Data Perturbation Research Articles

Lipschitz Stability in Convex Optimization

This work is concerned with the quantitative stability of the feasible and optimal solutions of convex semi-infinite optimization problems in different parametric settings. Regarding the constraint system, we deal with convex inequalities under right-hand side (RHS), affine, and free perturbations of the original data. In these three frameworks, our focus is on quatifying the local variation of feasible solutions with respect to data perturbations (which could come, for instance, from rounding in computational processes). We formalize our analysis through the so-called Aubin property and the computation of the corresponding Lipschitz modulus. In a second stage, we tackle the Aubin property of the optimal set mapping under canonical perturbations (i.e., linear perturbations of the objective function together with RHS perturbations of the constraint system). Special attention to didactical aspects is paid; in particular, to the technics used to derive the main results gathered in this survey. Specifically, a supremum function approach and a linearization strategy based on the Fenchel-Legendre conjugate are discussed. Combining both strategies the paper provides a unified treatment (with a new proof) for the Lipschitz moduli of the feasible set mapping under RHS and affine perturbations.

A fake news detection model using the integration of multimodal attention mechanism and residual convolutional network

To improve the accuracy and efficiency of fake news detection, this study proposes a deep learning model that integrates residual networks with attention mechanisms. Building on traditional convolutional neural networks, the model incorporates multi-head attention mechanisms to enhance the extraction of key features from multimodal data such as text, images, and videos. Additionally, residual connections are introduced to deepen the network architecture, mitigate the vanishing gradient problem, and improve the model’s learning depth and stability. Compared with existing approaches, this study introduces several key innovations. First, it constructs a multimodal feature fusion module that integrates text, image, and video data. Second, it designs a cross-modal alignment mechanism to better connect information across different data types. Third, it optimizes the feature fusion structure for more effective integration. Finally, the study employs attention mechanisms to highlight and enhance the representation of salient features. Experiments were conducted using three representative datasets: the LIAR dataset for political short texts, the FakeNewsNet dataset for English multimodal news, and the Weibo dataset from a Chinese social media platform. These were selected to comprehensively evaluate the model’s performance across different scenarios. Baseline models used for comparison include Bidirectional Encoder Representations from Transformers (BERT), Robustly Optimized Bidirectional Encoder Representations from Transformers Approach (RoBERTa), Generalized Autoregressive Pretraining for Language Understanding (XLNet), Enhanced Representation through Knowledge Integration (ERNIE), and Generative Pre-trained Transformer 3.5 (GPT-3.5). In terms of four key performance metrics—accuracy, precision, recall, and F1 score—the proposed model achieved best-case values of 0.977, 0.986, 0.969, and 0.924, respectively, outperforming the aforementioned baseline models overall. Furthermore, simulated experiments were conducted to evaluate the model’s real-world applicability from four dimensions: robustness, generalization ability, response time, and resource consumption. The results demonstrate that the model maintains strong stability and adaptability under data perturbations and diverse input conditions, with a response time controllable within 0.02 s. The model also shows significant computational advantages when handling large-scale datasets. Therefore, this study presents a high-performance and deployment-friendly solution for fake news detection in multimodal contexts. The study also offers valuable theoretical insights and practical guidance for applying deep learning to public opinion governance and text classification.

Locally Differentially Private Frequency Estimation via Joint Randomized Response

Local Differential Privacy (LDP) has been widely recognized as a powerful tool for providing a strong theoretical guarantee of data privacy to data contributors against an untrusted data collector. Under a typical LDP scheme, each data contributor independently randomly perturbs their data before submitting them to the data collector, which in turn infers valuable statistics about the original data from received perturbed data. Common to existing LDP mechanisms is an inherent trade-off between the level of privacy protection and data utility in the sense that strong data privacy often comes at the cost of reduced data utility. Frequency estimation based on Randomized Response (RR) is a fundamental building block of many LDP mechanisms. In this paper, we propose a novel Joint Randomized Response (JRR) mechanism based on correlated data perturbations to achieve locally differentially private frequency estimation. JRR divides data contributors into disjoint groups of two members and lets those in the same group jointly perturb their binary data to improve frequency-estimation accuracy and achieve the same level of data privacy by hiding the group membership information in contrast to the classical RR mechanism. Theoretical analysis and detailed simulation studies using both real and synthetic datasets show that JRR achieves the same level of data privacy as the classical RR mechanism while improving the frequency-estimation accuracy in the overwhelming majority of the cases by up to two orders of magnitude.

RNA-ligand interaction scoring via data perturbation and augmentation modeling.

Despite recent advances in RNA-targeting drug discovery, the development of data-driven deep learning models remains challenging owing to limited validated RNA-small molecule interaction data and scarce known RNA structures. In this context, we introduce RNAsmol, a sequence-based deep learning framework that incorporates data perturbation with augmentation, graph-based molecular feature representation and attention-based feature fusion modules to predict RNA-small molecule interactions. RNAsmol employs perturbation strategies to balance the bias between the true negative and unknown interaction space, thereby elucidating the intrinsic binding patterns between RNA and small molecules. The resulting model demonstrates accurate predictions of the binding between RNA and small molecules, outperforming other methods in ten-fold cross-validation, unseen evaluation and decoy evaluation. Moreover, we use case studies to visualize molecular binding profiles and the distribution of learned weights, providing interpretable insights into RNAsmol's predictions. In particular, without requiring structural input, RNAsmol can generate reliable predictions and be adapted to various drug design scenarios.

An accuracy-privacy optimization framework considering user’s privacy requirements for data stream mining

Data stream mining is a critical process utilized by organizations to derive insights from real-time data. Consequently, preserving the privacy of sensitive information while maintaining high accuracy remains a persistent challenge. Privacy-preserving data mining techniques modify data to increase privacy, a process that invariably decreases the accuracy of data mining algorithms. Though different techniques have been proposed to preserve privacy, there is a lack of well-formulated frameworks to optimize the trade-off between accuracy and privacy. This paper introduces a novel Accuracy-Privacy Optimization Framework (APOF) that allows users to define privacy requirements and predicts achievable accuracy levels, enabling fine-tuning of this balance. The logistic cumulative noise addition was used as the data perturbation method that has experimentally shown better performance and Hoeffding trees as the classifier. Additionally, a data fitting module using kernel regression is integrated, a unique approach that predicts accuracy levels based on user-defined privacy thresholds. Experimental results show that the proposed framework archives an optimal privacy level above 97% while minimising the accuracy loss across various datasets. By addressing critical gaps in privacy-preserving data mining, this study offers significant contributions to real-world applications, facilitating secure and efficient data utilization in dynamic environments.

Improving surrogate model robustness to perturbations for dynamical systems through machine learning and data assimilation

Many real-world systems are modelled using complex ordinary differential equations (ODEs). However, the dimensionality of these systems can make them challenging to analyse. Dimensionality reduction techniques like proper orthogonal decomposition (POD) can be used in such cases. However, these reduced-order models (ROMs) are susceptible to perturbations in the input. We propose a novel framework that combines machine learning (ML) and data assimilation (DA) techniques to improving surrogate models to handle perturbations in input data effectively. Through rigorous experiments on dynamical systems modelled on graphs, we demonstrate that our framework substantially improves the accuracy of surrogate models under input perturbations. Furthermore, we evaluate the framework’s efficacy on alternative surrogate models, including neural ODEs, and the empirical results consistently show enhanced performance.

Boosting 3D Object Detection with Adversarial Adaptive Data Augmentation Strategy

In real-world applications, autonomous driving systems need to handle a variety of complex scenarios, such as object occlusion and lighting changes. In these scenarios, accurately identifying various objects is crucial for perceiving the surrounding environment and making reliable decisions. In this context, the fusion of Lidar and cameras is vital for the accuracy of object detection. To this end, we propose an adversarial adaptive data augmentation strategy that introduces virtual adversarial perturbations during the image feature extraction process, effectively enhancing the robustness of 3D object detection methods and enabling them to maintain stable performance when facing environmental changes and data perturbations. Experimental results on the nuScenes-mini and KITTI datasets show that, compared with previous 3D object detection methods, our method not only improves detection accuracy but also demonstrates stronger stability.

A Robustness Study of Multi-Layer Perceptrons and Logistic Regression to Data Perturbation: MNIST Dataset

This study systematically evaluates the robustness of Multi-Layer Perceptrons (MLPs) And Logistic Regression (LR) models against data pertubations using the MNIST handwritten digit dataset. While MLPs and LR are foundational in machine learning, their comparative resilience to diverse pertubations-noise, geometric distortions, and adversarial attacks-remains underexplored,despite implications for real-world applications with imperfect data., whe test three pertubations categories : Gaussian noise (σ=0.1 to 1.0), salt and pepper noise (p=0.1 to 0.5), rotational distorsions (5° to 30°), and adversial attacks (FGSM with ϵ=0.005 to0.30). both models were trained on 60.000 MNIST samples and tested on 10.000 pertubed images. Results demonstrate that MLPs exhibit superior robustness under moderate noise and rotations, achieving baseline accuracies of 97.07% (vs. LR’s 92.63%). For Gaussian noise (σ=0.5), MLP retained 35.35% accuracy compared to LR’s 23.91% . however, adversarial attacks (FGSM, ϵ= 0.30) reduced MLP accuracy to 0.20%, revealing critical vulnerabilities. Statistical analysis (paired t-test, p < 0.05) confirmed significant performance differences across pertubations levels. Alinear regressions (R^2 = 0.98) further quantified MLP’s predictable accuracy decline with Gaussian noise intensity. These findings underscore MLP’s suitability for noise-prone environments but highlight urgent needs for adversarial defense mechanisms. Practitioners are advised to prioritize MLPs for tasks with moderate distortions, while future work should integrate robustness enhancements like adversarial training.

Exposome-Scale Investigation of Cl-/Br-Containing Chemicals Using High-Resolution Mass Spectrometry, Multistage Machine Learning, and Cloud Computing.

Over 70% of organic halogens, representing chlorine- and bromine-containing disinfection byproducts (Cl-/Br-DBPs), remain unidentified after 50 years of research. This work introduces a streamlined and cloud-based exposomics workflow that integrates high-resolution mass spectrometry (HRMS) analysis, multistage machine learning, and cloud computing for efficient analysis and characterization of Cl-/Br-DBPs. In particular, the multistage machine learning structure employs progressively different heavy isotopic peaks at each layer and capture the distinct isotopic characteristics of nonhalogenated compounds and Cl-/Br-compounds at different halogenation levels. This innovative approach enables the recognition of 22 types of Cl-/Br-compounds with up to 6 Br and 8 Cl atoms. To address the data imbalance among different classes, particularly the limited number of heavily chlorinated and brominated compounds, data perturbation is performed to generate hypothetical/synthetic molecular formulas containing multiple Cl and Br atoms, facilitating data augmentation. To further benefit the environmental chemistry community with limited computational experience and hardware access, above innovations are incorporated into HalogenFinder (http://www.halogenfinder.com/), a user-friendly, web-based platform for Cl-/Br-compound characterization, with statistical analysis support via MetaboAnalyst. In the benchmarking, HalogenFinder outperformed two established tools, achieving a higher recognition rate for 277 authentic Cl-/Br-compounds and uniquely identifying the number of Cl/Br atoms. In laboratory tests of DBP mixtures, it identified 72 Cl-/Br-DBPs with proposed structures, of which eight were confirmed with chemical standards. A retrospective analysis of 2022 finished water HRMS data revealed insightful temporal trends in Cl-DBP features. These results demonstrate HalogenFinder's effectiveness in advancing Cl-/Br-compound identification for environmental science and exposomics.

PonziHunter: Hunting Ethereum Ponzi Contract via Static Analysis and Contrastive Learning on the Bytecode Level

In recent years, blockchain technology has developed rapidly and received widespread attention. However, its pseudonymous and decentralized nature has also attracted many criminal activities. Ponzi schemes, a kind of classic financial scam, also hide their true face in smart contracts, causing massive financial losses to blockchain users. Although several methods have been proposed to detect Ponzi contracts, there are still limitations in broad applicability, semantics understanding, and adversarial robustness. In this paper, we propose PonziHunter, an intelligent framework for hunting Ponzi contracts on Ethereum. To tackle the problem of broad applicability, we train a detection model that does not require expert experience based on publicly available on-chain bytecode and off-chain contract labels. To tackle the problem of semantics understanding, we employ cross-function control flows and state variable dependencies to understand the logic of Ponzi contracts. Specifically, we decompile bytecodes into higher-order representations to analyze control flows and state variable dependencies and model the information as graph data. By combining the idea of code slicing, we identify the basic blocks related to Ponzi contract recognition. To tackle the problem of adversarial robustness, we model Ponzi contract recognition as a graph classification problem based on contrastive pre-training. We propose a data augmentation method for control flow graphs (CFGs), which preserves the basic blocks related to Ponzi contract recognition as much as possible during data perturbation. Experimental results show that PonziHunter outperforms state-of-the-art tools with average improvements of at least 4.77% on real-world ground-truth data, and can newly discover 85 Ponzi contracts in the wild. More importantly, PonziHunter is robust against adversarial examples and can locate the critical basic blocks for smart Ponzi detection.

A Robust Electricity Demand Forecasting of Rajshahi Metropolitan City of Bangladesh

Analysis of electricity demand serves as a foundation to understand the trend of demand and related variability which helps to plan for the distribution of electricity to a region. Fuzzy linear regression uses fuzzy parameters to represent ambiguous and imprecise relationships between dependent and independent variables which omits the limitations of conventional linear model. In this paper, robust optimization has been used to increase the feasibility while working with uncertainty in input data. A combined Interval-Ellipsoidal Robust Counterpart of the fuzzy linear regression model is used to analyze the data. To formulate the Robust Fuzzy Linear Regression model (RFLR), the quantity of customers and the average yearly temperature have been regarded as independent variables, while the consumption of electricity has been regarded as a dependent variable. The proposed method has been applied to forecast electricity demand of Rajshahi City. The optimal solution of the linear model has been obtained using the Lingo software. The results show the capability of the RFLR to analyze data uncertainty with more accuracy and it is demonstrated that the inaccuracy in forecasting and estimated fuzzy bands grow as the data perturbation increases.

Fault diagnosis of axle-box bearing in electric multiple unit based on transfer learning network with feedback control input strategy

Abstract Aiming at several problems in the practical application of axle-box bearing fault diagnosis for electric multiple unit, a fault diagnosis method based on a transfer learning (TL) network with a feedback control input strategy is proposed. First, given the unavailability of large amounts of labeled data, the proposed method employs a TL network as its central framework, incorporating a lightweight architecture. Maximum mean discrepancy and classification loss are also introduced to align the extracted features, allowing the method to perform fault diagnosis across various bearings under different operating conditions. Then, considering the various perturbations in the vibration data, an adaptive input strategy combining signal processing methods is proposed to screen out important information through denoising and deharmonizing, thereby enhancing input data quality. Furthermore, a feedback control input strategy is proposed to reduce redundant health information input, inspired by feedback control principles. Finally, a fault identification strategy is implemented to recognize fault types in unknown signals. The accuracy and practicality of the proposed method are validated through experiments on three datasets and random fault vibration signals.

Discriminative Consistency Semi‐Supervised Carotid Ultrasound Plaque Segmentation by Exploiting Global Context

ABSTRACTCarotid plaques in ultrasound images are a routine indicator for stroke accident risk evaluation. However, plaque segmentation for diagnosis is a difficult task because artifacts and heterogeneity can obfuscate the plaque boundaries. Moreover, pixel‐level labeling of numerous images can be time‐consuming and laborious. In this paper, we propose a discriminative consistency semi‐supervised method by employing global contexts, named DCGC‐Net, to segment carotid ultrasound plaques. Firstly, student‐teacher consistency learning is adopted to leverage unlabeled images using data perturbations. However, the unsupervised outputs may suffer from a lack of shape constraint. Thus, we introduce an adversarial network to enforce the outputs of unlabeled images more reliably. Finally, a global dilated convolution block (GDCB), embedded in U‐Net, is designed to obtain global contexts for reducing the effect of artifacts. Extensive experiments are performed on 1400 images of 1259 patients using 1/2, 1/4, and 1/8 labeled training images. Compared to cutting‐edge semi‐supervised methods, the proposed method can acquire more outstanding results on metrics of DSC and MHD (p value < 0.05). Ablation experiments demonstrate the validity of each proposed module. Besides, plaque clinical parameters are automatedly calculated as a short diagnostic report. Our proposed semi‐supervised method can be useful for clinically segmenting carotid ultrasound plaques by using limited labeled images and numerous unlabeled images.

Bi-attribute utility preference robust optimization: A continuous piecewise linear approximation approach

Bi-attribute utility preference robust optimization: A continuous piecewise linear approximation approach

On the Robustness of Individual Tree Segmentation to Data Adversarial Attacks from Remote Sensing Point Clouds

Forests play a vital role in maintaining ecological balance, making accurate forest monitoring technologies essential. Remote sensing point cloud data always capture distinctive geometric features of forests, including the cylindrical symmetry of tree trunks and the radial symmetry of canopies. However, the inherent complexity of point cloud data, combined with their vulnerabilities to adversarial attacks, often disrupts these symmetrical patterns, significantly limiting the practical application of deep learning models in forest monitoring. This research presents a novel approach to enhance the robustness of individual tree segmentation networks by combining data augmentation and adversarial training techniques. Our method employs the FGSM algorithm and Gaussian noise attack to generate adversarial samples while utilizing data denoising and controlled noise injection for data augmentation. A dynamic adversarial training framework can adaptively adjust the proportion of adversarial samples during the network training stage to optimize the model. Using remote sensing point cloud datasets from Wisconsin, the experimental results demonstrate the effectiveness of the individual tree segmentation networks, PointNet++ and DBSCAN, in reducing attack success rates whilst improving the stability and accuracy of segmentation results under various adversarial conditions. This study highlights the potential for more robust forest monitoring systems capable of maintaining accuracy even when faced with data perturbations or intentional interference.

A Robust Behavioral Biometrics Framework for Smartphone Authentication via Hybrid Machine Learning and TOPSIS

Significant vulnerabilities in traditional authentication systems have been demonstrated due to the high dependence on smartphone hardware devices to execute many different and complicated tasks. PINs, passwords, and static biometric techniques have been shown to be subjected to various serious attacks, such as environmental limitations, spoofing, and brute force attacks, and this in turn mitigates the security level of the entire system. In this study, a robust framework for smartphone authentication is presented. Touch dynamic pattern recognitions, including trajectory curvature, touch pressure, acceleration, two-dimensional spatial coordinates, and velocity, have been extracted and assessed as behavioral biometric features. The TOPSIS (Technique for Order of Preference by Similarity to Ideal Solution) methodology has also been incorporated to obtain the most affected and valuable features, which are then fed as input to three different Machine Learning (ML) algorithms: Random Forest (RF), Gradient Boosting Machines (GBM), and K-Nearest Neighbors (KNN). Our analysis, supported by experimental results, ensure that the RF model outperforms the two other ML algorithms by getting F1-Score, accuracy, recall, and precision of 95.1%, 95.2%, 95.5%, and 94.8%, respectively. In order to further increase the resiliency of the proposed technique, the data perturbation approach, including temporal scaling and noise insertion, has been augmented. Also, the proposal has been shown to be resilient against both environmental variation-based attacks by achieving accuracy above 93% and spoofing attacks by obtaining a detection rate of 96%. This emphasizes that the proposed technique provides a promising solution to many authentication issues and offers a user-friendly and scalable method to improve the security of the smartphone against cybersecurity attacks.

Stability and global analysis of the Vlasov–Maxwell–Einstein system in astrophysical plasmas

This paper provides an analysis of the relativistic Vlasov–Maxwell–Einstein (RVME) system, which models the dynamics of relativistic plasmas in the presence of strong gravitational fields. We establish the existence and uniqueness of global classical solutions using a fixed-point iteration method combined with a priori estimates. The stability of these solutions is analyzed with respect to small perturbations in the initial data to derive explicit stability bounds. The implications of our results are briefly discussed in the context of astrophysical scenarios, particularly focusing on the dynamics of plasmas near compact objects such as black holes and neutron stars.

LXR: Learning to eXplain Recommendations

Recommender systems have become integral to many online services, leveraging user data to provide personalized recommendations. However, as these systems grow in complexity, understanding the rationale behind their recommendations becomes increasingly difficult. Explainable Artificial Intelligence (XAI) has emerged as a crucial field addressing this challenge, particularly in ensuring transparency and trustworthiness in automated decision-making processes. In this paper, we introduce Learning to eXplain Recommendations (LXR) , a scalable, model-agnostic framework designed to generate counterfactually correct explanations for recommender systems. LXR generates explanations for recommendations produced by any differentiable recommender system. By leveraging both factual and counterfactual loss terms, LXR offers robust, accurate, and computationally efficient explanations that reflect the model’s internal decision-making process. A key feature of LXR is its focus on the factual correctness of explanations through counterfactual reasoning, bridging the gap between plausible and accurate explanations. Unlike traditional approaches that rely on exhaustive perturbations of user data, LXR uses a self-supervised learning method to generate explanations efficiently, without sacrificing accuracy. LXR operates in two stages: a pre-training step and a novel Inference-Time Fine-tuning (ITF) step that refines explanations at the individual recommendation level, significantly improving accuracy with minimal computational overhead. Additionally, LXR is applied to hybrid recommender models incorporating demographic data, demonstrating its versatility across real-world scenarios. Finally, we also showcase LXR’s ability to explain recommendations at various ranks within a user’s recommendation list. As a secondary contribution, we introduce several novel evaluation metrics, inspired by saliency maps from computer vision, to rigorously assess the counterfactual correctness of explanations in recommender systems. Our results demonstrate that LXR sets a new benchmark for explainability, providing accurate, transparent, and interpretable explanations. The code is available on our GitHub repository: https://github.com/DeltaLabTLV/LXR.

Abstract 2343: CancerSpot: a multi-cancer early detection test using targeted methylation sequencing

Abstract Background: Liquid biopsy-based cancer detection is a noninvasive method to identify cell-free DNA (cfDNA) fragment based signatures in the bloodstream characterized by distinct methylation patterns. Tumor-derived fragments (ctDNA) form a tiny fraction of all cfDNA fragments in the bloodstream. ctDNA originating from some specific genomic regions can have altered methylation patterns. Reliable and cost-effective detection of altered methylation pattern in these regions requires a broad targeted sequencing approach. Methods: We developed CancerSpot, a multi-cancer early detection (MCED) test based on cfDNA methylation signatures in a cohort of 688 cancer patients and 326 healthy controls representing 10 cancer types. A custom genomics workflow prioritizes high-quality reads by rejecting poorly aligned or imperfectly converted data while rescuing reliable reads to calculate methylation scores across genomic regions at the read level accounting for detection of a weak signal from tumor derived DNA. Machine learning (ML) models were developed using methylation signatures from ∼7000 genomic markers. A gradient-boosted trees approach with repeated cross-validation incorporated data perturbations to account for variability in lab protocols and sequencing platforms ensuring robust prediction accuracy and error tolerance. Results: CancerSpot achieved 79% sensitivity and 97% specificity for detecting Stage I-III tumors across 10 cancer types. It also demonstrated a 77% accuracy in identifying the tumor tissue of origin within the top two predictions. The robust workflow enhances detection of tumor-derived cfDNA signals even in samples with low conversion efficiency while mitigating background noise from normal cfDNA. The regions used by the test show strong differential methylation in tumor-normal data from a diverse population in TCGA indicating its wide applicability. Conclusions: CancerSpot provides a cost-effective and accurate MCED solution tailored to diverse cohorts. Citation Format: Urvashi Bahadur, Divya Vishwanath, Prakash Hiremath, Nihesh Rathod, Swaraj Basu, Sampuran Chakraborty, Aastha Tripathi, Preetha R T, Aditi Chatterjee, Arunima Ghosh, Sameer Phalke, Subuhi Yete, Ankita Khan, Sweta Sanguri, Mohammed Rahman, Sewanti Limaye, Charles Cantor, Vamsi Veeramachaneni, Ramesh Hariharan. CancerSpot: a multi-cancer early detection test using targeted methylation sequencing [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 2343.

Enhanced Physical Layer Security Based on Joint Encryption of Similar Polarization State Rotation Perturbations and 6-D Hyperchaotic Scrambling for Polarization Division Multiplexing System

ABSTRACT The paper proposes an optical physical layer encryption scheme based on the combination of similar rotation of the state of polarization (SRSOP) and a six-dimensional (6D) hyperchaotic system. By introducing similar polarization state rotation to disturb vector data at the transmitting end, the security of optical transmission systems can be effectively enhanced. In addition, the data perturbation of polarization state rotation, combined with the 6D hyperchaotic system with unpredictability and sensitivity, can extend the key space of the encryption scheme to 10180, making it resistant to brute force attacks. The encryption process of this scheme is as follows: Firstly, the chaotic sequences generated by the 6D hyperchaotic system were used to encrypt various parameters in the polarization encryption matrix, and then these parameters were used to generate the corresponding polarization encryption matrix. Finally, the encryption matrix was used to perturb the data to be sent. The scheme was successfully tested on a polarization division multiplexing (PDM)-QPSK system at a rate of 112 Gbps. The results indicate that authorized users can correctly decrypt the encrypted information, while unauthorized users can only obtain useful information with a bit error rate (BER) of approximately 0.5, verifying the effectiveness and security of the encryption scheme.