LFRA-Net: Dual Lightweight Models

• LFRA-Net is a dual-purpose lightweight architecture that employs focal modulation and region-aware attention to achieve efficient retinal vessel segmentation.
• In wireless authentication, LFRA-Net uses rapid LoRA aggregation to adapt to varying channel conditions, resulting in improved AUC and reduced EER.
• Both implementations focus on resource-constrained environments, leveraging compact designs and attention mechanisms to enhance overall performance.

LFRA-Net is an overloaded model name in the recent arXiv literature. In one usage, it denotes the “Lightweight Focal and Region-Aware Attention Network” for retinal vessel segmentation in color fundus images (Mehmood et al., 15 Sep 2025). In another, it denotes a Rapid LoRA Aggregation-based framework for open-set radio frequency fingerprinting (RFF) under varying wireless channels, described as a channel-adaptive RFF neural network using Low-Rank Adaptation modules and rapid aggregation (Zhang et al., 14 Apr 2026). The shared acronym reflects different expansions and distinct problem settings rather than a single research lineage. Accordingly, LFRA-Net refers either to a compact encoder–decoder segmentation architecture for ophthalmic imaging or to a lightweight adaptation framework for physical-layer wireless authentication.

1. Terminological scope and naming ambiguity

The retinal LFRA-Net is introduced explicitly as “LFRA-Net: A Lightweight Focal and Region-Aware Attention Network for Retinal Vessel Segmentatio” and is framed as a lightweight, attention-enhanced encoder–decoder model for vessel segmentation in fundus images (Mehmood et al., 15 Sep 2025). Its defining elements are multiscale convolutions, focal modulation attention at the bottleneck, and region-aware attention in selective skip connections.

The wireless LFRA-Net is associated with Rapid LoRA Aggregation for Wireless Channel Adaptation in Open-Set Radio Frequency Fingerprinting (Zhang et al., 14 Apr 2026). The paper states that the name “LFRA-Net” is not explicitly used in the body, but also states that the system is exactly a channel-adaptive RFF neural network that uses LoRA modules per environment and rapid aggregation of these modules for new channel conditions. This suggests that the acronym operates more as a compact label for the proposed adaptation framework than as a separately instantiated backbone architecture.

A plausible implication is that the term should be interpreted contextually: in medical image analysis, LFRA-Net denotes a retinal vessel segmentation network; in wireless security, it denotes a LoRA-based channel adaptation framework for open-set RFF authentication. Because the two models solve unrelated tasks and rely on different primitives, they are best treated as homonymous entries within the technical literature rather than variants of one architecture.

2. LFRA-Net in retinal vessel segmentation

In the ophthalmic imaging context, LFRA-Net addresses retinal vessel segmentation, defined as producing a binary mask of arteries and veins from a fundus image (Mehmood et al., 15 Sep 2025). The paper situates this task as critical for the early diagnosis of diabetic retinopathy, glaucoma, age-related macular degeneration (AMD), neurodegenerative disorders such as dementia and Alzheimer’s disease, and cardiovascular and systemic diseases through vessel tortuosity, caliber, and branching patterns. Manual annotation is described as laborious, expertise-intensive, and error-prone, which motivates robust automatic segmentation.

The model is designed for scenarios with limited computational resources, where existing deep segmentation methods face a dual constraint: difficulty extracting tiny vessels, especially thin and low-contrast peripheral branches, and high computational overhead from heavy encoders, transformers, or stacked refinement networks (Mehmood et al., 15 Sep 2025). The stated objective is to deliver competitive or superior segmentation accuracy while keeping the model small enough for real-time deployment in resource-constrained environments.

Architecturally, the retinal LFRA-Net is an encoder–decoder (U-Net–like) model with three principal components (Mehmood et al., 15 Sep 2025):

• an encoder composed of three multiscale convolution blocks with downsampling,
• a bottleneck containing focal modulation attention,
• a decoder with three upsampling stages and selective skip connections modulated by region-aware attention.

The input image $I$ is resized to $512 \times 512$, and the output $I_{\text{out}}$ is a sigmoid-activated vessel probability map that is thresholded to obtain the segmentation mask (Mehmood et al., 15 Sep 2025). The network therefore follows the standard U-shaped pattern while replacing deeper or more expensive contextual mechanisms with compact attention modules chosen for low compute and high spatial sensitivity.

3. Retinal LFRA-Net architecture and attention mechanisms

The retinal model uses multiscale convolution blocks throughout the encoder and decoder. The module is defined as (Mehmood et al., 15 Sep 2025):

$$C_{ms} = C^{1\times 1}(I) \oplus C^{3\times 3}(I) \oplus C_{di}^{3\times 3}(I), \tag{1}$$

followed by

$$C_1 = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C_{ms}))\big), \tag{2}$$

$$MS\_Conv(I) = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C^{3\times 3}(C_1)))\big). \tag{3}$$

Here, $1 \times 1$ convolution performs channel mixing, standard $3 \times 3$ convolution extracts local spatial structure, and dilated $3 \times 3$ convolution expands the receptive field. The paper interprets this design as enabling the model to capture fine vessel boundaries and broader context without deepening the network.

The encoder comprises three multiscale blocks and $2 \times 2$ downsampling convolutions (Mehmood et al., 15 Sep 2025):

$512 \times 512$0

$512 \times 512$1

$512 \times 512$2

$512 \times 512$3

$512 \times 512$4

$512 \times 512$5

The bottleneck contains Focal Modulation Attention (FMAM), inserted as

$512 \times 512$6

where $512 \times 512$7 denotes the focal modulation attention module (Mehmood et al., 15 Sep 2025). Instead of explicit pairwise self-attention, FMAM applies hierarchical depth-wise convolutions to $512 \times 512$8,

$512 \times 512$9

followed by global context aggregation

$I_{\text{out}}$0

and gated fusion

$I_{\text{out}}$1

The final modulation is

$I_{\text{out}}$2

The paper contrasts this with conventional self-attention by noting that self-attention has $I_{\text{out}}$3 complexity, whereas FMAM is closer to $I_{\text{out}}$4 with small kernel sizes (Mehmood et al., 15 Sep 2025). In the retinal setting, FMAM is intended to inject global vascular context into decoder features while avoiding the computational cost of full pairwise attention.

The Region-Aware Attention Mechanism (RAAM) is applied only to early skip connections, specifically $I_{\text{out}}$5 and $I_{\text{out}}$6 (Mehmood et al., 15 Sep 2025):

$I_{\text{out}}$7

$I_{\text{out}}$8

For a skip tensor $I_{\text{out}}$9, RAAM computes

$$C_{ms} = C^{1\times 1}(I) \oplus C^{3\times 3}(I) \oplus C_{di}^{3\times 3}(I), \tag{1}$$0

then successive max and average pooling,

$$C_{ms} = C^{1\times 1}(I) \oplus C^{3\times 3}(I) \oplus C_{di}^{3\times 3}(I), \tag{1}$$1

$$C_{ms} = C^{1\times 1}(I) \oplus C^{3\times 3}(I) \oplus C_{di}^{3\times 3}(I), \tag{1}$$2

feature fusion,

$$C_{ms} = C^{1\times 1}(I) \oplus C^{3\times 3}(I) \oplus C_{di}^{3\times 3}(I), \tag{1}$$3

cross-channel semantic averaging,

$$C_{ms} = C^{1\times 1}(I) \oplus C^{3\times 3}(I) \oplus C_{di}^{3\times 3}(I), \tag{1}$$4

attention-map computation,

$$C_{ms} = C^{1\times 1}(I) \oplus C^{3\times 3}(I) \oplus C_{di}^{3\times 3}(I), \tag{1}$$5

and feature reweighting,

$$C_{ms} = C^{1\times 1}(I) \oplus C^{3\times 3}(I) \oplus C_{di}^{3\times 3}(I), \tag{1}$$6

The output layer is

$$C_{ms} = C^{1\times 1}(I) \oplus C^{3\times 3}(I) \oplus C_{di}^{3\times 3}(I), \tag{1}$$7

with $$C_{ms} = C^{1\times 1}(I) \oplus C^{3\times 3}(I) \oplus C_{di}^{3\times 3}(I), \tag{1}$$8 denoting the sigmoid function (Mehmood et al., 15 Sep 2025). The selective placement of RAAM on early, high-resolution skips is justified in the paper by the claim that these paths retain thin-vessel detail, whereas later skips are more semantic and offer diminishing returns for attention.

4. Retinal LFRA-Net training protocol, benchmarks, and results

The retinal model is trained with a weighted Dice loss to address vessel–background imbalance (Mehmood et al., 15 Sep 2025):

$$C_{ms} = C^{1\times 1}(I) \oplus C^{3\times 3}(I) \oplus C_{di}^{3\times 3}(I), \tag{1}$$9

where $$C_1 = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C_{ms}))\big), \tag{2}$$0 is the predicted mask and $$C_1 = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C_{ms}))\big), \tag{2}$$1 is the ground-truth mask. The optimizer is Adam, the learning rate is 0.002, and the batch size is 8. Training and implementation are reported in TensorFlow/Keras, using an NVIDIA RTX A4000 (16 GB GDDR6 VRAM) (Mehmood et al., 15 Sep 2025).

Evaluation is conducted on three standard retinal vessel benchmarks: DRIVE, STARE, and CHASE_DB (Mehmood et al., 15 Sep 2025). All images are resized to $$C_1 = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C_{ms}))\big), \tag{2}$$2. The data protocol includes augmentation with rotations ($$C_1 = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C_{ms}))\big), \tag{2}$$3) and contrast modifications, implemented with CLoDSA and imgaug. Augmented dataset sizes are reported as 1080 for DRIVE, 1024 for STARE, and 1080 for CHASE_DB (Mehmood et al., 15 Sep 2025).

The reported model complexity is central to the paper’s positioning:

Measure	LFRA-Net
Parameters	0.17 M
FLOPs	10.50 G
Model size	0.66 MB

For comparison, the paper reports U-Net at 7.76 M parameters, 96.68 G FLOPs, 29.60 MB, U-Net++ at 9.04 M, 238.52 G, 34.49 MB, Attention U-Net at 9.25 M, 371.68 G, 35.33 MB, IterNet at 13.60 M, 194.40 G, 94.70 MB, and FS-UNet at 0.87 M, 47.60 G, 3.50 MB (Mehmood et al., 15 Sep 2025). The paper also notes that LMBiS-Net has the same parameter count as LFRA-Net, 0.17 M, but higher FLOPs and lower accuracy.

Reported segmentation performance is as follows (Mehmood et al., 15 Sep 2025):

Dataset	Dice	Jaccard
DRIVE	84.28%	72.86%
STARE	88.44%	79.31%
CHASE_DB	85.50%	74.70%

The same source reports sensitivity and specificity of 82.43% and 98.08% on DRIVE, 88.75% and 98.56% on STARE, and 84.36% and 98.20% on CHASE_DB (Mehmood et al., 15 Sep 2025). The paper states that Dice and Jaccard are consistently higher than those of U-Net, U-Net++, Attention U-Net, MultiRes-UNet, FR-UNet, SegNet, IterNet, OCE-Net, MAGF-Net, DCNet, FS-UNet, G-Net Light, and LMBiS-Net.

The ablation study on DRIVE isolates the effect of multiscale convolutions, skip connections, RAAM, and FMAM (Mehmood et al., 15 Sep 2025). Notable configurations include LU-NS with 0.07 M parameters and 80.30% Dice, MLU with 0.10 M and 81.32% Dice, MLU + FMAM in bottleneck only with 0.15 M and 83.61% Dice, and the final LFRA-Net with 0.17 M parameters, 84.28% Dice, 72.86% Jaccard, 96.09% accuracy, 82.43% sensitivity, and 98.09% specificity. The paper concludes that FMAM provides the largest single gain and that RAAM in early skips contributes additional improvement, especially in Jaccard. This suggests a division of labor in which bottleneck focal modulation chiefly captures global context, while skip-level region-aware attention mainly sharpens fine-structure recovery.

5. LFRA-Net in open-set radio frequency fingerprinting

In the wireless-authentication context, LFRA-Net denotes a lightweight framework for open-set physical-layer authentication based on Rapid LoRA Aggregation (RLA) (Zhang et al., 14 Apr 2026). The underlying problem is that radio frequency fingerprints (RFFs) arise from device-specific hardware distortions, but the received signal is also modified by the wireless channel. The paper formalizes transmission as

$$C_1 = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C_{ms}))\big), \tag{2}$$4

where $$C_1 = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C_{ms}))\big), \tag{2}$$5 denotes hardware impairments and $$C_1 = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C_{ms}))\big), \tag{2}$$6 denotes channel effects (Zhang et al., 14 Apr 2026). An RFF extractor maps the received signal to an embedding,

$$C_1 = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C_{ms}))\big), \tag{2}$$7

The target setting is open-set authentication, where many deployment devices are unknown at training time and channel conditions vary across location, LoS/NLoS status, mobility, aging, and SNR (Zhang et al., 14 Apr 2026). Authentication uses cosine distance,

$$C_1 = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C_{ms}))\big), \tag{2}$$8

with the rule

$$C_1 = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C_{ms}))\big), \tag{2}$$9

(Zhang et al., 14 Apr 2026). Robust performance therefore requires features that are discriminative across devices while remaining stable under channel variation.

The paper positions existing methods as limited by overfitting to specific channels, imperfect robustness from synthetic augmentation or disentanglement methods, and the high cost of full fine-tuning (Zhang et al., 14 Apr 2026). The proposed framework keeps a base RFF backbone $$MS\_Conv(I) = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C^{3\times 3}(C_1)))\big). \tag{3}$$0, pretrains environment-specific LoRA modules $$MS\_Conv(I) = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C^{3\times 3}(C_1)))\big). \tag{3}$$1, and adapts to a new environment $$MS\_Conv(I) = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C^{3\times 3}(C_1)))\big). \tag{3}$$2 by constructing

$$MS\_Conv(I) = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C^{3\times 3}(C_1)))\big). \tag{3}$$3

The backbones used are ML-RFF and DR-RFF, and the paper explicitly states that LFRA-Net does not change the backbone architecture; it adds LoRA modules to the base models (Zhang et al., 14 Apr 2026).

For a weight matrix $$MS\_Conv(I) = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C^{3\times 3}(C_1)))\big). \tag{3}$$4, the LoRA update is

$$MS\_Conv(I) = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C^{3\times 3}(C_1)))\big). \tag{3}$$5

with $$MS\_Conv(I) = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C^{3\times 3}(C_1)))\big). \tag{3}$$6, $$MS\_Conv(I) = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C^{3\times 3}(C_1)))\big). \tag{3}$$7, and $$MS\_Conv(I) = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C^{3\times 3}(C_1)))\big). \tag{3}$$8 (Zhang et al., 14 Apr 2026). The adapted layer output is

$$MS\_Conv(I) = \mathrm{LeakyReLU}\big(Dr^{(0.5)}(\mathrm{BN}(C^{3\times 3}(C_1)))\big). \tag{3}$$9

The paper evaluates ranks $1 \times 1$0 and reports that performance peaks at $1 \times 1$1, while $1 \times 1$2 causes slight overfitting (Zhang et al., 14 Apr 2026).

Environments are defined by channel conditions such as location, LoS/NLoS, room separation, device aging, and SNR differences (Zhang et al., 14 Apr 2026). The dataset contains 59 ZigBee CC2530 devices, a USRP N210 receiver, 2.4 GHz, 19 dBm transmit power, and a sampling rate of 10 MS/s. Each preamble contains $1 \times 1$3 samples. Training and validation use devices 1–45 under LoS conditions at 0.3–1 m and SNR $1 \times 1$4 dB; evaluation spans six channel environments, T1–T3 and U1–U3, including new devices, device aging, NLoS with one or two rooms of separation, and long-range LoS at 40 m (Zhang et al., 14 Apr 2026).

6. Rapid LoRA aggregation, optimization, and empirical performance

The core of the wireless LFRA-Net is the aggregation stage. Instead of learning a fresh LoRA module for a new environment, the method forms

$1 \times 1$5

and optimizes the scalar coefficients $1 \times 1$6 through

$1 \times 1$7

where $1 \times 1$8 is a small labeled set from the new environment (Zhang et al., 14 Apr 2026). The paper uses 20% of a test set for adaptation and the remaining 80% for evaluation.

Optimization is performed by Covariance Matrix Adaptation Evolution Strategy (CMA-ES) rather than gradient descent (Zhang et al., 14 Apr 2026). The method samples candidates

$1 \times 1$9

with hyperparameters

$3 \times 3$0

initial step size $3 \times 3$1, and a maximum of 20 iterations (Zhang et al., 14 Apr 2026). The paper emphasizes that this is a very low-dimensional optimization because $3 \times 3$2 in the experiments.

Performance is evaluated using AUC and Equal Error Rate (EER). EER is defined at the threshold $3 \times 3$3 where the false accept rate equals the false reject rate:

$3 \times 3$4

(Zhang et al., 14 Apr 2026). The paper reports the following base-model performance without adaptation: ML-RFF achieves AUC = 97.11% and EER = 8.22%, while DR-RFF achieves AUC = 98.87% and EER = 3.92% (Zhang et al., 14 Apr 2026).

For the proposed RLA-based LFRA-Net, the main average results across test sets are (Zhang et al., 14 Apr 2026):

Backbone / method	AUC	EER
ML-RLA ($3 \times 3$5)	98.94 ± 0.88%	4.33 ± 2.33%
DR-RLA ($3 \times 3$6)	99.19 ± 0.90%	3.32 ± 2.88%

The paper states that all LoRA variants outperform full fine-tuning and SSL-FT on the ML-RFF backbone, and that RLA achieves best AUC and lowest EER (Zhang et al., 14 Apr 2026). For the stronger DR-RFF backbone, naive full fine-tuning and some LoRA variants can degrade performance, whereas DR-RLA improves beyond the non-adapted base model. Relative to non-finetuned DR-RFF with EER = 3.92%, DR-RLA lowers EER to 3.32%, described as an approximately 15% relative reduction (Zhang et al., 14 Apr 2026).

The efficiency claim is equally central. The paper reports an 83% decrease in training time relative to full fine-tuning, using the same training dataset, and further states that RLA requires only 1/6 of the training time to achieve comparable performance to a strong DR-FT baseline (Zhang et al., 14 Apr 2026). This follows from the fact that, at deployment, the method trains only the mixing coefficients $3 \times 3$7 while keeping the backbone and pretrained LoRA modules frozen.

The wireless paper identifies several limitations: the need for environment-specific pretraining, manual definition or grouping of environments, possible degradation under extreme unseen conditions that differ strongly from all $3 \times 3$8, and growing complexity in LoRA module pool management as the number of environments increases (Zhang et al., 14 Apr 2026). It also points toward extensions such as online learning of new LoRA modules, more sophisticated environment detection and weighting, integration with higher-layer security mechanisms, and application to OFDM, LTE, 5G, and federated or privacy-preserving settings.

Taken together, the two uses of LFRA-Net illustrate a broader pattern in contemporary model design: both prioritize lightweight adaptation or attention mechanisms to improve performance under resource constraints, but they do so in entirely different domains and with distinct technical substrates. In retinal imaging, LFRA-Net is a compact attention-augmented U-shaped segmenter (Mehmood et al., 15 Sep 2025). In wireless authentication, LFRA-Net is a LoRA aggregation framework for channel-robust open-set RFF extraction (Zhang et al., 14 Apr 2026).