Spectrum Sensing Meets ISAC: An Spectrum Detection Scheme for ISAC Services Based on Improved Denoising Auto-Encoder and CNN

Abstract

Integrated Sensing and Communications (ISAC) has attracted increasing attention due to more efficient utilization of both radio spectrum and hardwares. However, ISAC can only relieve the shortage of the spectrum, especially in the situation of exponential growth of wireless terminals. Efficient spectrum utilization strategy is still an important direction for the continuous evolution of wireless communication technology. As such, spectrum sensing (SS) is discussed in ISAC scenarios, and a novel cooperative SS scheme is proposed by as an improved auto-encoder for more efficient spectrum utilization. More specifically, the parameters of each local spectrum spectrum-sensing network are encoded and sent to the central server, and the local network parameters are decoded, fused, and returned to each local node at the central server. The simulations are designed, and the experiment results demonstrate the effectiveness of the proposed scheme.

1. Introduction

In recent years, with the rapid development of wireless communication technology, the types and quantities of wireless devices are have growing grown exponentially, leading to the increasing shortage of spectrum resources. Cognitive Radio (CR) is one of the potential technologies to improve spectrum efficiency by opportunistic spectrum access [1,2]. Additionally, the spectrum of wireless sensing and wireless communication gradually coincides. Integrated Sensing and Communications (ISAC) works as a new paradigm for improving spectrum efficiency by unified sensing and communication waveform design [3,4]. When CR meets ISAC, two gains will be gained. Firstly, the spectrum efficiency will be further improved by the dual function of idle spectrum reuse and integrated enabling of the occupying spectrum. Moreover, the hardware cost of wireless service will be greatly reduced by spectrum sharing of wireless sensing and wireless communication services.

As a key technology of CR, Spectrum Sensing (SS) is a well-known idle spectrum monitoring technology, including traditional SS and intelligent SS. Traditional SS is the mathematical model-driven sensing strategy, such as energy detection (ED) [5,6], matched filtering (MF) [7], cyclic spectrum feature detection (CSFD) [8], covariance matrix detection (CMD) [9], etc. Intelligent SS relies on artificial intelligence technology to extract high-order features of the observed spectrum. As a result, intelligent SS is data-driven and better than traditional SS in performance. False alarm probability and detection probability are two basic indicators to evaluate the performance of SS. Generally, detection probability denotes the ability of protectingto protect authorized users, while false alarm probability is positively related to the system throughput of cognitive users.

On the other hand, with the development of 5G and Beyond, the radio spectrum continues to evolve to high frequency, resulting in the increasing spectrum overlap between wireless sensing and wireless communication. This provides the possibility of spectrum sharing for wireless sensing and wireless communication, namely ISAC. Currently, the focus of ISAC consists inconsists of ISAC waveform design and signal processing, communication-assisted sensing, and sensing-empowered communication [10,11,12], where ISAC waveform design mainly focuses on the waveform design of sensing and communication information in the same frequency band. Distributed sensing is mainly considered for communication-assisted sensing, where the sensing performance of each node is enhanced by mutual information sharing. In addition, sensing-empowered communication is devoted to the sensing of wWireless channels by the echo of the wireless signal, thereby providing a more reasonable configuration of RF parameters and improving the communication performance as a consequence.

Although SS and ISAC ISAC, respectivelyrespectively, show the ability to improve spectrum efficiency, the contradiction between the growing frequency demand and the limited spectrum resources is still the bottleneck of wireless services [13]. Motivated by this, this paper considers the fusion of ISAC and SS, and provides an opportunistic available spectrum for ISAC services. More specifically, an improved denoising auto-encoder is considered for cooperative spectrum sensing to discover an available spectrum of ISAC service, where the encoder is deployed on the local nodes, and the decoder is deployed on the central server. At the local node, the observed signal is input to the designed convolutional neural network (CNN) [14] to decide the status of observed spectrum bands, and then the network parameters are encoded and sent to the central server. At the central server, the encoded network parameters of local nodes are decoded and fused to the weighted network parameters, and then the weighted network parameters are distributed to local nodes. Note that the federated learning (FL) framework [15,16,17] is considered for this cooperative SS scenario, and the denoising auto-encoder is trained locally in advance. The main contributions of this paper are concluded as follows.

(1)

A more efficient spectrum utilization scheme is discussed by the fusion of SS and ISAC, and this scheme may provide more available spectrum for ISAC services, especially for 5G and beyond [18].

(2)

An improved denoising auto-encoder is proposed for joint spectrum detection based on FL, where the local network parameters of this improved scheme are encoded locally. As such, the communication overhead of cooperative sensing is greatly reduced [19,20].

2. Open Challenges When SS Meets ISAC

ISAC is a cutting-edge research area that combines wireless communication and radar-like sensing into a unified framework. SS plays a critical role in ISAC, particularly for dynamic spectrum access, interference mitigation, and environment-aware communication. However, there are several open challenges in integrating SS with ISAC.

2.1. Hardware and RF Front-End Limitations

ISAC requires a shared RF front-end for both sensing and communication, which can lead to hardware impairments like phase noise, nonlinearity, and limited dynamic range. Balancing the design between high-fidelity sensing and high-speed communication remains an unresolved issue.

2.2. Spectrum Sharing and Dynamic Access

ISAC must intelligently allocate spectrum resources for both functions, which requires advanced spectrum- sensing techniques to detect available frequencies in real time. Coexistence with legacy systems such as Wi-Fi and 5G networks poses spectrum management challenges.

2.3. Interference Management

ISAC systems suffer from self-interference and cross-interference with other users, making real-time mitigation techniques essential. Cognitive radio techniques can help adapt transmission parameters dynamically, but robust algorithms are still needed.

2.4. Machine Learning for Spectrum Sensing in ISAC

Deep learning and reinforcement learning offer promising solutions for dynamic SS, but require extensive training data and real-time adaptability. Developing lightweight AI models for low-latency ISAC applications remains an open problem.

2.5. Security and Privacy Concerns

Spectrum sensing in ISAC raises potential eavesdropping and spoofing threats, where attackers can exploit sensing data to manipulate the system. Secure spectrum- sensing techniques, such as privacy-preserving FL, need further exploration.

3. Proposed Scheme

To protect the SS data privacy, an FL-enabled SS framework is applied, where the spectrum spectrum-sensing data isdata are trained locally, and the updated model parameters are transmitted to the server for global model aggregation. Traditionally, to achieve better training performance, the complexity of the model is getting higher and higher, leading to a huge communication overhead in FL. To reduce the communication overhead during FL, model compression methods are widely applied where the local model is compressed before updating to the server. The traditional model compression method, such as principal component analysis (PCA), directly reduces the feature space and abandons the local feature, resulting in the loss of model information. Compared to the traditional model compression methods, auto-encoder achieves efficient reduction and reconstruction of high-dimensional data through the structure of encoder and decoder, can automatically learn effective representations based on data characteristics and adapt to different types of data compression requirements, and is applied in this paper to alleviate the communication overhead. In the proposed framework, the SS nodes firstlyfirst train the local CNN model, the updated model parameters are then encoded with an local auto-encoder, the encoded local model parameters are uploaded into the central server, the server decodes the received local model parameters, and a global model is aggregated finally. The overall workflow is shown in Figure 1.

3.1. System Model

The spectrum detection system for ISAC services in this paper consists of n nodes, each of which holds a local dataset of observed signals. The local node is represented as $C=\{c_1,c_2,\dots ,c_n\}$. Initially, an auto-encoder within each node generates model parameters, which are then uploaded to a central server. This server oversees the FL process, decoding and aggregating these parameters. Let’s Let us denote w as the global model parameter and $w_n$ as the local model weight of client n. The global model of the CNN network is initialized with random weights and distributed to all nodes.

Each node updates its local model parameters using gradient descent. Specifically, at time step t, the local model weight $w_n^t$ of client n is updated according to:

$$w_n^t=w^{t-1}-\eta \nabla F_n\left(w^{t-1}\right),$$

(1)

where $\eta$ represents the learning rate of the client-side model. $F_n\left(w\right)$ stands for the local loss function at client n, and $\nabla F_n\left(w\right)$ denotes the local gradient at client n. Essentially, each node optimizes its local model parameters based on its own dataset to drive convergence.

After each round of training, the server receives updated model parameters from each node and computes the difference between the local and global models. Parameter aggregation occurs next, with the server aggregating local updates from nodes using a weighted average. Suppose $C_t$ represents the collection of clients selected at round t, and $\left|C_t\right|$ is the number of clients selected. The global model $w^t$ at time step t is updated as follows:

$$w^t=w^{t-1}+{\displaystyle \frac{1}{\left|C_t\right|}}\sum _{n\in C_t}{\alpha }_n{w}_n^t-w^{t-1},$$

(2)

where ${\alpha }_n$ signifies the weight of client n. Subsequently, the latest global model parameters are disseminated to each node to initiate the next round of model training. This iterative process continues until the spectrum-aware model converges or achieves a predefined accuracy threshold.

3.2. Communication Model

In cooperative training, the data transmission of the uplink channel is easily limited. In addition, the model parameters are subject to noise interference during channel transmission. To reduce the communication overhead between the node and the central server, an improved auto-encoder model is considered in the communication process of the FL.

In the communication process of global aggregation, it is assumed that n nodes participate in FL and send model parameters through a shared wireless channel. To ensure a fast, reliable, and privacy-preserving transmission process, the node first compresses and denoises the sent data through an auto-encoder. Upon receiving the data, the server decodes it to retrieve the actual parameters, facilitating model fusion. Specifically, node-to-server communication learns the mapping from input to itself through the encoding and decoding phases. Let $E(.)$ represent the encoding function and $w_n$ be the model weight of client n. The channel input for client n is represented as $H_n$:

$$S_n=H_{n}E\left(w_n\right)+N_n,$$

(3)

where $S_n$ is the channel output received by the server, and $N_n$ is the noise term. $D(.)$ is the decoding function. According to the received channel output, the decoding process of the server can be represented as:

$${\widehat{w}}_n=D\left(S_n\right).$$

(4)

After decoding, $\widehat{w}$ is the model update information for client n. The server obtains the model update information of each client to complete the model fusion. The n nodes will obtain the updated spectrum-aware model and repeat the process until the model accuracy is as required.

3.3. Network Framework

In this paper, CNN network models are deployed at each node to perform the spectral image classification task. Note that the covariance matrix of the observed signal is firstly first obtained, and then it is converted to the spectral image. The CNN network architecture used in this paper is shown in

Table 1. CNN network architecture at each node.

Layer Name	Parameters
Convolutional Layer 1	Kernel Size: $(3,10,5,5)$
Max- Pooling Layer 1	Pool Size: $(2,2)$
ReLU Activation	-
Convolutional Layer 2	Kernel Size: $(10,20,5,5)$
Max- Pooling Layer 2	Pool Size: $(2,2)$
Dropout Layer	Default Dropout Probability
Flatten Layer	-
Fully Connected Layer 1	in features: $\left(320\right)$, out features: $\left(40\right)$
ReLU Activation	-
Dropout Layer	Default Dropout Probability
Output Layer	in features: $\left(40\right)$, out features: $\left(2\right)$

and Figure 2.

This study focuses on signal covariance images with relatively few pixels and simple features. For low-pixel images, local patterns are crucial for recognition. CNN can effectively capture these local details with fewer parameters. Compared to more complex models, CNN achieves high performance within a shorter training time. In federated learning, local nodes transmit trained model parameters to a central server. Using a CNN with fewer parameters helps reduce communication overhead.

The basic structure of auto-encoders can be divided into two categories: fully connected auto-encoder and convolutional auto-encoder. Fully connected auto-encoders consist of multiple fully connected layers and are suitable for processing structured data. Convolutional auto-encoders, on the other hand, use a convolutional neural network structure and are suitable for processing image data.

The aim of this paper is to compress the model parameters, so a fully connected auto-encoder is chosen. The network architecture of the fully connected auto-encoder is shown in

Table 2. The network architecture of the fully connected auto-encoder.

Layer Name	Input/Output Size
Encoder	-
Fully Connected Layer 1	Input: 18,702, Output: 2048
Tanh Activation	Input/Output: 2048
Fully Connected Layer 2	Input: 2048, Output: 512
Tanh Activation	Input/Output: 512
Fully Connected Layer 3	Input: 512, Output: 256
Tanh Activation	Input/Output: 256
Fully Connected Layer 4	Input: 256, Output: X (Defaults to 128)
Decoder	-
Fully Connected Layer 1	Input: X(Defaults to 128), Output: 256
Tanh Activation	Input/Output: 256
Fully Connected Layer 2	Input: 256, Output: 512
Tanh Activation	Input/Output: 512
Fully Connected Layer 3	Input: 512, Output: 2048
Tanh Activation	Input/Output: 2048
Fully Connected Layer 4	Input: 2048, Output: 18,702
Tanh Activation	Input/Output: 18,702

.

Assume that the output size of the encoder is X. The size of the model parameter to be compressed is 18,702, and the compression ratio of the encoder is changed by setting the value of X. With X defaulting to 128, for example, the compression ratio is 146.

4. Simulations and Discussions

In this section, we first describe the details of the radio spectrum image datasets for our experiments and the parameters setting. Then, we evaluate the performance of the proposed scheme through simulations.

4.1. Parameter Setting

Orthogonal Frequency Division Multiplexing (OFDM) is a multi multi-carrier modulation technique. It divides a broadband signal into multiple orthogonal sub-carriers, each of which transmits an independent data stream, and ultimately superimposes the signals of all sub-carriers for transmission. It has strong resistance to multi-path fading, high spectrum utilization, strong resistance to inter inter-symbol interference, and can flexibly allocate resources. As a result, the signal type considered in this paper is the OFDM signal due to its universality and versatility.

To verify the performance of our proposed scheme, an SS dataset is designed first, and the theory of the designed dataset is introduced as follows. An initial training and test dataset is created, where the OFDM signal is treated as the primary user’s signal, and the transmission medium is characterized by a Rayleigh fading channel. Noise is introduced to the OFDM signal after it passes through the Rayleigh fading channel. The signal-to-noise ratio (SNR) varies from a certain value to another, with a specified increment. For the multi-antenna system employed in the experiment, the number of antennas is fixed at 10, and the number of sampling points is determined as 10. The sampling covariance matrix is computed based on the real and imaginary components of the received signals, leading to a covariance matrix of a particular size of $10\times 10\times 2$. For the true color plot, there are three channels, with the first two representing the real and imaginary parts of the sampling covariance matrix, and the third being a matrix of zeros. The dimensions of this matrix are dictated by the number of antennas in the system. Subsequently, the true color plot is resized to a specified dimension.

During the dataset creation, two scenarios are taken into account: one with the signal noise and the other with noise alone. Consequently, each dataset is split into two sections, with 3000 samples for each scenario. The entire dataset comprises 45,000 samples, of which 30,000 are allocated for training, and the remaining 15,000 samples form the test set. The dataset contains a total of 30,000 images, where the ratio of the training set to the test set is 2:1. The training and test sets are evenly distributed across 10 nodes. In addition, we consider different auto-encoder compression ratios and noise scenarios.

The number of training rounds is set to 100 to ensure sufficient learning. There are 10 participating nodes. The batch size is 1, allowing fine-grained parameter updates and reducing memory usage, but it may increase training time. Each node has 2000 training samples and 1000 test samples. The Adam optimizer is used. It adaptively adjusts the learning rate and speeds up convergence. The learning rate is set to 0.001. This value ensures stable updates while enabling efficient optimization. The comprehensive parameters of the proposed algorithm is shown in

Table 3. Parameters of the proposed algorithm.

Parameters	Values
Number of training epochs	100
Number of nodes	10
Batch size	1
Number of samples in the training set per node	2000
Number of samples in the test set per node	1000
Optimizer	Adam
Learning rate	0.001

.

In establishing the CNN framework, this study explores different configurations of convolutional and linear layers. The results are shown in Figure 3. The left figure compares the performance of 2-layer and 3-layer convolutional networks. In the early training stages, the 2-layer model shows a clear advantage. As the models approach convergence, the difference becomes negligible. Additionally, the 2-layer model has fewer parameters, which helps reduce communication overhead in federated learning. The right figure compares 2-layer and 3-layer linear networks, where the 2-layer model demonstrates superior performance.

In determining key parameters for the CNN framework, this study conducts experiments on kernel sizes, pooling layers, and dropout values. The experimental results for different parameter settings are shown in Figure 4. The left figure compares the effects of max- pooling and average pooling. The data clearly show that max max-pooling outperforms average pooling overall. The right figure presents the impact of different convolutional kernel sizes. As the models approach convergence, the performance difference between 3 × 3 and 5 × 5 kernels becomes negligible. However, the 5 × 5 kernel exhibits more stable performance during training, making it the preferred choice. For the dropout layer, each neuron has two states: retained or dropped. Given the shallow network used in this study, both states are typically assigned equal probability, which aligns with the default setting.

4.2. Simulation Results

Firstly, we evaluate the performance of the auto-encoder without considering noise. The comparison of test accuracy at no compression and compression ratio of 143 is shown in Figure 5. The results show that after 30 rounds of training with a compression ratio of 143, the performance of the model compression scheme is similar to the performance of the scheme without compression. This indicates that our scheme reduces the communication overhead while maintaining the learning accuracy.

Further, Furthermore, to evaluate the denoising performance of the proposed scheme, we consider the case where the model parameter transmission is contaminated by noise. Gaussian noise with a mean of 0 and standard deviation of 0.1 is superimposed on the model parameters transmitted by each node. First, we train the denoising auto-encoder using model parameters contaminated with noise and model parameters without noise. The trained denoising aotuauto-encoder is then applied to the FL framework.

In Figure 6 and Figure 7, the accuracy of applying the denoised auto-encoder is significantly higher than the undenoised case. The results show that under a compression ratio of 143, our auto-encoder achieves a terrific denoising effect, although it cannot completely eliminate the noise.

Moreover, we evaluate the performance of the denoising auto-encoder under different compression ratios. The test accuracy and average accuracy are shown in Figure 8 and Figure 9. The results show a decrease in performance as the compression ratio increases. This may be due to the fact that as the compression ratio increases, the more information is lost after encoding and decoding.

In the communication process, Gaussian white noise, channel fading, and various other interference factors coexist. The combined effects of these factors lead to variations in the bit error rate (BER) of data received by the central server, resulting in discrepancies between the data sent by local nodes and the data received in the cloud.

To simplify the network architecture and enhance the feasibility of the study, an equivalent simulation approach is adopted. Specifically, Gaussian white noise of varying intensities is superimposed to simulate cloud-received data under different BER conditions. This equivalent experiment enables a simulation-based analysis of system performance under different levels of communication interference. The results are shown in Figure 10. As the level of communication interference increases, the accuracy of federated learning exhibits a significant downward trend.

To evaluate the denoising performance of the proposed denoising auto-encoderencoder, we use median filtering as a baseline and compare both methods under the same noise conditions. Median filtering is a classical denoising algorithm that operates by sorting all data points within a local neighborhood and replacing the central value with the median of the sorted values. This method effectively removes isolated noise points while preserving edge information, ensuring minimal distortion to the original data structure and features.

The results are shown in Figure 11. The blue curve represents accuracy under noise-free conditions, the green curve corresponds to accuracy with median filtering, the black curve represents accuracy without any noise processing, and the red curve indicates accuracy with the proposed denoising auto-encoderencoder. The results demonstrate that the proposed auto-encoderencoder significantly reduces the impact of noise and outperforms the median filtering approach.

5. Conclusions

This paper presents a novel spectrum detection scheme designed to enable opportunistic spectrum access for ISAC services. To enhance the accuracy and reliability of SS, an improved denoising auto-encoder is employed, leveraging its capability to extract useful features from noisy spectrum data. The proposed scheme incorporates cooperative spectrum sensing, where multiple sensing nodes collaborate to improve detection performance, mitigating the effects of fading and shadowing. Furthermore, FL is integrated into the framework to enable distributed model training while preserving data privacy, reducing the need for centralized data aggregation. The simulation results validate the effectiveness of the proposed approach, demonstrating its superior detection accuracy, robustness to noise, and adaptability to dynamic spectrum environments, making it a promising solution for efficient spectrum utilization in ISAC systems.