasingly require joint treatment of perception, communications, mobility, and threat modeling within common experimental and optimization frameworks.