arXiv:2604.15936v1 [cs.NI] 17 Apr 2026
Federated Parameter-Efficient Adaptation for Interference Mitigation at the Wireless Edge Evar Jones
Daniel J. Jakubisin
Sanmay Das
Virginia Tech National Security Institute USA [email protected]
Virginia Tech National Security Institute USA [email protected]
Virginia Polytechnic Institute and State University USA [email protected]
Abstract
an increasingly critical challenge [16]. Interference suppression is commonly addressed through signal (source) separation–based mitigation techniques, which aim to recover the signal of interest (SOI) with high fidelity to support reliable detection, demodulation, and decoding [11]. Data-driven methods based on deep learning (DL) have demonstrated superior performance for source separation over traditional methods [7, 10, 25]. However, the traditional centralized learning methods may no longer be feasible due to privacy concerns, computational burden, and communication overhead [23]. Additionally, the rapid changing wireless environment requires online adaptation [27]. Federated learning (FL) is a distributed ML paradigm in which multiple clients collaboratively train a shared model by computing updates locally on their private data and periodically aggregating these updates at a coordinating server, without transferring raw data off the devices. This approach decouples model training from centralized data collection, reducing communication overhead and privacy risks while enabling learning from distributed data sources [12]. Various works have explored FL for online adaptation at the physical (PHY)-layer, including for neural receivers [27] to perform well in different channel environments, channel estimation [3], and modulation classification [19]. FL for RF interference suppression remains largely unexplored, despite the task being inherently distributed across base stations that each encounter distinct local interference environments. Meanwhile, the Open RAN paradigm disaggregates traditionally monolithic, vendor-proprietary RAN equipment into modular components, enabling data-driven network operation [17] and the capability of training ML models [23]. This architectural shift exposes per-site signal data at distributed base stations (gNB in 5G) capable of localized learning, creating a natural setting for federated learning across the RAN [1]. However, gNBs may have limited resources [14], making full model fine-tuning and transmission costly at the wireless edge. Federating all model parameters each round strains bandwidth-limited backhaul links, while local backpropagation through the entire model demands compute and memory that may exceed gNB capacity.
Dense wireless deployments face co-channel interference from heterogeneous sources that vary across base stations (gNBs in 5G). While centralized DNN-based approaches to interference mitigation have shown strong performance, deploying and adapting these models across distributed gNBs via federated learning (FL) requires transmitting full model updates each round, resulting in a cost that scales poorly with network density. Parameter-efficient fine-tuning (PEFT) reduces this burden by training and communicating only a small fraction of parameters. While traditionally applied to large foundation models, we adapt Low-Rank Adaptation (LoRA) to temporal convolutional neural network architectures for interference suppression, placing low-rank adapters on the dilated convolutional layers. This placement enables LoRA to learn local interference-specific temporal patterns, while the frozen backbone retains the shared signal extraction capability. These lightweight adapters (5.1% of backbone parameters) are federated via FedAvg, reducing per-round communication by up to 20× compared to federating full model updates. We evaluate various PEFT strategies across simulated distributed gNBs with non-IID interference environments. Results show that local LoRA achieves 12.8% average BER improvement over the frozen backbone, while Fed-LoRA achieves comparable performance (12.6%). FedLoRA outperforms local adaptation on data-starved nodes where federated knowledge transfer compensates for limited samples, all while avoiding the catastrophic degradation observed with full-model FedAvg under heterogeneous conditions.
Keywords Federated Learning, Parameter-Efficient Fine-Tuning, LoRA, Interference Mitigation
1
Introduction
The densification of wireless networks toward 6G intensifies co-channel interference from heterogeneous sources, making robust interference suppression at the physical layer 1
Parameter-efficient fine-tuning (PEFT) addresses both bottlenecks: by training and communicating only a small set of adapter parameters, it reduces local training cost and per-round communication simultaneously [6]. Among PEFT methods, Low-Rank Adaptation (LoRA) [9] is particularly well-suited to this setting. LoRA augments frozen pretrained weights with low-rank decompositions, training only compact adapter matrices that can be transmitted directly between gNBs and the aggregation server without reconstruction or post-hoc compression. There is a lack of lightweight adaptation strategies for federated interference suppression at the wireless edge that simultaneously reduces local training cost and communication overhead while preserving the backbone’s learned signal representations. In this work, we apply LoRA to the dilated convolutional layers of a WaveNet-based signal separation model, and evaluate its integration within a FL framework in the context of O-RAN systems for interference suppression. In our framework, distributed gNBs act as federated clients that collaboratively adapt a shared signal separation backbone by training and exchanging only lightweight LoRA adapters, orchestrated by a central aggregation server. By restricting finetuning to low-rank modules on the interference-sensitive dilated convolutional layers, our approach simultaneously reduces local training cost, memory footprint, and per-round communication. This approach enables efficient federated adaptation across resource-constrained gNBs. The contributions of this paper are summarized as follows: (1) We apply LoRA to dilated convolutional layers of a WaveNet-based signal separation backbone, enabling nodespecific adaptation of temporal interference filtering while training and communicating only 5.1% of the model parameters, yielding a 20× reduction in per-round communication cost compared to full-model federated averaging. (2) We provide an empirical study across 𝐾 number of gNBs with heterogeneous interference environments under two non-IID regimes: (1) balanced, where all gNBs have equal data volume but heterogeneous interference types, and (2) imbalanced, where data-scarce nodes have limited exposure to rare interference sources. We show that local LoRA applied to the dilated convolutional layers captures 90% of full fine-tuning gains, while federated LoRA outperforms local adaptation on data-starved nodes and unseen interference, all while avoiding the catastrophic degradation observed with FedAvg.
filtering, and adaptive methods such as successive interference cancellation (SIC), which face limitations in adaptability and scalability in dynamic RF environments [16]. Deep learning approaches have emerged as model-free alternatives that learn interference characteristics directly from data, achieving superior performance across diverse interference conditions. The ICASSP 2024 RF Challenge [11] established a benchmark for data-driven single-channel signal separation, demonstrating that deep learning architectures such as WaveNet and U-Net achieve orders-of-magnitude improvement in MSE and BER over traditional baselines including matched filtering and LMMSE across multiple interference types. Tian et al. [25] introduced learnable dilation parameters into the WaveNet architecture to adaptively modulate the receptive field for RF signal processing, achieving optimal signal separation. Naseri et al. [13] developed a U-Net operating in the time-frequency domain via STFT, exploiting OFDM signal grid structure for a 63% average MSE improvement. Henneke [7] extended the WaveNet baseline with SOI-matched autoencoders that learn the demodulation and re-synthetization of the signal of interest. The WaveNet architecture, originally proposed for speech generation [26] and later adapted for speech denoising [20], employs dilated causal convolutions with exponentially growing receptive fields, enabling efficient multi-scale temporal processing well-suited to RF waveforms. Naseri et al. [14] explored architectural compression techniques including depthwise separable convolutions and quantization to enable high-throughput interference cancellation on edge devices, demonstrating the need for computationally efficient signal separation models in resource-constrained deployments. However, these approaches are trained and deployed centrally, assuming access to all interference types during training. In practice, gNBs may encounter heterogeneous and location-specific interference environments, motivating distributed fine-tuning strategies that can tailor a pre-trained backbone to local conditions.
2.2
FL in O-RAN
Federated Learning (FL), introduced by McMahan et al. [12], enables distributed model training by exchanging model updates rather than raw data, addressing both privacy and bandwidth concerns. The O-RAN architecture disaggregates the radio access network into modular components, including the O-RU, O-DU, and O-CU. These components are managed by near-real-time and non-real-time RICs that enable data-driven optimization through open interfaces [1]. In the O-RAN architecture, where disaggregated gNBs are managed by the RICs over bandwidth-limited backhaul links, FL has emerged as a natural paradigm for collaborative intelligence at the wireless edge. Recent works have applied FL
2 Background and Motivation 2.1 DL-based RF Signal Separation Traditional interference suppression techniques rely on domain knowledge of the interference characteristics, matched 2
to diverse O-RAN tasks including resource allocation and network slicing [28], anomaly detection within digital twin security architectures [21], and joint task offloading with fronthaul routing [15]. However, deploying FL efficiently in O-RAN remains challenging. Singh and Nguyen [23] demonstrated that the stringent latency requirements and limited compute resources of RICs necessitate communication-efficient FL methods, proposing compressed gradient techniques to reduce backhaul overhead. More fundamentally, Zhao et al. [29] show that non-IID data distributions across wireless nodes cause severe performance degradation in standard FedAvg, proposing federated data sharing and weight divergence bounds to mitigate this effect. In our setting, non-IID heterogeneity arises not from label distribution skew but from physically distinct interference environments across gNBs. Additionally, FL for interference suppression in an O-RAN context is largely unexplored, despite it being critical for downstream PHY-layer tasks.
2.3
[5] compresses full model updates via orthogonal subspace superposition. FAH-QLoRA [4] dynamically adapts LoRA ranks across FL rounds and assigns heterogeneous ranks to devices, reducing training time by up to 45% on LLM finetuning tasks. In this work, we show that standard federated averaging of LoRA parameters is sufficient for our setting, achieving strong performance without specialized aggregation or compression techniques. To date, federated LoRA has been studied primarily for large language models and vision transformers. No prior work has applied federated LoRA to dilated convolutional architectures or to RF signal separation, where the adapter design must preserve the per-layer dilation structure critical for multi-scale temporal filtering.
3 System Design 3.1 System Architecture We consider a disaggregated O-RAN deployment consisting of 𝐾 gNBs, each comprising an O-RU for RF front-end processing and an O-DU for baseband computation. A Near-RT RIC serves as the central coordination point, connected to all gNBs via bandwidth-limited midhaul/backhaul interfaces. The signal separation model executes at the gNB, operating directly on time-domain IQ baseband samples. The NearRT RIC acts as the federated aggregation server, collecting adapter parameter updates from participating or clustered gNBs and distributing the aggregated global adapter each communication round. This deployment is illustrated in Fig. 2. Each gNB 𝑘 ∈ {1, . . . , 𝐾 } operates in a distinct RF environment characterized by a local interference profile. In practice, a gNB near an industrial facility may experience electromagnetic interference (EMI), while another in a dense urban area may encounter co-channel communication signals. This spatial heterogeneity produces a naturally non-IID data distribution across the network, where the non-IID structure arises from physically distinct interference environments rather than artificial label skew.
Parameter-Efficient Fine-Tuning
Parameter-Efficient Fine-Tuning (PEFT) methods adapt pretrained models to downstream tasks by training only a small subset of parameters while keeping the backbone frozen [6]. Houlsby et al. [8] introduced adapter modules for NLP, demonstrating near full fine-tuning performance with 3.6% of the parameters. LoRA [9] further simplifies adaptation by injecting trainable low-rank decompositions into frozen weight matrices, avoiding additional inference latency. FiLM [18] provided an even lighter alternative through channelwise affine transformations, and has been applied to federated few-shot learning with fewer than 1% of updateable parameters [22]. While PEFT originated in NLP and vision transformers, recent work has extended LoRA to convolutional architectures. Conv-LoRA [30] integrates convolutional experts into LoRA’s bottleneck structure for image segmentation, and LoRA-C [2] proposes layer-wise low-rank decomposition for CNN fine-tuning on resource-constrained IoT devices. However, these methods target spatial convolutional architectures for vision tasks. Dilated convolutional networks, which capture temporal patterns across exponentially growing time scales, require adapters that preserve the per-layer dilation structure. This is a design consideration absent from prior work and essential for RF interference suppression. In federated settings, LoRA’s compact parameterization naturally reduces communication cost, as only the low-rank adapter matrices are transmitted rather than full model updates. Several works have proposed refinements to federated LoRA. FFA-LoRA [24] freezes the A matrices to simplify aggregation and halve communication cost, while CoCo-Fed
3.2
Signal Model
At each gNB 𝑘, the received baseband signal is modeled as: 𝑦𝑘 (𝑡) = 𝑠 (𝑡) + 𝑖𝑘 (𝑡),
𝑡 = 1, . . . ,𝑇 ,
(1)
where 𝑠 (𝑡) ∈ C is the signal of interest (SOI) and 𝑖𝑘 (𝑡) ∈ C is the node-specific interference. Following the RF Challenge convention [11], we use signal-to-interference-plus-noise ratio (SINR) as the interference signals are derived from recorded RF captures that inherently contain additive noise. Each interference signal is scaled to achieve a target SINR and undergoes a random phase rotation before being added to the SOI. During training, the SINR is drawn uniformly at random from [−10, +10] dB; at test time, performance is evaluated at 3
Round 𝑡 : distribute 𝜃¯ (𝑡 ) → local train → upload Δ𝜃 𝑘 → aggregate
Near-RT RIC Aggregation Server
residual + skip
FedAvg on LoRA parameters
Residual Block 𝑖 1×1 Conv (frozen)
Δ𝜃𝑘
𝜃¯ (𝑡 ) 𝜎 (·) ⊙ tanh(·) +
gNB1
gNB2
gNB3
gNB𝐾
Frozen WaveNet
Frozen WaveNet
Frozen WaveNet
Frozen WaveNet
Dilated Conv (frozen)
LoRA Adapter
LoRA Adapter
LoRA Adapter
CommSignal2
CommSignal3
CS2 + EMI
···
B𝑖 A𝑖
x
LoRA Adapter
Frozen
rank 𝑟 Trainable
EMISignal1
Heterogeneous interference environments (non-IID)
Figure 1: Federated LoRA framework for RF interference mitigation within an O-RAN deployment. The Near-RT RIC serves as the aggregation server, coordinating federated learning across 𝐾 distributed gNBs. Each gNB trains lightweight LoRA adapters on its local interference data while the WaveNet backbone remains frozen. Right: LoRA placement on the dilated convolutional layer within each residual block. 11 discrete levels from −10 to +10 dB in 2 dB increments. The SOI is an OFDM signal with QPSK modulation, consisting of 𝑇 = 40,960 complex-valued samples at a 20 MHz sampling rate. Each signal spans 512 OFDM symbols carrying 𝐵 = 57,344 uncoded information bits. The interference 𝑖𝑘 (𝑡) is drawn from a node-specific distribution I𝑘 , which may include communication signals (CommSignal2, CommSignal3), electromagnetic interference (EMISignal1), or mixtures thereof. The SINR is defined as the ratio of SOI power to interference-plus-noise power: SINR = 10 log10
𝑃𝑠 𝑃𝑖
[dB],
local objective at gNB 𝑘 is to minimize the mean squared error between the estimated and true SOI over its local dataset D𝑘 : ∑︁ 1 L𝑘 (𝜃 ) = ∥ 𝑓𝑊 ,𝜃 (𝑦) − 𝑠 ∥ 2, (3) |D𝑘 | (𝑦,𝑠 ) ∈ D𝑘
where the backbone 𝑊 is frozen after centralized pretraining and only the adapter parameters 𝜃 are optimized. This separation reflects the assumption that the backbone has learned general signal extraction capabilities from diverse training data, while the adapters capture node-specific interference characteristics.
(2)
3.4
where 𝑃𝑠 = 𝑇1 ∥𝑠 ∥ 2 and 𝑃𝑖 = 𝑇1 ∥𝑖𝑘 ∥ 2 denote the average signal and interference power, respectively. The objective is to recover an estimate 𝑠ˆ(𝑡) from the mixture 𝑦𝑘 (𝑡), from which information bits are obtained via OFDM demodulation and QPSK demapping. Performance is measured by the bit error rate (BER) between the estimated and true bit sequences.
3.3
Federated Adapter Aggregation
In the federated setting, the global objective is to find adapter parameters that minimize the weighted average of local losses across all 𝐾 gNBs: min L (𝜃 ) = 𝜃
𝐾 ∑︁ |D𝑘 | Í L𝑘 (𝜃 ). 𝑗 |D 𝑗 |
(4)
𝑘=1
We solve this via federated averaging (FedAvg) [12] restricted to the adapter parameters 𝜃 . At each communication round 𝑡:
Problem Formulation
Let 𝑓𝑊 ,𝜃 denote the signal separation model parameterized by backbone weights 𝑊 and adapter parameters 𝜃 . The model maps a mixture input to an estimated SOI: 𝑠ˆ = 𝑓𝑊 ,𝜃 (𝑦). The
(1) The aggregation server distributes the current global adapter 𝜃¯(𝑡 ) to all 𝐾 gNBs. 4
(2) Each gNB 𝑘 initializes its local adapter from 𝜃¯(𝑡 ) and performs 𝐸 epochs of local SGD on D𝑘 , producing updated parameters 𝜃 𝑘(𝑡 ) . (3) The server aggregates local adapters via weighted averaging:
𝜃¯(𝑡 +1) =
𝐾 ∑︁ |D𝑘 | Í 𝜃 𝑘(𝑡 ) . 𝑗 |D 𝑗 |
4 Adapter Design for Dilated Convolutional Networks 4.1 WaveNet Architecture The backbone signal separation model is based on the WaveNet architecture [26], originally proposed for speech generation and subsequently adapted for speech denoising [20]. The model operates directly on time-domain IQ samples, taking the real and imaginary components of the mixture signal 𝑦𝑘 (𝑡) as a 2-channel input and producing a 2-channel output representing the estimated SOI 𝑠ˆ(𝑡). The architecture consists of an input projection, a stack of 𝑅 = 15 residual blocks, and an output projection. The input projection is a 1 × 1 convolution that maps the 2-channel input to 𝐶 = 48 residual channels. Each residual block 𝑖 contains two convolution layers: • A dilated convolution (kernel size 3, dilation 𝑑𝑖 = 2𝑖 mod 𝑚 with cycle length 𝑚 = 5) that maps 𝐶 → 2𝐶 channels, performing temporal filtering at an exponentially growing receptive field scale. • A 1 × 1 output projection that maps 𝐶 → 2𝐶 channels, performing channel mixing. The 2𝐶 output of the dilated convolution is split into two halves and passed through a gated activation:
(5)
𝑘=1
Crucially, only 𝜃 is transmitted each round, not the full model {𝑊 , 𝜃 }. When 𝜃 consists of LoRA adapter matrices, the communication cost per round per node is |𝜃 |, compared to |𝑊 | + |𝜃 | for standard FedAvg over the full model. In our architecture, |𝜃 | = 14,400 at rank=4 while |𝑊 | = 281,954, yielding a ∼20× reduction in per-round communication. After local training at round 𝑡, each gNB retains its locally adapted model 𝑓𝑊 ,𝜃 (𝑡 ) for inference, while contributing 𝜃 𝑘(𝑡 ) 𝑘 to the global aggregate. This structure implicitly provides personalization: the frozen backbone encodes shared signal structure learned during centralized pretraining, while each node’s locally trained adapter captures its site-specific interference characteristics. The global aggregate 𝜃¯(𝑡 +1) seeds the next round of local adaptation, enabling knowledge transfer across nodes — particularly benefiting data-scarce nodes that have limited local exposure to certain interference types.
3.5
h𝑖 = 𝜎 (g𝑖 ) ⊙ tanh(v𝑖 ),
(6)
where g𝑖 and v𝑖 are the gate and filter components obtained by splitting the dilated convolution output along the channel dimension. The activated output is then passed through the 1 × 1 projection and split into a residual connection and a skip connection: x𝑖 + r𝑖 x𝑖+1 = √ , s𝑖 = skip𝑖 , (7) 2 where r𝑖 and s𝑖 are the residual and skip components. The skip connections from all 𝑅 blocks are summed, normalized √ by 1/ 𝑅, and passed through a final skip projection, ReLU activation, and output projection to produce the 2-channel estimated SOI. The dilation cycle {1, 2, 4, 8, 16} repeats three times across the 15 blocks, giving the model multi-scale temporal receptive fields ranging from 3 to 33 samples per cycle. The full lightweight model contains 281,954 parameters and was pretrained centrally for 151,200 steps on 56,000 signal mixtures (28,000 CommSignal2 + 28,000 CommSignal3) using Adam optimization. Notably, the pretraining data contains no EMI samples, so adaptation to EMI-containing nodes relies entirely on the PEFT methods.
End-to-End Workflow
The complete workflow proceeds in two phases. In the pretraining phase, a WaveNet backbone is trained centrally on a large corpus of signal mixtures spanning multiple interference types. This backbone learns general-purpose signal extraction capabilities and is subsequently distributed to all 𝐾 gNBs as a frozen feature extractor. In the federated adaptation phase, each gNB augments the frozen backbone with lightweight LoRA adapters and participates in 𝑅 rounds of federated learning. At each round, gNBs perform local adapter training on their private interference data for 𝐸 epochs, upload the updated adapter parameters to the Near-RT RIC, and receive the aggregated global adapter to seed the next round. After adaptation, each gNB retains its locally trained adapter for real-time inference on incoming signal mixtures, applying OFDM demodulation to the separated signal to recover information bits. The entire adaptation phase communicates only adapter parameters (|𝜃 | ≪ |𝑊 |), making it feasible over bandwidth-limited backhaul links.
4.2
LoRA on Dilated Convolutions
Each WaveNet residual block contains two convolution layers: a dilated convolution for temporal filtering and a 1 × 1 5
initialized to 𝜸 𝑖 = 1, 𝜷 𝑖 = 0 (identity), preserving the pretrained backbone at initialization. With 𝑅 = 15 blocks and 𝐶 = 48 channels, FiLM adds only 2 × 𝑅 × 𝐶 = 1,440 trainable parameters (0.51% of the backbone). Unlike LoRA, which introduces new temporal filter responses via low-rank convolutions on the dilated layers, FiLM can only rescale and shift existing channel activations. This distinction is important: FiLM modulates what the backbone already computes, while LoRA augments how the backbone filters in time. As we show in Section 6, this limits FiLM’s effectiveness for interference suppression, where adapting temporal filtering behavior is critical.
projection for channel mixing. We apply LoRA to the dilated convolutions rather than the 1 × 1 projections. This design choice is motivated by the role each layer plays in signal separation: dilated convolutions capture temporal interference patterns at multiple time scales, while 1 × 1 convolutions mix channel representations. Since interference mitigation requires adapting the model’s temporal filtering behavior to site-specific interference characteristics, targeting the dilated convolutions allows the adapter to directly modify the temporal receptive field response at each scale. We add a parallel low-rank branch to the dilated convolution in each residual block 𝑖: 𝛼 (8) y𝑖 = DilConv𝑖 (x) + B𝑖 (A𝑖 (x)), 𝑟
5 Experimental Setup 5.1 Dataset and Node Configuration
where A𝑖 ∈ Conv1d is a down-projection with 𝐶 → 𝑟 channels, kernel size 3, and dilation 𝑑𝑖 ; and B𝑖 ∈ Conv1d is a pointwise up-projection with 𝑟 → 2𝐶 channels and kernel size 1. The scaling factor 𝛼/𝑟 controls the magnitude of the adapter’s contribution. Critically, A𝑖 preserves the kernel size and dilation rate 𝑑𝑖 = 2𝑖 mod 𝑚 of the original dilated convolution, ensuring the low-rank branch operates over the same temporal receptive field as the frozen layer it augments. Following standard LoRA initialization [9], A𝑖 is initialized with Kaiming uniform and B𝑖 is initialized to zeros, so that B𝑖 (A𝑖 (x)) = 0 at the start of training. This preserves the pretrained backbone’s behavior before any adaptation occurs. With rank 𝑟 = 4 and 𝐶 = 48 channels, each block contributes 𝑟 × 𝐶 × 3 + 2𝐶 × 𝑟 × 1 = 576 + 384 = 960 parameters for A𝑖 and B𝑖 respectively. Across all 𝑅 = 15 blocks, the total adapter size is 15 × 960 = 14,400 trainable parameters, representing 5.1% of the 281,954 backbone parameters. This compact parameterization is what enables efficient federated communication — each round requires transmitting only 14,400 parameters per node rather than the full model, yielding a ∼20× reduction in per-round communication cost. We evaluate ranks 𝑟 ∈ {2, 4, 8} and find diminishing returns beyond 𝑟 = 4: increasing to 𝑟 = 8 doubles the adapter size but yields only 0.9 percentage points of additional BER improvement, confirming 𝑟 = 4 as an efficient operating point (see Fig. 2).
4.3
Signal mixtures are generated following the ICASSP 2024 RF Signal Separation Challenge protocol [11]. Three interference types are considered: CommSignal2, CommSignal3, and EMISignal1, drawn from recorded RF captures in the challenge dataset. Training data is distributed across 𝐾 = 5 gNBs in a non-IID manner reflecting heterogeneous interference environments. Nodes 1 and 2 encounter only communication interference (CommSignal2 and CommSignal3, respectively), Nodes 3 and 4 face mixed communication and electromagnetic interference, and Node 5 encounters only EMISignal1. Crucially, the centrally pretrained backbone has no EMI exposure, so adaptation to EMI-containing nodes relies entirely on the PEFT methods. We evaluate two non-IID data regimes: • Balanced: Each node has 3,000 training samples drawn from its local interference distribution. Nodes differ in interference type but not data volume. • Imbalanced: EMI samples are reduced to 200 per node (from 1,000 or 3,000) while communication-interference samples remain unchanged, simulating rare interference exposure at the edge. Test sets are constructed at 11 discrete SINR levels from −10 to +10 dB in 2 dB increments. Each node is evaluated on a local test set matching its interference profile, as well as a global test set containing all three interference types.
FiLM Conditioning
As a lightweight baseline, we additionally evaluate Featurewise Linear Modulation (FiLM) [18], which applies a learned channel-wise affine transformation to the residual stream between successive blocks: x𝑖′ = 𝜸 𝑖 ⊙ x𝑖 + 𝜷 𝑖 ,
5.2
Model Training
The WaveNet backbone was pretrained centrally for 151,200 steps on 56,000 signal mixtures (28,000 CommSignal2 + 28,000 CommSignal3) using Adam with a learning rate of 5 × 10−4 and batch size 8 with FP16 mixed precision. The pretrained backbone serves as the frozen feature extractor for all subsequent adaptation methods.
(9)
where 𝜸 𝑖 , 𝜷 𝑖 ∈ R𝐶 are per-block scale and shift parameters and ⊙ denotes channel-wise multiplication. Parameters are 6
Table 1: Training configuration. Avg. BER improvement (%)
Parameter
15
Value Backbone Pretraining
Training steps Training data Learning rate
151,200 56,000 (28k CS2 + 28k CS3) 5 × 10−4 Local Adaptation Epochs 20 Scheduler ReduceLROnPlateau Federated Adaptation Communication rounds 𝑅 10 Local epochs 𝐸 2 Aggregation Weighted FedAvg Shared Batch size 8 Optimizer Adam LR (FedAvg / Full-FT) 10−4 LR (LoRA / FiLM) 10−3 LoRA rank 𝑟 4 FP16 ✓ Train/val split 90/10 Params per round per node FedAvg 281,954 Fed-LoRA (𝑟 =4) 14,400 Fed-FiLM 1,440
Fed-FiLM Fed-LoRA FedAvg
𝑟 =8 𝑟 =4 𝑟 =2
10 1.4k 282k
5
0
103
104
105
Parameters communicated per round
Figure 2: Communication–performance tradeoff for federated methods.
BER by 12.8% while training only 5.1% of the model parameters (14,400 vs. 281,954). Fed-LoRA performs comparably at 12.6%, confirming that federated aggregation of adapter parameters preserves most of the local adaptation quality. Both LoRA variants substantially outperform FiLM-based methods (L-FiLM: 6.3%, Fed-FiLM: 5.9%), which lack the capacity to synthesize new filter responses and are limited to rescaling existing features. Full fine-tuning provides a modest upper bound. Local full fine-tuning achieves 14.3% average improvement, meaning L-LoRA captures approximately 90% of the full finetuning gain with 20× fewer trainable parameters. The gap is largest on EMI-heavy nodes (Node 3: 59.3% vs. 52.0%; Node 5: 76.1% vs. 71.0%), suggesting that EMI suppression benefits from additional model capacity, though LoRA still captures the majority of the gain. FedAvg degrades performance on well-represented nodes. While FedAvg achieves a reasonable 4.6% average improvement, this masks a harmful trade-off: it improves Node 5 (EMI) by 50.1% but degrades Node 1 (CS2) by 152.6%. Aggregating all 281,954 backbone parameters causes catastrophic interference between the heterogeneous node objectives. In contrast, Fed-LoRA improves every node without degrading any, demonstrating that restricting federation to a small adapter subspace avoids this failure mode. CommSignal3 remains a hard interference type. Node 2 (CS3-only) shows negligible improvement across all methods, including full fine-tuning (+0.3%). This suggests that CS3 interference is structurally difficult to suppress given the current backbone architecture, and the bottleneck lies in the model’s representational capacity rather than insufficient adaptation.
We compare six adaptation strategies: (1) Backbone: the frozen pretrained model with no adaptation; (2) FedAvg: standard federated averaging over all 281,954 model parameters; (3) L-FiLM and (4) Fed-FiLM: local and federated FiLM conditioning (1,440 params); (5) L-LoRA and (6) FedLoRA: local and federated LoRA on dilated convolutions (14,400 params); and (7) Full-FT: local full fine-tuning of all parameters as an upper bound. For local adaptation (L-LoRA, L-FiLM, Full-FT), each node trains for 20 epochs with a ReduceLROnPlateau scheduler and early stopping based on validation loss. For federated methods (FedAvg, Fed-LoRA, Fed-FiLM), training proceeds for 𝑅 = 10 communication rounds with 𝐸 = 2 local epochs per round per node. All methods use a 90/10 train/validation split. Table 1 summarizes the key hyperparameters.
6 Results and Analysis 6.1 Balanced Data Regime Table 2 reports the average BER across all 11 SINR levels for each adaptation method evaluated on per-node local test sets. Several findings emerge. Local LoRA achieves the best average BER among all lightweight methods, reducing the backbone’s average 7
Table 2: Average BER across all SINR levels for each adaptation method under balanced dataset. Full-FT fine-tunes all 281,954 parameters; L-LoRA adapts only 14,400. Bold = best adapter method (excluding Full-FT upper bound). Node 1 (CS2) 2 (CS3) 3 (CS2+EMI) 4 (CS3+EMI) 5 (EMI) Avg
Backbone .00447 .12596 .01397 .07476 .02319 .04847
FedAvg .01128 .12716 .01142 .06975 .01157 .04624
L-FiLM .00418 .12596 .01131 .07252 .01304 .04540
Fed-FiLM .00420 .12596 .01175 .07311 .01303 .04561
L-LoRA .00411 .12586 .00670 .06783 .00672 .04225
Fed-LoRA .00414 .12591 .00706 .06828 .00653 .04238
Full-FT .00399 .12558 .00569 .06702 .00554 .04156
Table 3: Average BER under data scarcity (imbalanced regime). Node 1 (CS2) 2 (CS3) 3 (CS2+EMI) 4 (CS3+EMI) 5 (EMI) Avg
Backbone .00447 .12596 .01397 .07476 .02319 .04847
FedAvg .01068 .12627 .01357 .07167 .01632 .04770
CommSignal2 0.14
Full-FT .00397 .12558 .00720 .06800 .01229 .04341
0.030
Backbone L-LoRA Fed-LoRA
BER
0.08 0.06
EMISignal1
Backbone L-LoRA Fed-LoRA
0.12
0.025
0.10
0.020
0.08
BER
0.10
BER
Fed-LoRA .00416 .12593 .00863 .06970 .01237 .04416
CommSignal3 Backbone L-LoRA Fed-LoRA
0.12
0.015
0.06
0.04
0.04
0.02
0.02
0.00
L-LoRA .00411 .12589 .00814 .06856 .01335 .04401
Node 1 (CS2)
Node 2 (CS3)
Node 3 Node 4 (CS2+EMI) (CS3+EMI)
Node 5 (EMI)
0.00
0.010 0.005 Node 1 (CS2)
Node 2 (CS3)
Node 3 Node 4 (CS2+EMI) (CS3+EMI)
Node 5 (EMI)
0.000
Node 1 (CS2)
Node 2 (CS3)
Node 3 Node 4 (CS2+EMI) (CS3+EMI)
Node 5 (EMI)
Figure 3: Per-interference-type BER on the global test set
6.2
Imbalanced Data Regime
Federation provides a clear advantage on the most data-starved node. On Node 5 (200 EMI-only samples), FedLoRA achieves 46.6% improvement versus L-LoRA’s 42.5%—a 4.1 percentage point gap. This is the only node where FedLoRA consistently outperforms L-LoRA, and the advantage is attributable to knowledge transfer: the federated aggregate incorporates EMI-relevant gradient information from Nodes 3 and 4, effectively augmenting Node 5’s limited local data. FedAvg is catastrophic under scarcity. FedAvg degrades Node 1 by 139.1%, consistent with the catastrophic degradation observed in the balanced regime (-152.6%). Its average
To evaluate robustness under data scarcity, we reduce EMI training samples to 200 per node (from 3,000) while keeping CS2 and CS3 volumes unchanged. This simulates a realistic deployment scenario where certain interference types are rarely observed. Table 3 reports the results. L-LoRA remains the strongest method overall (9.2% average improvement), but the gap with Fed-LoRA narrows compared to the balanced regime (0.3pp vs. 0.2pp). Both methods degrade gracefully: L-LoRA’s Node 3 improvement drops from 52.0% to 41.8%, while Fed-LoRA drops from 49.5% to 38.2%. 8
improvement of just 1.6% makes it unsuitable for heterogeneous deployments. Full fine-tuning under scarcity. Local full fine-tuning achieves 10.5% average improvement, outperforming all other methods. On Node 5, Fed-LoRA (46.6%) nearly matches full fine-tuning (47.0%) despite training 20× fewer parameters, showing that federated knowledge transfer can largely compensate for limited local model capacity under data scarcity.
6.3
less improvement, making it Pareto-dominated across the entire rank sweep.
6.5
Fig. 4 shows per-node validation loss across federated communication rounds. Fed-FiLM diverges under federation. Fed-FiLM’s validation loss increases monotonically after the first 1–2 rounds on four of five nodes, with only Node 5 (EMI) showing sustained improvement. With only 1,440 scale-and-shift parameters, FiLM lacks sufficient capacity to reconcile conflicting adaptation directions from heterogeneous nodes during aggregation. Fed-LoRA converges stably. Fed-LoRA stabilizes by round 3–4 across all nodes with no divergence, confirming that 14,400 low-rank parameters provide enough expressiveness for stable federated aggregation. MSE–BER disconnect in FedAvg. FedAvg converges to the lowest validation loss on most nodes, yet produces the worst BER on Node 1. Minimizing aggregate MSE across heterogeneous interference types does not guarantee correct signal recovery for demodulation at individual nodes.
Per-Interference-Type Analysis
Figure 3 disaggregates BER by interference type on the global test set, revealing how each node’s locally adapted model generalizes across interference types it may not have trained on. CommSignal2: Most nodes achieve near-backbone BER, except Node 2 (trained only on CS3), which exhibits 0.120 BER under L-LoRA. Fed-LoRA reduces Node 2’s CS2 BER to 0.024—an 80% reduction—demonstrating that federation transfers CS2 suppression knowledge from other nodes. This is the clearest example of cross-node knowledge transfer in our experiments. CommSignal3: All nodes cluster near the backbone BER (∼0.13), confirming that CS3 is uniformly difficult and neither local nor federated adaptation provides meaningful improvement on this interference type. EMISignal1: Both L-LoRA and Fed-LoRA reduce BER well below the backbone across all nodes. The improvement is largest on nodes that include EMI in their training data (Nodes 3, 4, 5), but even CS-only nodes (1, 2) show substantial EMI suppression, indicating that the backbone already captures partial EMI structure and LoRA refines it.
6.4
Convergence Analysis
7
Conclusion and Future Work
This paper investigated parameter-efficient adaptation strategies for RF interference mitigation in federated wireless deployments. We proposed applying LoRA to dilated convolutional layers of a WaveNet-based signal separation backbone, enabling node-specific adaptation with only 5.1% of the backbone parameters. We evaluated six adaptation strategies— local and federated variants of LoRA, FiLM, and full finetuning—across balanced and imbalanced data regimes on a 5-node heterogeneous interference scenario. Our results demonstrate that L-LoRA is the strongest lightweight method overall, capturing approximately 90% of full fine-tuning’s gain while training 20× fewer parameters. FedLoRA performs comparably in most settings and provides a measurable advantage on data-starved nodes, where crossnode knowledge transfer compensates for limited local observations. In contrast, standard FedAvg over the full backbone suffers catastrophic degradation on well-represented nodes, confirming that restricting federation to a compact adapter subspace is essential for heterogeneous interference environments. The per-interference-type analysis further reveals that federation enables cross-node generalization: Fed-LoRA reduces a CS3-trained node’s CommSignal2 BER by 80% compared to L-LoRA, demonstrating effective transfer of interference suppression capabilities across nodes with non-overlapping data distributions. The communication–performance tradeoff analysis shows that Fed-LoRA at rank 2 already outperforms FedAvg while transmitting 39× fewer parameters per
Communication–Performance Trade-off
Fig. 2 plots average BER improvement plots average BER improvement against per-round communication cost (parameters transmitted per node) for all federated methods across ranks 𝑟 ∈ {2, 4, 8}. Fed-LoRA achieves 11.3% improvement at rank 2 (7,200 params/node/round), increasing to 12.6% at rank 4 (14,400) and 13.5% at rank 8 (28,800). This represents a favorable tradeoff: Fed-LoRA at rank 2 already outperforms both FedAvg (4.6% at 281,954 params) and Fed-FiLM (5.9% at 1,440 params) while transmitting 39× fewer parameters than FedAvg. Diminishing returns are evident: doubling the rank from 4 to 8 yields only 0.9pp additional improvement for Fed-LoRA. L-LoRA follows the same trend (11.7% → 12.8% → 13.7%), consistently outperforming Fed-LoRA by 0.2–0.4pp at each rank, confirming that the federation overhead is small but nonzero. FedAvg occupies an unfavorable position: it transmits 10– 39× more parameters than Fed-LoRA while achieving 2–3× 9
Per-Node Convergence Node 1 (CS2)
Node 2 (CS3)
0.08 0.06
0.1900 0.1850 0.1825 0.1800
5
10 Round
15
20
0.1750
0.06
5
10 Round
15
20
5
10 Round
15
20
Node 5 (EMI) FedAvg Fed-FiLM Fed-LoRA
FedAvg Fed-FiLM Fed-LoRA
0.055 0.050 Val Loss (MSE)
Val Loss (MSE)
0.08
0.04
Node 4 (CS3+EMI)
0.1350 0.1325 0.1300 0.1275 0.1250 0.1225 0.1200 0.1175
FedAvg Fed-FiLM Fed-LoRA
0.10
0.1875
0.1775
0.04
0.12
FedAvg Fed-FiLM Fed-LoRA
Val Loss (MSE)
0.10
Val Loss (MSE)
Val Loss (MSE)
0.12
Node 3 (CS2+EMI)
0.1925
FedAvg Fed-FiLM Fed-LoRA
0.045 0.040 0.035 0.030 0.025
5
10 Round
15
20
5
10 Round
15
20
Figure 4: Per-node validation loss (MSE) across federated communication rounds.
Acknowledgments
round, making it practical for bandwidth-constrained fronthaul links. Convergence analysis further reveals that naïve federation of lightweight adapters (Fed-FiLM) can diverge under non-IID interference distributions, whereas Fed-LoRA maintains stable convergence across all nodes. Several directions remain for future investigation. First, our evaluation uses a simulated federated setting with sequential node training. Real-world deployment and testing on a real O-RAN testbed with asynchronous communication and Near-RT RIC orchestration would validate the practical feasibility of federated adapter aggregation. Second, CommSignal3 proved resistant to all adaptation methods including full fine-tuning, suggesting that architectural modifications to the backbone may be necessary, such as an increased receptive field or attention mechanisms. Third, adaptive rank allocation across nodes or layers could improve efficiency. Nodes with simple interference may require lower rank than mixed-interference nodes, and deeper layers may benefit from higher rank than shallow layers. Finally, extending this framework to additional PEFT methods such as adapter layers or prompt tuning, and to other PHY-layer tasks beyond signal separation (e.g., channel estimation, equalization), would broaden the applicability of federated parameterefficient adaptation at the wireless edge.
This research was supported by the U.S. Department of Commerce’s National Telecommunications and Information Administration (NTIA) under the Public Wireless Supply Chain Innovation Fund Grant Program (Award 24-60-IF2415: ASPEN - Advanced Signal Processing Enhancement for NextGeneration Open Radio Units), administered by the National Institute of Standards and Technology.
References [1] Bouziane Brik, Hatim Chergui, Lanfranco Zanzi, Francesco Devoti, Adlen Ksentini, Muhammad Shuaib Siddiqui, Xavier Costa-Pérez, and Christos Verikoukis. 2025. Explainable AI in 6G O-RAN: A Tutorial and Survey on Architecture, Use Cases, Challenges, and Future Research. IEEE Communications Surveys & Tutorials 27, 5 (2025), 2826–. https: //doi.org/10.1109/COMST.2024.3510543 [2] Chuntao Ding, Xu Cao, Jianhang Xie, Linlin Fan, Shangguang Wang, and Zhichao Lu. 2024. LoRA-C: Parameter-Efficient Fine-Tuning of Robust CNN for IoT Devices. arXiv preprint arXiv:2410.16954 (2024). [3] Ahmet M. Elbir and Sinem Coleri. 2021. Federated Learning for Channel Estimation in Conventional and RIS-Assisted Massive MIMO. IEEE Transactions on Wireless Communications 20, 10 (2021), 6518–6533. https://doi.org/10.1109/TWC.2021.3081836 [4] Zhidong Gao, Zhenxiao Zhang, Yuanxiong Guo, and Yanmin Gong. 2025. Federated Adaptive Fine-Tuning of Large Language Models with Heterogeneous Quantization and LoRA. In IEEE INFOCOM 2025 – IEEE 10
Conference on Computer Communications. IEEE. [5] Zhiheng Guo, Zhaoyang Liu, Zihan Cen, Chenyuan Feng, Xinghua Sun, Xiang Chen, Tony Q. S. Quek, and Xijun Wang. 2025. CoCo-Fed: A Unified Framework for Memory- and Communication-Efficient Federated Learning at the Wireless Edge. arXiv preprint arXiv:2601.00549 (2025). [6] Zhen Han, Ning Ding, Yuxian Wu, Xiang Zhao, Kai Lv, Yixuan Gu, Zhiyuan Liu, and Maosong Sun. 2024. Parameter-Efficient Fine-Tuning for Large Models: A Comprehensive Survey. Comput. Surveys (2024). https://doi.org/10.1145/3638552 [7] Lukas Henneke. 2024. Improving Data-Driven RF Signal Separation with SOI-Matched Autoencoders. In Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW). IEEE, 45–46. https://doi.org/10.1109/ICASSPW62465.2024. 10626245 [8] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-Efficient Transfer Learning for NLP. In Proceedings of the 36th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 97). 2790–2799. [9] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations (ICLR). [10] Hiten Prakash Kothari and R. Michael Buehrer. 2025. Interference Mitigation using U-Net Autoencoder based system. arXiv preprint arXiv:2512.13844 (2025). [11] Alejandro Lancho, Amir Weiss, Gary C. F. Lee, Tejas Jayashankar, Binoy G. Kurien, Yury Polyanskiy, and Gregory W. Wornell. 2025. RF Challenge: The Data-Driven Radio Frequency Signal Separation Challenge. IEEE Open Journal of the Communications Society 6 (2025), 4083–4100. https://doi.org/10.1109/OJCOMS.2025.3556319 [12] H Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. 2017. Communication-Efficient Learning of Deep Networks from Decentralized Data. In Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS), Vol. 54. 1–10. [13] Mostafa Naseri, Jaron Fontaine, Ingrid Moerman, Eli De Poorter, and Adnan Shahid. 2024. A U-Net Architecture for Time-Frequency Interference Signal Separation of RF Waveforms. In Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW). IEEE, 91–92. https://doi.org/10.1109/ ICASSPW62465.2024.10626228 [14] Mostafa Naseri, Eli De Poorter, Ingrid Moerman, H. Vincent Poor, and Adnan Shahid. 2025. High-Throughput Adaptive Co-Channel Interference Cancellation for Edge Devices Using Depthwise Separable Convolutions, Quantization, and Pruning. IEEE Open Journal of the Communications Society 6 (2025), 656–. https://doi.org/10.1109/ OJCOMS.2024.3523797 [15] Anselme Ndikumana, Kim Khoa Nguyen, and Mohamed Cheriet. 2023. Federated Learning Assisted Deep Q-Learning for Joint Task Offloading and Fronthaul Segment Routing in Open RAN. IEEE Transactions on Network and Service Management 20, 3 (2023), 3261–3273. https://doi.org/10.1109/TNSM.2023.3245544 [16] Taiwo Oyedare, Vijay K. Shah, Daniel J. Jakubisin, and Jeffrey H. Reed. 2022. Interference Suppression Using Deep Learning: Current Approaches and Open Challenges. IEEE Access 10 (2022), 66238–66266. https://doi.org/10.1109/ACCESS.2022.3185124 [17] Qingrui Pan, Mahesh K. Marina, and Thanos Triantafyllou. 2025. On NextG Open RAN as a Sensing Infrastructure. In Proceedings of the 26th International Workshop on Mobile Computing Systems and Applications (HotMobile ’25). ACM, CA, La Quinta, USA. https://doi.org/10.1145/
3708468.3711894 [18] Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron Courville. 2018. FiLM: Visual Reasoning with a General Conditioning Layer. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence. 3942–3951. [19] Ratun Rahman and Dinh C. Nguyen. 2024. Improved Modulation Recognition Using Personalized Federated Learning. IEEE Transactions on Vehicular Technology 73, 12 (2024), 19937–19952. https://doi.org/10. 1109/TVT.2024.3447569 [20] Dario Rethage, Jordi Pons, and Xavier Serra. 2018. A WaveNet for Speech Denoising. arXiv preprint arXiv:1706.07162 (2018). [21] Yasintha Rumesh, Dinaj Attanayaka, Pawani Porambage, Jarno Pinola, Joshua Groen, and Kaushik Chowdhury. 2024. Federated Learning for Anomaly Detection in Open RAN: Security Architecture Within a Digital Twin. In Proceedings of the European Conference on Networks and Communications and 6G Summit (EuCNC/6G Summit). 1–6. https: //doi.org/10.1109/EuCNC/6GSummit60053.2024.10597083 [22] Aliaksandra Shysheya, John Bronskill, Massimiliano Patacchiola, Sebastian Nowozin, and Richard E. Turner. 2023. FiT: Parameter-Efficient Few-Shot Transfer Learning for Personalized and Federated Image Classification. In Proceedings of the International Conference on Learning Representations. [23] Amardip Kumar Singh and Kim Khoa Nguyen. 2024. Communication Efficient Compressed and Accelerated Federated Learning in Open RAN Intelligent Controllers. IEEE/ACM Transactions on Networking 32, 4 (2024), 3361–3376. https://doi.org/10.1109/TNET.2024.3384839 [24] Youbang Sun, Zitao Li, Yaliang Li, and Bolin Ding. 2024. Improving LoRA in Privacy-preserving Federated Learning. In Proc. ICLR. [25] Yu Tian, Ahmed Alhammadi, Abdullah Quran, and Abubakar Sani Ali. 2024. A Novel Approach to WaveNet Architecture for RF Signal Separation with Learnable Dilation and Data Augmentation. In Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW). IEEE, 79–80. https: //doi.org/10.1109/ICASSPW62465.2024.10627586 [26] Aäron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. 2016. WaveNet: A Generative Model for Raw Audio. arXiv preprint arXiv:1609.03499 (2016). [27] Shuo Wang, Tianxin Wang, and Xudong Wang. 2025. FedPDA: Collaborative Learning for Reducing Online-Adaptation Frequency of Neural Receivers. In Proceedings of the IEEE International Conference on Computer Communications (INFOCOM). 1–10. https://doi.org/10. 1109/INFOCOM55648.2025.11044747 [28] Han Zhang, Hao Zhou, and Melike Erol-Kantarci. 2022. Federated Deep Reinforcement Learning for Resource Allocation in O-RAN Slicing. In 2022 IEEE Global Communications Conference (GLOBECOM). IEEE, 958–963. https://doi.org/10.1109/GLOBECOM48099.2022.10001658 [29] Zhongyuan Zhao, Chenyuan Feng, Wei Hong, Jiamo Jiang, Chao Jia, Tony Q. S. Quek, and Mugen Peng. 2022. Federated Learning With Non-IID Data in Wireless Networks. IEEE Transactions on Wireless Communications 21, 3 (2022), 1927–1940. https://doi.org/10.1109/TWC. 2021.3108197 [30] Zihan Zhong, Zhiqiang Tang, Tong He, Haoyang Fang, and Chun Yuan. 2024. Convolution Meets LoRA: Parameter Efficient Finetuning for Segment Anything Model. In Proc. ICLR.
11