PERFECT: Personalized Federated Learning for CBRS Radar Detection Shafi Ullah Khan∗ , Madan Baduwal† , Vini Chaudhary † , and Debashri Roy∗
arXiv:2605.03199v1 [cs.NI] 4 May 2026
∗ The University of Texas at Arlington, † Mississippi State University Emails: [email protected], [email protected], [email protected], [email protected]
Abstract—The Citizens Broadband Radio Service (CBRS) band is pivotal for expanding next-generation wireless services, but its success hinges on robustly protecting incumbent users, such as naval radar systems, from interference. This task is delegated to a network of Environmental Sensing Capability (ESC) sensors, which must detect faint radar signals amidst heavy co-channel interference from commercial LTE and 5G users. Traditional centralized detection models raise significant data privacy concerns and are ill-suited for the Non-Independent and Identically Distributed (non-IID) nature of data from geographically dispersed sensors. To overcome these limitations, we propose a novel Federated Learning (FL) framework PERFECT that leverages ESC level personalization for robust and efficient radar detection. PERFECT preserves privacy by training models locally on ESC sensors. Furthermore, our framework is the first to effectively handle non-IID scenarios through model personalization where different ESCs observe distinct radar types. We demonstrate through extensive simulations that PERFECT achieves the mandated 99% recall for radar detection, matching centralized performance while significantly enhancing privacy, efficiency, and scalability for dynamic spectrum sharing. Index Terms—Federated learning, Personalization, Radar interference detection, Shared CBRS spectrum.
I. I NTRODUCTION The 3.5 GHz Citizens Broadband Radio Service (CBRS) band represents a paradigm shift in spectrum management, enabling dynamic sharing between federal incumbent users and new commercial wireless services. This three-tiered access model consists of Incumbent Access users, primarily naval Knowledge radar systems, Priority Access License (PAL) holders, typically deploying LTE networks, and General Authorized Access (GAA) users, which include private 5G networks [1]. The foundational principle of this framework is the absolute protection of incumbent operations, which necessitates a highly reliable and agile mechanism to detect their presence and vacate the channel accordingly. This critical function is performed by a network of Environmental Sensing Capability (ESC) sensors deployed along coastal areas. The Federal Communications Commission (FCC) mandates that this network must detect naval radar signals with extremely high reliability at the Spectrum Access System (SAS) specifically, a 99% recall rate at or above 20 dB Signal to Interference-plus-Noise Ratio (SINR) considering both environmental noise and the surrounding commercial traffic [2]. This presents a formidable technical challenge: the radar waveforms are often transient and have low duty cycles, making them difficult to distinguish from the noise floor, especially when buried under high-power, wideband LTE and 5G transmissions from PAL and GAA users, respectively.
(a) Problem Setting 5G
(b) Problem: Detect Harmful Overlap
Naval Radar (Type 1) ESC1
LTE
Type 1 Radar + Prominent 5G + Overlapped w/ Faded LTE @ ESC1 Local Dataset (D1)
Naval Radar (Type 2) ESC2
(c) PERFECT
Central Server (SAS) Aggregated Global Model
Local Training on D1 & Personalization ESC1
Type 2 Radar + Prominent LTE + Overlapped w/ Faded 5G @ ESC2 Local Dataset (D2)
Model Updates
Local Training on D2 & Personalization ESC2
Fig. 1: An overview of the radar interference detection problem and our proposed solution: (a) The operational scenario in the CBRS band, where a distributed network of ESC sensors must detect incumbent naval radar signals amidst commercial traffic. The data is inherently non-IID, as different sensors observe distinct radar types and interference levels. (b) The classification task based on spectrograms: the model must distinguish between non-harmful scenarios and the critical harmful overlap case. (c) Our proposed privacy-preserving PERFECT framework, where ESC sensors collaboratively train a model coordinated by a central server at Spectrum Access System (SAS). Personalization for radar presence in non-IID environment is performed locally at each sensor to create a robust model update.
A fundamental tension exists in designing such a detection system: the need for data privacy at each ESC sensor versus the need for a globally generalized detection model. On one hand, ESC networks may be operated by different entities, making the sharing of raw, sensitive spectral data with a central entity a significant privacy and security concern. Specifically, raw RF and spectral measurements can inadvertently reveal sensitive operational details such as ESC sensor locations and coverage footprints, as well as patterns of incumbent radar activity, and therefore transmitting them to a centralized server raises both security and regulatory concerns. Training a model exclusively on a sensor’s local data is the most private approach. On the other hand, a model trained only on local data will fail to generalize. The Radio Frequency (RF) environment is inherently non-Independent and Identically Distributed (non-IID); a sensor in one location may observe a specific type of naval radar and a particular mix of commercial traffic, while a sensor a few miles away sees a completely different radar type and interference profile. A locally trained model will overfit to its own unique environment and will be unable to detect unfamiliar radar signatures seen by other
sensors in the network, leading to a catastrophic failure of the protection mandate. This gives rise to two core challenges. First, how can we build a collaborative model that learns from the collective experience of all ESC sensors without ever exposing their private, local datasets? Second, how can this collaborative model be robust enough to provide high-recall detection for every sensor, even when the data distributions across sensors are highly skewed? A naive federated model that simply averages parameters will converge to a single global model that is not specialized enough for any particular sensor, especially those that observe rare but critical radar types. To address these challenges, we propose PERFECT, a novel Federated Learning (FL) framework that incorporates personalization for robust and private radar interference detection. Furthermore, to transition this framework from a theoretical model to a deployable solution, we also explicitly detail the architectural integration of PERFECT into the existing SAS and ESC sensor infrastructure. We expand upon the policy implications of distributed edge sensing, demonstrating how localized model training inherently satisfies the stringent Operational Security (OPSEC) requirements of federal incumbents by ensuring raw spectral data never traverses the network. Finally, to ensure our evaluation accurately reflects the complex, real-world circumstances of radar detection, we significantly extend our methodology beyond basic modeling. We utilize a highly realistic emulation testbed using National Institute of Standard and Technology (NIST) radar waveform profiles [3], multiplexed with 3GPP-compliant commercial signals: LTE as PAL and 5G as GAA users [4]. We rigorously validate PERFECT against the realistic, non-IID interference conditions and variable SINR that physically distributed ESC nodes experience in the field. Our approach resolves the privacy-generalization conflict as: • To ensure privacy, PERFECT adopts a federated learning architecture in which models are trained locally at each ESC sensor; only model parameters are exchanged. • To ensure generalization in non-IID settings, PERFECT introduces federated personalization that allows each sensor’s model to learn a shared, global representation from the collective framework while maintaining specialized personal layers that adapt to its unique local data, thereby enabling the detection of different radar types across the network. Fig. 1 provides a comprehensive overview of our work, illustrating the CBRS operational scenario with its inherent non-IID data challenge, the specific spectrogram-based detection problem, and our proposed federated solution. Our novel contributions are: (1) We design a FL framework for radar detection that preserves the privacy of ESC sensor data. We conduct a rigorous comparative analysis of our proposed approach against traditional centralized and local-only training paradigms, evaluating the trade-offs between detection performance and privacy. (2) We propose a federated personalization based generalization approach, PERFECT, which ensures robustness for
detecting radar interference in non-IID environment. To the best of our knowledge, this is the first work to design and validate a federated system for detecting multiple, distinct types of naval radars in a realistic non-IID environment, where different type of radar distribution varies across the network of ESC sensors. (3) Through extensive experimentation, we validate that PERFECT achieves 99% recall in a federated setting across five ESC sensors along a coastal region, considering two types of NIST generated radar waveforms [3], LTE as PAL, and 5G as GAA users, all non-identically observed at different ESC sensors. We have released the first-of-its-kind dataset, containing distributed ESCs with radar presence in non-IID environment, along with the associated FL codes1 . II. R ELATED W ORK Radar Detection in the CBRS Band. The reliable detection of incumbent radar signals in the CBRS band is a wellestablished problem. Early approaches relied on traditional signal processing techniques such as energy detection and matched filtering [5]. However, these methods often struggle in low SINR conditions and with the complex, non-stationary interference from PAL and GAA users. Recent works [6]–[8] apply deep learning using spectrograms or raw In-phase/Quadrature (IQ) data for radar detection, showing improved performance over the traditional methods. Further, the ESC+ framework in [9] utilizes YOLOv3 with spectrogram inputs to detect and localize CBRS radar signals with over 99% accuracy at SINR levels up to 17 dB, which they latter verify using a real-world dataset in [10]. The work [11] applies localized spectrogram scanning to narrower subbands and reports 99% recall for a single radar class at SINR values as low as 15 dB. SenseORAN framework in [12] develops a YOLO-based xApp integrated into the near-realtime RIC of a 5G gNB, converting uplink IQ data into spectrograms, aggregating multiple time slices per inference, and maintaining a dynamic channel occupancy list to reach 100% detection at SINR ≥ 12 dB. Khan et al. [13] recently propose an in-network fusion framework that combines features from raw IQ samples and spectrograms, achieving nearly 99% accuracy in high-interference environments, thereby highlighting the effectiveness of hybrid models for robust radar sensing. However, a major weakness of these advanced approaches is their reliance on a centralized training paradigm, where all sensor data is collected at a single location. This assumption overlooks the critical real-world constraints of data privacy and generalization. While the FaIR [14] leverages standard FL for radar detection and achieves 80% accuracy, our paper explicitly solves the key challenges of generalization over nonIID data through personalization, which also add another level of privacy to the edge ECS’s data, while achieving a higher detection accuracy of 99%. Communication efficiency and personalization in FL. FL enables privacy-preserving distributed training [15], [16], yet 1 Our dataset is available at https://twistlab.uta.edu/projects/.
vanilla FedAvg is frequently limited by uplink/downlink bandwidth and can underperform with non-IID client data [17]. Furthermore, personalization tackles non-IID drift by keeping a small client-specific component, adding an extra level of privacy to the clients while aggregating a shared backbone [18]. Innovation Opportunity. Radar detection in the CBRS band and FL personalization have been studied separately, but their intersection remains largely unexplored. To our knowledge, no prior work proposes an FL framework tailored to naval radar detection in CBRS. Centralized approaches lack scalability and privacy, and common FL benchmarks overlook RF realities such as strict recall targets and non-IID waveform diversity. We introduce the first FL framework for this setting with personalization, and show it achieves the required 99% recall under non-IID conditions while respecting privacy constraints across the ESC network. III. S YSTEM M ODEL AND P ROBLEM F ORMULATION A. System Model We consider a CBRS environment with a network of K geographically distributed ESC sensors, indexed by k ∈ {1, 2, . . . , K}. These sensors are responsible for detecting incumbent naval radar signals. The network operates under the coordination of a central server, often managed by a SAS. The RF environment observed by each ESC sensor is a superposition of signals from three tiers: 1) Incumbent Access: Naval radar systems. 2) Priority Access License (PAL): LTE-based commercial networks. 3) General Authorized Access (GAA): Private 5G or other commercial networks. Each ESC sensor k captures raw IQ samples and converts them into 2D spectrograms, which serve as the input to our detection model. These spectrograms are treated as RGB images, capturing the time-frequency characteristics of the received signal. In our FL setting, the ESC sensors act as clients that train local detection models on these private spectrograms. The central server acts as the aggregator, coordinating the learning process without ever accessing the raw data, thereby preserving data privacy. B. Signal Model and Detection Problem The signal received in the time domain on the ESC sensor k is yk (t), which is a composite of a potential incumbent radar signal sk (t), commercial traffic from PAL and GAA users ck (t), and Additive White Gaussian noise (AWGN) nk (t). The input to our detection model is a 2D spectrogram, Xk , generated from yk (t) by applying a Short-Time Fourier Transform (STFT). Let Sk (τ, ω) and Ck (τ, ω) be the STFTs of the radar and commercial signals, respectively, where τ is time and ω is frequency. The core task is to classify each spectrogram Xk based on whether a harmful time-frequency overlap between the incumbent and commercial users is occurring. This is formulated as a binary hypothesis test: • H1 (Harmful Overlap): A radar signal and a commercial signal are simultaneously present and their transmissions
overlap in the time-frequency domain. Let Supp(·) denote the support of a signal in the time-frequency plane. This hypothesis is true if: measure(Supp(Sk ) ∩ Supp(Ck )) > ∅ (1) Under this condition, the received signal is yk (t) = sk (t)+ck (t)+nk (t). This is the critical event that requires action to protect the incumbent. • H0 (No Harmful Overlap): This hypothesis covers all other conditions where no harmful overlap exists. This occurs if: measure(Supp(Sk ) ∩ Supp(Ck )) = ∅ (2) This single condition elegantly covers the four nonharmful scenarios: (i) only radar is present (Ck = 0 =⇒ Supp(Ck ) = ∅), (ii) only commercial traffic is present (Sk = 0 =⇒ Supp(Sk ) = ∅), and (iii) both radar and commercial signals are present but their transmissions are disjoint in time and/or frequency (Sk = 1, Ck = 1, however, Supp(Sk ) ∩ Supp(Ck )) = ∅). The non-IID nature of the data across the ESC network remains a key challenge, manifesting as: 1) Feature Skew: The visual characteristics of the PAL/GAA signals and the noise floor in the spectrograms can differ significantly for each sensor k due to its unique location. 2) Label Skew (Non-IID Environment): We consider a scenario with two distinct naval radar types, r1 and r2 . The radar signature appearing in a spectrogram at sensor k can be of either type, and the distribution of these types is not uniform across the ESCs, potentially due to varying, unknown ship-borne radar trajectories and time-of-occurrence patterns. C. Problem Formulation The objective is to design a distributed machine learning model that correctly classifies spectrograms (and thus, detects presence/absense of interference with radar) according to the hypotheses H0 and H1 , achieving a minimum required recall of 99% for the H1 class. Let fw denote a neural network model parameterized by weights w. In our federated setting, each client k (corresponding to ESC sensor k) has a local dataset k Dk = {(Xk,i , lk,i )}ni=1 , where Xk,i is the i-th spectrogram image and lk,i ∈ {0, 1} is the corresponding label, nk is number of samples from each ESCk . The goal of FL is to learn a global model by minimizing a global objective function F(W ), which is the weighted average of the local loss functions Lk of all clients: K X |Dk | Lk (W ), (3) min F(W ) = W |D| k=1 where W represent the parameters of the global model, PK |Dk | is the size of the local dataset at client k, and |D| = k=1 |Dk | is the total dataset size. The local loss function Lk for a classification task is the cross-entropy loss: |Dk | 1 X Lk (wk ) = − (lk,i log(fwk (Xk,i )) |Dk | i=1 + (1 − lk,i ) log(1 − fwk (Xk,i ))) ,
(4)
where fwk is the local model at client k with parameters wk . Our objective is to solve this optimization problem under the critical constraints of privacy and non-IID robustness. D. Proposed Solution Overview To solve the constrained optimization problem formulated in Sec. III-C, we propose a comprehensive framework that integrates FL with personalization. Our solution consists of two main components: • A baseline FL Framework that enables collaborative training while ensuring data privacy (see Sec. IV-A). • Federated Personalization to provide robustness against non-IID data distributions across the ESC sensors (see Sec.IV-B). E. System Architecture and Practical Deployment The integration of the proposed PERFECT framework into the CBRS ecosystem requires an architecture that satisfies the rigorous incumbent protection mandate formulated in Sec. III-C, while adhering to the practical communication and privacy constraints of distributed edge sensors. Building upon the solution overview in Sec. III-D, PERFECT is explicitly mapped onto the standardized SAS-ESC topology to bridge the gap between theoretical federated optimization and physical deployment. 1) Architectural Components: The framework operates across two primary infrastructural tiers: • ESC Edge Nodes: Geographically distributed along coastal boundaries, each ESC node k ∈ {1, 2, · · · , K} acts as an independent FL client. These nodes capture raw IQ samples and generate the local dataset Dk of spectrograms. During local training, the model fwk is bifurcated into a globally shared base network characterized by Nbase parameters, and a private, environment-specific personalized head containing Nhead parameters. • SAS as Central Aggregator: The SAS orchestrates the collaborative learning process by securely aggregating model parameter updates from the distributed ESCs, without ever accessing the raw spectral data Xk,i . 2) Deployment Workflow and Communication Efficiency: Standardizing ML deployment over remote edge sensors introduces significant communication bottlenecks. PERFECT addresses this by utilizing a lightweight neural network model comprising a total of Ntotal parameters, where Ntotal = Nbase +Nhead . The deployment workflow proceeds as follows: 1) Global Broadcast: The SAS initializes and broadcasts the global base model weights WBt to the participating ESC network at communication round t. 2) Edge Personalization: Each ESC node k trains the full model (wkt , vkt ) on its local dataset Dk . The personal head vkt adapts to local interference profiles and specific radar signatures, establishing robustness against the nonIID conditions defined by the feature and label skew. 3) Lightweight Uplink: To conserve bandwidth, ESC nodes decouple their trained models. They transmit only the updated base parameters wkt+1 back to the SAS,
Server-side Processing (at SAS) 5. Aggregate Updates Global Base Model SAS (Central Server)
4. Upload the Base Model
1. Broadcast Global Base Model ESC Sensors (Client)
...
...
Personalization (at each ESC Sensor) Global Base Model
Global Base Model
Personal Head Personal Head
2. Local Training on Private Data
Updated Full Model
3. Seperate the Head and Base Models
Fig. 2: The proposed PERFECT framework. In a single communication round, the central server at the SAS distributes a global base model. For the generalization via personalization, each ESC client performs local training on its private data using a personalized model (shared base + private head). The server aggregates only the shared base model weights.
securely retaining the Nhead personal head parameters on-device. 4) Global Aggregation: The SAS aggregates the received base parameters to formulate an improved global model WBt+1 via the defined weighted averaging objective. By sharing only the backbone network, PERFECT significantly reduces the per-round uplink payload to O(Nbase ) bytes per client. This lightweight footprint is critical for remote ESC deployments where high-bandwidth backhaul may be limited or cost-prohibitive. 3) Policy Implications and Security Compliance: Transmitting raw RF data to a centralized server poses a severe policy risk, potentially exposing sensitive operational parameters such as naval vessel trajectories and ESC node vulnerabilities. PERFECT structurally enforces privacy-by-design, aligning with the stringent OPSEC requirements of federal incumbents. Because raw spectrograms never traverse the network, the attack surface for data interception is fundamentally eliminated. Furthermore, the federated personalization approach ensures that localized features captured exclusively by the private head cannot be reverse-engineered from the shared global updates, adding a critical layer of defense-in-depth to the dynamic spectrum sharing paradigm. IV. P ROPOSED S OLUTION : FL & PERFECT Our overall proposed solution is illustrated in Fig. 2, which follows a series of communication rounds between the ESC clients and the central server during model training. A. Federated Learning Framework Our framework is built upon the FedAvg algorithm [15], a standard for collaborative model training. The process is iterative. In each communication round t, the central server first broadcasts the current global model weights, WBt , to a subset of participating ESC sensors (clients). Each client k then initializes its local model with WBt and performs several epochs of training using its private local dataset Dk . This local
training step computes a new set of weights, wkt+1 , that are optimized for the client’s specific data. These updated weights are then transmitted back to the central server. Finally, the server aggregates the weights from all participating clients, typically via a weighted average based on dataset size, to produce the improved global model WBt+1 for the next round. This cycle repeats, allowing the global model to learn from the collective knowledge of all clients without any raw data ever leaving the local sensors. B. PERFECT Framework Standard FedAvg struggles in non-IID settings as it forces a single global model on all clients. We introduce personalization using the FedPer algorithm [18]. At communication round t, each client’s model is split into a shared base (with weights wkt ) and a private head (with weights vk ), making the full local model (wkt , vkt ). In PERFECT, each ESC trains its full model (wkt , vkt ) on its local data, but only the base model weights wkt are returned to the server for aggregation; the personal head weights vkt remain private. This separation allows the system to learn collaboratively while preserving environment-specific adaptations: 1) Global Knowledge: The shared base layers learn general visual patterns from spectrograms common to all ESCs, such as frequency contours, temporal textures, and repetitive structures of radar and commercial signals. 2) Local Knowledge: The personal head layers capture high-level, environment-specific nuances present in the local data, allowing each ESC to specialize based on its unique operating conditions (e.g., localized clutter, specific radar signatures). The local optimization objective at ESC k is to find (wkt+1 , vkt+1 ) = arg min(wt+1 ,vt+1 ) Lk (wkt+1 , vkt+1 ) starting k k from (wkt , vkt ) = (WBt , vkt ), where Lk (.) represents loss function for ESC k. After local updates, the server aggregates the base weights WBt+1 from all ESCs: K X |Dk | (5) WBt+1 = γk wkt+1 , where γk = |D| k=1 This PERFECT procedure is formally described in Algorithms 1 and 2. Algorithm 1 FedPer ESC Local Update Require: Dk , learning rate ηkt , epochs E, batch size B t 1: Input: Global base weights WB from server 2: Initialize private head weights vk0 at random 3: for each round t = 1, 2, . . . do 4: Receive WBt from server 5: (wkt+1 , vkt+1 ) ← Train(model = (WBt , vkt ), data = Dk ) 6: Send wkt+1 to server (Call Algorithm 2) 7: end for V. DATASET To evaluate our PERFECT framework, we generate a synthetic dataset using MATLAB R2025a , considering a network of 5 ESC deployed at different locations along the sea coast, one LTE, and one 5G base station placed within fixed distances from these ESC sensors, as shown in Fig. 3.
Algorithm 2 FedPer Central Server Update Require: Total ESCs K, target recall R∗ 0 1: Initialize WB at random 2: for each ESC k = 1, . . . , K do 3: Receive sample count nk and compute γk = P nk / j nj 4: end for 0 5: Send WB to all ESCs 6: for each round t = 1, 2, . . . until recall > R∗ do 7: Receive wkt+1 from each ESC k (Call Algorithm 1) PK 8: Aggregate: WBt+1 ← k=1 γk wkt+1 9: Broadcast WBt+1 to all ESCs 10: end for
Fig. 3: A synthetically generated CBRS spectrum sharing scenario showing 5 geographically distributed ESC sensors sensing signals transmitted by different users in the shared band (radars, LTE PAL, 5G GAA). Spectrograms across each ESC show different strengths of 5G and radar Type 1 signals received at each ESC (as an example).
We consider two types of incumbent shipborne radar signal emitters, P0N#1 (denoted as Type 1) and P0N#2 (denoted as Type 2), generated using the NIST radar waveform generator [3]. Without loss of generality, we consider LTE base station to be a PAL user and 5G to be a GAA user. Each ESC captures signal snapshots from the radar, 5G, and LTE sources, with the observed signal strength and phase influenced by its distance from each transmitter. As a result, the SAS receives weight parameters updates from 5 ESC models locally trained on five distinct signal snapshots in each communication round. We assume that the ESCs in our PERFECT system monitor a 10 MHz sub-channel in the CBRS band and sample the RF environment at 10 MHz sampling rate every 20 ms. We synthetically generate type P0N#1 (Type 1) and P0N#2 (Type 2) radar waveforms using National Institute of Standards and Technology’s simulated radar waveform generator [19]. We consider the following parameters for Type 1 (Type 2) radar- pulse width 0.5 − 2.5 (13 − 52) µs, pulse repetition rate 1000 − 1100 (1000 − 2000) pulses per second, and pulses per burst 15−20 (10−20) pulses. These radar types have significant sidelobes, which makes interference scenarios with telecom signals extremely difficult to detect [20], samples are shown in Fig. 4. We generate 3GPP-compliant time division duplex downlink LTE and 5G signals of bandwidths 5 MHz each using MATLAB LTE and 5G toolboxes, respectively.
Fig. 4: Sample spectrograms featuring Type 1 and Type 2 radars with 5G and LTE signals. It is evident that both radar types exhibit significant sidelobes despite the Type 2 radar having longer pulse width (or lower bandwidth) than Type 1.
Fig. 5: Sample spectrograms of different types of signals corresponding to different subcategories in hypotheses classes H0 and H1 .
We construct a variety of signal scenarios corresponding to the hypotheses H1 (overlapping signals) and H0 (nonoverlapping signals). Under hypothesis H1 , we define three subcategories: (i) radar and 5G signals overlapping, (ii) radar and LTE signals overlapping, and (iii) radar overlapping with LTE or 5G signals when both telecom signals are present. For hypothesis H0 , we consider six subcategories: (i) radar only, (ii) 5G only, (iii) LTE only, (iv) simultaneous LTE and 5G without radar, (v) radar and 5G without LTE, and (vi) radar and LTE without 5G. Sample spectrograms corresponding to these subcategories are shown in Fig. 5. Our dataset contains 500 spectrogram frames per subcategory per ESC per radar type, thus generating a total of 22, 500 frames. For CBRS spectrum sharing, the FCC requires 99% radar detection accuracy for overlapping scenarios with radar peak power above −89 dBm/MHz and the aggregate telecom and noise interference (LTE/5G+noise) power below −109 dBm/MHz, that is, radar signal-to-(telecom)-interference and noise ratio (SINR) ≥ 20 dB [21]. To ensure this, we consider radar peak power in the range [−89, −85] dBm/MHz and the aggregate interference power in [−111, −109]) dBm/MHz, such that the SINR ≥ 20 dB is observed across all the ESCs after accounting for signal power variations that incur due to their different locations. To simulate a non-IID environment, we vary the power of different signals and the presence of specific radar types across the 5 ESC sensors 1 . Remark. This is the first-of-its-kind publicly available dataset that comprehensively captures all possible shared spectrum configurations within the CBRS band. It enables the development and benchmarking of ML models for accurate signal detection across distributed shared spectrum sensing. VI. P ERFORMANCE E VALUATION OF PERFECT In this section, we evaluate the performance of PERFECT using the dataset described in Sec. V. We first outline the experimental settings, followed by the performance metrics and baseline models used for comparison. A. Experimental Settings Our detection model is a Convolutional Neural Network (CNN) with residual connections implemented in PyTorch. The training is conducted for T = 150 communication rounds, with each client or ESC performing E = 1 local epoch of training in each round. We use the Adam optimizer with a
learning rate of η = 0.001. To evaluate our models, we use the following standard performance metrics for classification: • Recall (Probability of Detection): This is our primary metric, as the FCC mandates a 99% detection rate for the P harmful overlap class (H1 ) [1]. It is defined as T PT+F N. • Accuracy: The overall fraction of correct classifications. Here, T P , F P , and F N represent True Positives, False Positives, and False Negatives, respectively. Baseline Learning Paradigms. We compare our proposed framework with the following three baseline learning paradigms: 1) Local-only Learning: Each ESC trains a model only on its own local data, with no collaboration. This represents a privacy-preserving lower bound. 2) Centralized Learning: A single model trained on the entire dataset, aggregated from all ESCs. This represents the performance upper bound but violates data privacy. 3) Standard FedAvg: The federated averaging algorithm [15] without any personalization. Dataset. To evaluate the proposed PERFECT framework, we use the dataset with spectrograms collected from 5 ESC sensors, as discussed in Sec. V. Each spectrogram consists of either one or multiple signals of LTE, 5G, and one of two radar types. This results in 5 distinct local datasets, each containing samples from both the harmful overlap (H1 ) and no-overlap (H0 ) classes. Crucially, the distribution of the two radar types is non-identical across the 5 ESCs to simulate a realistic nonIID environment. To replicate real-world situations, we create local training and validation sets for each ESC sensor by splitting its local data (80%/10%). To test for generalization against unseen environments, we create a global test set by combining the leftover 10% of data from all ESC sensors. The overall dataset contains 22, 500 and 20, 250 local training and validation and 2, 250 global test samples, respectively. B. Architecture of the Proposed CNN Model Our proposed CNN architecture is a lightweight, 616, 002parameter model inspired by ResNet [22], as shown in Fig. 6. It begins with a convolutional layer to extract initial features, followed by four residual blocks that progressively increase feature channels while downsampling. The network concludes with a global average pooling layer and a fully connected classifier. This residual design allows for effective training by mitigating the vanishing gradient problem.
We present a detailed comparison of model performance under three baseline learning paradigms: Local, Centralized, and FL, as shown in Fig. 7, Fig. 8a, and Fig. 8b, respectively. Fig. 7 illustrates the performance of local learning at individual ESCs. It is evident that the recall and accuracy vary significantly across ESCs. For example, ESC4 achieves only around 73% recall for radar signals with no overlap, which is well below the FCC-mandated threshold of 99%. Similarly, the overall accuracy for ESC4 is just 83%, and even ESC2 records suboptimal performance with radar recall of only 88%. This inconsistency highlights the inability of local models to generalize due to data heterogeneity and limited training samples at each ESC.
$ F F X U D F \ 5 H F D O O
$ F F X U D F \ 5 H F D O O 1 R 2 Y H U O D S & O D V V 5 H F D O O 2 Y H U O D S & O D V V
( 6 &
( 6 &
( 6 &
( 6 & & D W H J R U \
( 6 &
( 6 &
Fig. 7: Performance of local learning: inference accuracy, recall for H0 (No overlap) and H1 (Harmful overlap) at each ESC sensor.
9 D O L G D W L R Q $ F F X U D F \ 5 H F D O O 1 R 2 Y H U O D S 5 H F D O O 2 Y H U O D S
( S R F K V
(a)
$ F F X U D F L H V 5 H F D O O V
$ F F X U D F L H V 5 H F D O O V
* O R E D O $ F F X U D F \ 5 H F D O O 1 R Q R Y H U O D S 5 H F D O O 2 Y H U O D S
& R P P X Q L F D W L R Q 5 R X Q G V
(b)
Fig. 8: (a) Centralized learning performance: achieves 99% validation accuracy across ESCs but requires data sharing, (b) FL (using standard FedAvg) validation performance: achieves 99% accuracy across ESCs using only the shared global model, preserving data privacy.
Fig. 8a illustrates the accuracy of the validation and the trends in class recall of the centralized model at all training epochs. The model achieves consistently high performance, with both the accuracy and recall values stabilizing above 98% in later epochs. This highlights the strength of centralized training in learning from the complete data distribution. Nonetheless, this approach assumes unrestricted access to all
5 D G D U 7 \ S H 5 D G D U 7 \ S H
( 6 &
( 6 &
( 6 &
( 6 & , '
( 6 &
( 6 &
6 D P S O H & R X Q W
C. Comparative Analysis of Learning Paradigms
6 D P S O H & R X Q W
Fig. 6: Deep residual CNN model architecture with an initial convolutional block, followed by four residual blocks.
local data, which may not be feasible in privacy-sensitive environments and imposes considerable communication overhead. Fig. 8b presents the results of our proposed FL framework using FedAvg [15] as a global aggregator. The accuracy and radar recall curves converge rapidly within the first 50 communication rounds and stabilize around 98-99%, closely matching the centralized performance. Notably, FL achieves this without direct data sharing, validating its effectiveness while preserving privacy. The FL-based model also significantly outperforms the local models shown in Fig. 7, particularly in recall, thereby satisfying regulatory constraints without centralized training. During this set of experiments, all the 616,002 parameters of our proposed CNN are being shared from each ESC with the SAS (central server) for each communication round.
5 D G D U 7 \ S H 5 D G D U 7 \ S H
( 6 &
( 6 &
( 6 &
( 6 & , '
( 6 &
( 6 &
(a) (b) Fig. 9: Radar sample distribution per ESCs under (a) non-IID and (b) IID setups. Each stacked bar shows counts for both radar types.
Observation 1. The proposed federated learning framework is able to detect harmful interference with radar signals (hypothesis H1 ) at FCC mandated regulation of 99% recall while preserving the data privacy (see Fig. 7, 8a, and 8b, validates Contribution 1). D. Robustness to Non-IID Environment via PERFECT Non-IID data distributions pose a fundamental challenge to FL systems, often leading to poor generalization and model bias when ESC data varies significantly across the network. This experiment highlights the necessity of personalization in FL when dealing with non-IID data using FedPer as global aggregator [18]. In our setup, each ESC observes radar signals with different distributions- ESC1 primarily sees radar Type 1, while ESC5 mostly encounters radar Type 2, as illustrated in Fig. 9. To understand the impact of this non-IID setup, we trained the network on data from one ESC and evaluated it across all others. The results, summarized in Table I, show that models perform best on the test data from the same ESC on which they were trained. For example, training and testing on ESC3 yields 98.98% accuracy and 98.00% recall, while similar consistency is observed for ESC2 and ESC5. However, performance drops significantly in cross-domain scenarios. Notably, when trained on ESC1 and tested on ESC5, the model achieves only 84.94% accuracy and 87.54% recall. Conversely, when trained on ESC5 and tested on ESC1, the performance further degrades to 75.47% accuracy and 74.19% recall. This performance gap highlights how global models struggle under domain shift, often biasing toward the dominant radar type in the training data. In contrast, our PERFECT framework adapts each ESC model to its local distribution. As shown in Fig. 10, all ESCs converge to an accuracy of over 99% and maintain high recall
( 6 & ( 6 & ( 6 & ( 6 & ( 6 &
& R P P X Q L F D W L R Q 5 R X Q G
( 6 & ( 6 & ( 6 & ( 6 & ( 6 &
(a) Per ESC (client) accuracy over rounds
& R P P X Q L F D W L R Q 5 R X Q G
5 H F D O O
5 H F D O O
$ F F X U D F \
( 6 & ( 6 & ( 6 & ( 6 & ( 6 &
(b) Recall for no-overlap Class (H0 )
& R P P X Q L F D W L R Q 5 R X Q G
(c) Recall for overlap class (H1 )
Fig. 10: PERFECT validation accuracy across each ESC and per-class recall, over communication rounds.
in both the overlapped and non-overlapped classes, even in the presence of significant heterogeneity. TABLE I: Accuracy and recall (%) for each test dataset across different training sets. The average accuracy is 86.57%, and the average recall is 92.49% across all ESCs. Test Data ESC1 ESC2 ESC3 ESC4 ESC5
ESC1
ESC2
Trained Data ESC3
ESC4
ESC5
99.14/98.02 98.95/97.58 89.26/89.09 85.12/90.42 84.94/87.54
90.49/98.47 92.46/98.70 86.46/99.25 87.37/95.95 94.20/88.79
75.95/65.20 84.97/86.80 98.98/98.00 78.14/76.06 71.75/59.03
79.21/94.46 79.78/95.91 83.53/98.66 87.22/98.71 84.48/98.75
75.47/74.19 75.84/78.07 83.71/91.98 98.45/97.24 98.40/98.35
Observation 2. PERFECT is able to provide a stable 99% recall for overlap class (H1 ) for cross-domain testing scenarios, and 15-26% improvement than standard FedAvg of Sec. VI-C (see Fig. 10, validates Contribution 2). E. Comparison Across ESC Clients Fig. 11 compares a PERFECT configuration, where each client keeps a small head local and shares only the backbone, with a non-personalized configuration, where all weights are synchronized. Across ESC-1 to ESC-5 the personalized setup sustains high radar recall in both conditions: the overlap class is typically 0.99 to 1.00 and the no-overlap class is consistently 1.00 (with the occasional 0.98 on a few ESCs). In terms of communication, sharing only the backbone reduces the perround payload from about 11.75 MB to about 11.43 MB per client. This is a reduction of roughly 0.32 MB per round, which accrues over many rounds and clients. The reduction comes from not transmitting the client-specific head, which also keeps client-unique features on device. Taken together, the results show that a PERFECT preserves the required ≥ 99% detection performance while trimming uplink traffic and adding a practical layer of privacy for each ESC. F. Comparison with State-of-the-Art (SOTA) Table II presents a comparative analysis of the parameter counts in our proposed architecture versus state-of-the-art radar signal detection models: Waldo [9], DeepRadar [23], and FaIR [14]. Waldo and DeepRadar consist of over 61M and 17M parameters, respectively. FaIR, a centralized model, includes 138M parameters, all of which are shared. In contrast, our model comprises only 616, 002 parameters—offering 100×, 27×, and 223× reduction compared to Waldo, DeepRadar, and FaIR, respectively. More importantly, the PERFECT, only 599, 232 base parameters are globally shared, while the 16, 770 local head parameters remain private. Unlike FaIR, which transmits its entire model in each
round, PERFECT reduces communication to less than 0.5% of FaIR’s size. This enables an efficient and privacy-preserving solution for non-IID environments and edge deployments. TABLE II: Parameter and privacy comparison of proposed work and SOTA. Our proposed work offers 100×, 27×, and 223× reduction in model parameters compared to Waldo [9], DeepRadar [23], and FaIR [14], respectively. Model Total Waldo [9] DeepRadar [23] FaIR [14] Proposed (FedAvg, Sec. VI-C) Proposed (PERFECT, Sec. VI-D)
61, 692, 331 16, 934, 018 138, 357, 544 616, 002 616, 002
Params Shared
Local
– – 138, 357, 544 – 599, 232
– – – – 16, 770
Privacy
Acc.
× × ✓ ✓ ✓
99% 99% 79% 99% 99%
G. Constraints on Real-world Experiments While validating on real-world RF data is ideal, conducting large-scale, distributed measurements of incumbent naval radar operations in the CBRS band is severely constrained by operational security and regulatory restrictions. Although there exist a few real-world datasets capturing radar and interference signals for the CBRS band, none of them consider a geographically distributed ESC sensor network where data is simultaneously collected from multiple sensors. Raw RF and spectral measurements inherently reveal highly sensitive operational details, including naval vessel trajectories and patterns of incumbent radar activity. Consequently, capturing and coordinating real-time, non-IID distributions of these rare events across a physical, multi-node ESC network is not only logistically complex but also poses severe privacy risks. To overcome these practical barriers while ensuring rigorous evaluation, our emulation setup utilizes the NIST simulated radar waveform generator [3] within MATLAB. This approach allows us to accurately model the complex side-lobe characteristics, pulse widths, and repetition rates of different naval radars, multiplexed with 3GPP-compliant LTE and 5G signals. By precisely controlling the SINR to align with FCC compliance thresholds (SIN R ≥ 20dB), our synthetic testbed provides a reproducible, standardized environment. This tightly controlled setup is essential for isolating and evaluating the specific non-IID label and feature skews that the PERFECT framework is designed to mitigate, conditions that are virtually impossible to systematically guarantee or groundtruth in an existing, unclassified measurement campaign. VII. C ONCLUSIONS In this paper, we address the critical challenge of incumbent radar interference detection in the CBRS band under strict privacy and non-IID data constraints. We propose PERFECT,
16 12 8 4
A
B
ESC-1
A
B
ESC-2
A
B
ESC-3
A
B
ESC-4
A
B
0
Parameters Size (MB)
11.75 1.00 0.95
11.43 0.99 1.00
Parameters size (MB)
11.43 1.00 1.00
11.75 0.99 1.00
11.43 1.00 1.00
11.75 0.98 1.00
11.43 1.00 1.00
11.75 0.70 1.00
11.43 0.99 1.00
Radar Recall
1.2 1.0 0.8 0.6 0.4 0.2 0.0
B: No Personalization
Radar No-overlap Recall
11.75 0.98 1.00
A: PERFECT Radar Overlap Recall
ESC-5
Fig. 11: Comparison across ESC-1–ESC-5 for two FL setups: A, PERFECT, uses a personalized head kept local while sharing only the backbone; B shares the full model without personalization. Personalized FL (A) is consistent on non-IID data, maintaining ≥ 99% overlap recall across ESCs; FedAvg (B) is inconsistent across ESCs (see ESC-1, B).
a FL framework that successfully detects faint radar pulses and leverages client level personalization to improve generalization. Our extensive simulations demonstrate that our approach achieves the FCC-mandated 99% recall, performing on par with a non-private centralized model while significantly outperforming standard FL and local-only learning approaches. We show that our PERFECT framework effectively mitigates the negative impact of non-IID data. Future work will focus on deploying this framework on a hardware testbed with over-theair transmissions to validate its performance in real-world conditions, exploring online and continual learning mechanisms to adapt to evolving interference patterns, and investigating more advanced model compression and communication-efficient FL techniques to further enhance its efficiency and robustness under high-interference scenarios. ACKNOWLEDGMENT The authors gracefully acknowledge the funding from the US National Science Foundation (CNS 2526490). R EFERENCES [1] Federal Communications Commission, “Amendment of the commission’s rules with regard to commercial operations in the 3550-3650 mhz band,” Federal Communications Commission, Report and Order GN Docket No. 12-354, 2015. [2] Wireless Innovation Forum, “Requirements for commercial operations in the u.s. 3.5 ghz citizens broadband radio service band,” The Wireless Innovation Forum, Technical Specification WINNF-TS-0112, 2018. [3] R. Caromi, M. Souryal, and T. A. Hall, “Rf dataset of incumbent radar signals in the 3.5 ghz cbrs band,” Journal of Research of the National Institute of Standards and Technology, vol. 124, p. 1, 2019. [4] MathWorks, “Dynamic spectrum sharing for 5G NR and LTE coexistence,” https://www.mathworks.com/help/5g/ug/ dss-for-5g-nr-and-lte-coexistence.html, 2024. [5] R. Caromi, M. Souryal, and W.-B. Yang, “Detection of incumbent radar in the 3.5 ghz cbrs band,” in 2018 IEEE Global Conference on Signal and Information Processing (GlobalSIP). IEEE, 2018, pp. 241–245. [6] R. Caromi, A. Lackpour, K. Kallas, T. Nguyen, and M. Souryal, “Deep learning for radar signal detection in the 3.5 ghz cbrs band,” in 2021 IEEE International Symposium on Dynamic Spectrum Access Networks (DySPAN), 2021, pp. 1–8. [7] R. Vanukuri, S. U. Khan, T. T. Sarı, G. Secinti, D. Patiño, and D. Roy, “Waves of imagination: Unconditional spectrogram generation using diffusion architectures,” in MILCOM 2025-2025 IEEE Military Communications Conference (MILCOM). IEEE, 2025, pp. 1468–1473. [8] S. U. Khan, M. Kulhandjian, and D. Roy, “Pushing the boundaries in cbrs band: Robust radar detection within high 5g interference,” in MILCOM 2025-2025 IEEE Military Communications Conference (MILCOM), 2025, pp. 1278–1283. [9] N. Soltani, V. Chaudhary, D. Roy, and K. Chowdhury, “Finding waldo in the cbrs band: Signal detection and localization in the 3.5 ghz spectrum,” in 2022 IEEE Global Communications Conference (GLOBECOM). IEEE, 2022, pp. 1–6.
[10] C. Tassie, A. Gaber, V. Chaudhary, N. Soltani, M. Belgiovine, M. Loehning, V. Kotzsch, C. Schroeder, and K. R. Chowdhury, “Detection of coexisting rf signals in cbrs using ml: Dataset and api-based collection testbed,” IEEE Communications Magazine, vol. 61, no. 9, pp. 82–88, 2023. [11] R. Hazari, G. Singh, D. Renjith, D. Krishnan, P. B, H. Olufowobi, and D. Roy, “Spec-scan: Spectrum learning in shared channel using neural networks,” in 2025 IEEE 22nd Consumer Communications & Networking Conference (CCNC), 2025, pp. 1–6. [12] G. Reus-Muns, P. S. Upadhyaya, U. Demir, N. Stephenson, N. Soltani, V. K. Shah, and K. R. Chowdhury, “Senseoran: O-ran-based radar detection in the cbrs band,” IEEE Journal on Selected Areas in Communications, vol. 42, no. 2, pp. 326–338, 2024. [13] S. U. Khan, M. Kulhandjian, and D. Roy, “In-Network Fusion for High Interference Signal Detection within CBRS Band,” in IEEE International Conference on Computer Communications (INFOCOM), May 2025. [14] M. Troglia, J. Melcher, Y. Zheng, D. Anthony, A. Yang, and T. Yang, “Fair: Federated incumbent detection in cbrs band,” in 2019 IEEE International Symposium on Dynamic Spectrum Access Networks (DySPAN), 2019, pp. 1–6. [15] B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” in Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS). PMLR, 2017, pp. 1273–1282. [16] M. Baduwal, P. Paudel, and V. Chaudhary, “Federated learning: A survey of core challenges, current methods, and opportunities,” Computers, vol. 15, no. 3, 2026. [Online]. Available: https://www.mdpi. com/2073-431X/15/3/155 [17] P. Kairouz, H. B. McMahan, B. Avent, A. Bellet, M. Bennis, A. N. Bhagoji, K. Bonawitz, Z. Charles, G. Cormode, H. Eichner et al., “Advances and open problems in federated learning,” Foundations and Trends® in Machine Learning, vol. 14, no. 1–2, pp. 1–210, 2021. [18] M. G. Arivazhagan, V. Aggarwal, A. K. Singh, and S. Choudhary, “Federated learning with personalization layers,” CoRR, vol. abs/1912.00818, 2019. [Online]. Available: http://arxiv.org/abs/1912.00818 [19] R. Caromi, A. Lackpour, K. Kallas, T. Nguyen, and M. Souryal, “Deep Learning for Radar Signal Detection in the 3.5 GHz CBRS Band,” in 2021 IEEE International Symposium on Dynamic Spectrum Access Networks (DySPAN). IEEE, 2021, pp. 1–8. [20] F. H. Sanders, “Effects of radar interference on lte base station receiver performance,” Institute for Telecommunication Sciences, Tech. Rep., 2013. [21] F. H. Sanders, J. E. Carroll, G. A. Sanders, R. L. Sole, J. S. Devereux, and E. F. Drocella, “Procedures for laboratory testing of environmental sensing capability sensor devices,” National Telecommunications and Information Administration, Technical Memorandum TM, pp. 18–527, 2017. [22] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, June 2016, pp. 770–778. [23] S. Sarkar, M. Buddhikot, A. Baset, and S. K. Kasera, “Deepradar: A deep-learning-based environmental sensing capability sensor design for cbrs,” in Proceedings of the 27th annual international conference on mobile computing and networking, 2021, pp. 56–68.