Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

Journal of Lightwave Technology

Multi-Task Learning With Physics-Informed Neural Network for Bending Eavesdropping Detection and Localization

Wenshuai Qin, Xiaoxue Gong, Weigang Hou, Lu Gan, Yuxin Xu, Lei Guo

Published Sep 24, 2026• 0 citations
Publisher

Abstract

This paper proposes a multi-task learning based physics-informed neural network (MTL-PINN) scheme for detecting and localizing bending eavesdropping in coherent optical communication systems. The scheme employs MTL to develop a unified neural network that simultaneously addresses both the detection and localization tasks, thereby enhancing the model generalization capabilities. By integrating physical information such as linear dispersion, birefringence, and nonlinear effects into a convolutional neural network (CNN) architecture, the proposed PINN model improves both detection and localization accuracy. To validate the scheme, we construct an experimental platform for bending eavesdropping detection and localization in an 80 km, 42 GBaud quadrature phase shift keying (QPSK) coherent optical communication system. Polarization data are collected for bending radii of 10.8 mm and 15 mm under both bending eavesdropping and normal transmission scenarios. The performance of MTL-CNN, MTL-CNN-BiLSTM, and MTL-PINN models in terms of detection and localization is evaluated for both bending radii. Experimental results show that the MTL-PINN achieves localization and detection accuracies of 100% under bending radii of 10.8 mm and 15 mm. Our work provides an innovative perspective for bending eavesdropping detection and localization in optical communication systems.

DOI
10.1109/jlt.2026.3733656

Related papers

Experimental Demonstration of Eavesdropping Localization Based on Physics-Informed Neural Network

Wenshuai Qin, Xiaoxue Gong, Weigang Hou +3 more

This paper proposes a physics-informed neural network (PINN) scheme for localizing bending eavesdropping in coherent optical communication systems. First, we establish a signal transmission model under bending eavesdropping based on the Manakov equations to analyze its impact on physical information such as linear birefringence and nonlinearity. Subsequently, a PINN is developed by incorporating physical information such as linear birefringence and nonlinear effects into a convolutional neural network-bidirectional long short-term memory (CNN-BiLSTM) architecture. To validate the effectiveness of the scheme, we construct an experimental platform for bending eavesdropping localization in an 80 km, 168 Gbps quadrature phase shift keying (QPSK) coherent optical communication system. Polarization data during eavesdropping are collected from nine positions with bending radii of 10.8 mm and 15 mm. The performance of three models, including CNN, CNN-BiLSTM and PINN is evaluated for bending eavesdropping localization under both bending radii. Experimental results demonstrate that the PINN achieves localization accuracies of 100% and 99.8% under bending radii of 10.8 mm and 15 mm, respectively. This work offers a novel theoretical framework and methodological approach for localizing bending eavesdropping in optical fiber communication systems.

Journal of Lightwave Technology•Feb 12, 2026•0 citations

Experimental Demonstration of Bending Eavesdropping Detection in Optical Communications Using a Physics-Informed Convolutional Network

Wenshuai Qin, Xiaoxue Gong, Weigang Hou +2 more

This paper proposes a physics-informed convolutional network (PICN) scheme to detect bending eavesdropping attacks in dual-polarization coherent optical communication systems. We present a theoretical model for optical signal transmission under bending eavesdropping, analyzing the impact of bending eavesdropping on fiber physical characteristics such as dispersion and nonlinear effect. These physical characteristics are embedded into a convolutional neural network (CNN) to construct PICN, which automatically captures subtle variations of the signal features under bending eavesdropping. To validate the effectiveness of the scheme, we first develop an eavesdropping experimental platform in an 80-km 168 Gbps dual-polarization quadrature phase shift keying (QPSK) coherent optical communication system. Polarization data are then collected under normal transmission, 10.8 mm and 15 mm bending radius. Finally, the detection performance of four classifiers including PICN, random forest (RF), support vector machine (SVM), and K-nearest neighbor (KNN) are evaluated at single and mixed bending radii. Experimental results demonstrate that PICN achieves detection accuracies of 100%, 98.53%, and 99.02% under 10.8 mm, 15 mm, and mixed bending radii, respectively. Our work provides novel theoretical foundations and innovative perspectives for bending eavesdropping detection in optical fiber communication systems.

Journal of Lightwave Technology•Dec 24, 2025•2 citations

Convolutional neural network for detecting fiber-bending eavesdropping attacks in optical communication systems

Wenshuai Qin, Xiaoxue Gong, Weigang Hou +3 more

Optical fiber communication networks are the backbone of information transmission, carrying over 90% of the world's data. However, optical fibers are susceptible to fiber-bending eavesdropping attacks, which can result in massive data leaks. The detection of fiber-bending eavesdropping attacks has become one of the effective strategies to safeguard data security. In this paper, we propose a detection scheme based on a convolutional neural network (CNN) for detecting fiber-bending eavesdropping attacks in dual-polarization coherent optical communication systems. To validate the feasibility and effectiveness of the proposed scheme, we first establish the relationship between the fiber-bending radius and the bending power loss coefficient. Based on this relationship, a 100 km, 25 GBaud dual-polarization 16 quadrature amplitude modulation (16QAM) coherent optical communication system is constructed. This system collects both normal polarization data and bending eavesdropping polarization data. Subsequently, the performance of the CNN algorithm in detecting fiber-bending eavesdropping attacks is evaluated. The results show that under bending radii of 10.8 mm, 12.1 mm, 15 mm, and mixed bending radii, the eavesdropping detection accuracy rates of the CNN reach 99.8%, 99.67%, 99.22%, and 96.15%, respectively.

Optics Express•Apr 29, 2025•5 citations

Convolutional Neural Networks for Fiber-Bending Eavesdropping Attacks Detection in Coherent Optical Communication Systems

Wenshuai Qin, Qihan Zhang, Weigang Hou +2 more

Optical fiber communication is the only technical means of supporting high transmission rates. Since optical fibers are insulators and can resist electromagnetic interference, they were once considered to be highly secure. However, due to the endless means of eavesdropping, optical fibers are vulnerable to eavesdropping, which poses a serious threat to the data security of optical fiber communication infrastructures. Optical fiber eavesdropping detection is one of the strategies to effectively safe transmission of data. Due to insufficient mining of eavesdropping information or limited by the performance of the sensor, it cannot be detected with high accuracy. To deal with this problem, we propose a convolutional neural networks-based scheme for de-tecting fiber bending eavesdropping in 16QAM coherent optical communication systems. Numerical results show 99.1%, 99%, and 98.5% detection accuracy at bending radii of 10.8 mm, 12.1 mm, and 15 mm, respectively.

International Conference on Ubiquitous Communication (Ucom)•Jul 5, 2024•7 citations

Multi-parameter deep learning for robust fiber eavesdropping detection and temporal localization

Yingbo Fan, yuang Li, Yuyuan Liang +5 more

As the backbone of global communication infrastructure, optical fibers carry vast amounts of sensitive data, making them prime targets for eavesdropping attacks that can compromise data confidentiality and national security. However, optical fibers are highly vulnerable to eavesdropping attacks-such as fiber bending or evanescent coupling-which can eavesdrop on data without interrupting the service. Accurate detection of eavesdropping remains a critical challenge, particularly under complex environmental disturbances. Environmental factors such as fluctuation and disturbances can induce signal distortions and polarization variations that mask or mimic eavesdropping characteristics, thereby posing a significant challenge to the training and generalization of machine learning-based classifiers. In this Letter, we propose a sensing framework based on a hybrid gated recurrent unit-convolutional neural network model that jointly analyzes polarization parameters and error vector magnitude. The proposed method enables accurate and robust eavesdropping detection and temporal localization under complex physical environments. Experiments conducted across seven practical scenarios demonstrate that the proposed scheme achieves a detection accuracy of 97.2%. Furthermore, we developed real-time and offline optical fiber monitoring platforms integrating mainstream detection techniques to facilitate comprehensive evaluation and deployment.

Optics Letters•Mar 12, 2026•0 citations

Physics-Informed and Convolutional Neural Network Hybrid Model for Intrusion Detection in Quantum Communication Networks

Ian Jasper Agulo, Crismar C. Calanoga, Mariah Kaliegee M. Jardinico

This study presents a comparative analysis of two deep learning architectures-a Physics-Informed Neural Network (PINN)-Convolutional Neural Network (CNN) hybrid model and a CNN-Only model-for the classification of various cyber intrusions, including Man-in-the-Middle, Quantum Attack, and Side Channel threats. The primary objective was to evaluate whether the integration of a PINN component enhances the detection performance, stability, and generalization capability of the conventional CNN approach. The results demonstrate that the PINN-CNN hybrid model achieved a higher overall test accuracy of 0.9750, marginally outperforming the CNN-Only model's accuracy of 0.9650. Crucially, the hybrid model exhibited superior robustness and generalization, evidenced by the close convergence of its training and validation loss curves, which effectively mitigates the overfitting observed in the CNN-Only model. Furthermore, the PINN-CNN hybrid achieved perfect recall (1.00) for critical attacks such as Man-in-the-Middle and Side Channel, alongside a significantly higher f1-score (0.98 vs. 0.94) for Quantum Attack detection. This work concludes that incorporating physics-informed regularization into the deep learning framework yields a more stable and reliable intrusion detection system, marking a significant advancement in developing robust cyber defenses capable of accurately identifying evolving network threats.

International Conference on Artificial Intelligence, Computer, Data Sciences and Applications (ACDSA)•Feb 5, 2026•0 citations
This paper proposes a multi-task learning based physics-informed neural network (MTL-PINN) scheme for detecting and localizing bending eavesdropping in coherent optical communication systems. The scheme employs MTL to develop a unified neural network that simultaneously addresses both the detection and localization tasks, thereby enhancing the model generalization capabilities. By integrating physical information such as linear dispersion, birefringence, and nonlinear effects into a convolutional neural network (CNN) architecture, the proposed PINN model improves both detection and localization accuracy. To validate the scheme, we construct an experimental platform for bending eavesdropping detection and localization in an 80 km, 42 GBaud quadrature phase shift keying (QPSK) coherent optical communication system. Polarization data are collected for bending radii of 10.8 mm and 15 mm under both bending eavesdropping and normal transmission scenarios. The performance of MTL-CNN, MTL-CNN-BiLSTM, and MTL-PINN models in terms of detection and localization is evaluated for both bending radii. Experimental results show that the MTL-PINN achieves localization and detection accuracies of 100% under bending radii of 10.8 mm and 15 mm. Our work provides an innovative perspective for bending eavesdropping detection and localization in optical communication systems.
This paper proposes a physics-informed neural network (PINN) scheme for localizing bending eavesdropping in coherent optical communication systems. First, we establish a signal transmission model under bending eavesdropping based on the Manakov equations to analyze its impact on physical information such as linear birefringence and nonlinearity. Subsequently, a PINN is developed by incorporating physical information such as linear birefringence and nonlinear effects into a convolutional neural network-bidirectional long short-term memory (CNN-BiLSTM) architecture. To validate the effectiveness of the scheme, we construct an experimental platform for bending eavesdropping localization in an 80 km, 168 Gbps quadrature phase shift keying (QPSK) coherent optical communication system. Polarization data during eavesdropping are collected from nine positions with bending radii of 10.8 mm and 15 mm. The performance of three models, including CNN, CNN-BiLSTM and PINN is evaluated for bending eavesdropping localization under both bending radii. Experimental results demonstrate that the PINN achieves localization accuracies of 100% and 99.8% under bending radii of 10.8 mm and 15 mm, respectively. This work offers a novel theoretical framework and methodological approach for localizing bending eavesdropping in optical fiber communication systems.
This paper proposes a physics-informed convolutional network (PICN) scheme to detect bending eavesdropping attacks in dual-polarization coherent optical communication systems. We present a theoretical model for optical signal transmission under bending eavesdropping, analyzing the impact of bending eavesdropping on fiber physical characteristics such as dispersion and nonlinear effect. These physical characteristics are embedded into a convolutional neural network (CNN) to construct PICN, which automatically captures subtle variations of the signal features under bending eavesdropping. To validate the effectiveness of the scheme, we first develop an eavesdropping experimental platform in an 80-km 168 Gbps dual-polarization quadrature phase shift keying (QPSK) coherent optical communication system. Polarization data are then collected under normal transmission, 10.8 mm and 15 mm bending radius. Finally, the detection performance of four classifiers including PICN, random forest (RF), support vector machine (SVM), and K-nearest neighbor (KNN) are evaluated at single and mixed bending radii. Experimental results demonstrate that PICN achieves detection accuracies of 100%, 98.53%, and 99.02% under 10.8 mm, 15 mm, and mixed bending radii, respectively. Our work provides novel theoretical foundations and innovative perspectives for bending eavesdropping detection in optical fiber communication systems.
Optical fiber communication networks are the backbone of information transmission, carrying over 90% of the world's data. However, optical fibers are susceptible to fiber-bending eavesdropping attacks, which can result in massive data leaks. The detection of fiber-bending eavesdropping attacks has become one of the effective strategies to safeguard data security. In this paper, we propose a detection scheme based on a convolutional neural network (CNN) for detecting fiber-bending eavesdropping attacks in dual-polarization coherent optical communication systems. To validate the feasibility and effectiveness of the proposed scheme, we first establish the relationship between the fiber-bending radius and the bending power loss coefficient. Based on this relationship, a 100 km, 25 GBaud dual-polarization 16 quadrature amplitude modulation (16QAM) coherent optical communication system is constructed. This system collects both normal polarization data and bending eavesdropping polarization data. Subsequently, the performance of the CNN algorithm in detecting fiber-bending eavesdropping attacks is evaluated. The results show that under bending radii of 10.8 mm, 12.1 mm, 15 mm, and mixed bending radii, the eavesdropping detection accuracy rates of the CNN reach 99.8%, 99.67%, 99.22%, and 96.15%, respectively.
Optical fiber communication is the only technical means of supporting high transmission rates. Since optical fibers are insulators and can resist electromagnetic interference, they were once considered to be highly secure. However, due to the endless means of eavesdropping, optical fibers are vulnerable to eavesdropping, which poses a serious threat to the data security of optical fiber communication infrastructures. Optical fiber eavesdropping detection is one of the strategies to effectively safe transmission of data. Due to insufficient mining of eavesdropping information or limited by the performance of the sensor, it cannot be detected with high accuracy. To deal with this problem, we propose a convolutional neural networks-based scheme for de-tecting fiber bending eavesdropping in 16QAM coherent optical communication systems. Numerical results show 99.1%, 99%, and 98.5% detection accuracy at bending radii of 10.8 mm, 12.1 mm, and 15 mm, respectively.
As the backbone of global communication infrastructure, optical fibers carry vast amounts of sensitive data, making them prime targets for eavesdropping attacks that can compromise data confidentiality and national security. However, optical fibers are highly vulnerable to eavesdropping attacks-such as fiber bending or evanescent coupling-which can eavesdrop on data without interrupting the service. Accurate detection of eavesdropping remains a critical challenge, particularly under complex environmental disturbances. Environmental factors such as fluctuation and disturbances can induce signal distortions and polarization variations that mask or mimic eavesdropping characteristics, thereby posing a significant challenge to the training and generalization of machine learning-based classifiers. In this Letter, we propose a sensing framework based on a hybrid gated recurrent unit-convolutional neural network model that jointly analyzes polarization parameters and error vector magnitude. The proposed method enables accurate and robust eavesdropping detection and temporal localization under complex physical environments. Experiments conducted across seven practical scenarios demonstrate that the proposed scheme achieves a detection accuracy of 97.2%. Furthermore, we developed real-time and offline optical fiber monitoring platforms integrating mainstream detection techniques to facilitate comprehensive evaluation and deployment.
This study presents a comparative analysis of two deep learning architectures-a Physics-Informed Neural Network (PINN)-Convolutional Neural Network (CNN) hybrid model and a CNN-Only model-for the classification of various cyber intrusions, including Man-in-the-Middle, Quantum Attack, and Side Channel threats. The primary objective was to evaluate whether the integration of a PINN component enhances the detection performance, stability, and generalization capability of the conventional CNN approach. The results demonstrate that the PINN-CNN hybrid model achieved a higher overall test accuracy of 0.9750, marginally outperforming the CNN-Only model's accuracy of 0.9650. Crucially, the hybrid model exhibited superior robustness and generalization, evidenced by the close convergence of its training and validation loss curves, which effectively mitigates the overfitting observed in the CNN-Only model. Furthermore, the PINN-CNN hybrid achieved perfect recall (1.00) for critical attacks such as Man-in-the-Middle and Side Channel, alongside a significantly higher f1-score (0.98 vs. 0.94) for Quantum Attack detection. This work concludes that incorporating physics-informed regularization into the deep learning framework yields a more stable and reliable intrusion detection system, marking a significant advancement in developing robust cyber defenses capable of accurately identifying evolving network threats.