arXiv subjects

Xinyang Deng

Publications and source records attributed to Xinyang Deng.

At least 19 recordsLinked to original sources

arXiv2026

Quantum Incremental Learning with Mixed State Prototypes

Incremental learning models are required to learn new classes sequentially without catastrophic forgetting, while operating under parameter and memory constraints. In the Noisy Intermediate-Scale Quantum (NISQ) era, although quantum neural networks offer advantages in feature mapping, hardware limitations restrict circuit width. Furthermore, traditional quantum classifiers are constrained by the number of orthogonal basis states, limiting their capacity to accommodate a continually growing number of categories. Thus, we introduce a novel quantum incremental learning framework based on trainable mixed-state prototypes. Its original design incorporates new classes by adding class prototypes rather than increasing the circuit width of the shared quantum backbone. The use of mixed-state prototypes is another key contribution, since they have representation capabilities to represent information than a single pure-state prototype. And the decomposable mixed-state calculation provides lower production costs and a convenient Hilbert-Schmidt (HS) distance metric for classification. Simulation results show that our model achieves high-dimensional feature concentration using a minimal number of qubits, while demonstrating lower computational complexity and robust representation in incremental learning tasks compared with classical baselines.

cs.AI↗

arXiv2026

DataRx: Missingness-Aware Sampling for Safer Large Language Model Task-Specific Fine-Tuning

Task-specific fine-tuning can improve the performance of large language models (LLMs) on downstream tasks. However, our study reveals that task-specific fine-tuning can also weaken the safety guardrails of aligned LLMs. A widely adopted strategy for preserving safety during fine-tuning is to incorporate safety data. Although previous studies have shown that randomly mixing safety data can alleviate safety degradation, the underlying principle determining why some safety examples are more effective than others still remains unclear. In this paper, we propose DataRx, a missingness-aware sampling method for selecting safety-critical examples. DataRx is based on the hypothesis that a safety sample is more effective when the selected examples provide safety signals that fill the missing parts of LLMs' safety capabilities. DataRx's key insight is leveraging high-dimensional hidden representations rather than discrete tokens to quantify the safety signal gap between the target model's native response and the safety reference response. The results show that, with only 1% additional safety samples from BeaverTails, DataRx reduces the average attack success rate of Llama3-8B-Instruct across seven downstream tasks from 59.23% under random sampling to 13.70%. In addition, DataRx can be combined with the existing safety data synthesis method to further enhance safety defenses during fine-tuning. We hope that DataRx will inspire more data-centric defense research.

cs.CL↗

arXiv2026

DOW-KE: Anchor-Free Multi-Layer Knowledge Editing via Direct End-to-End Weight Optimization

Multi-layer locate-then-edit methods for knowledge editing first optimize target residual-stream activations (anchors) at selected layers, then realize them layer by layer as weight updates. This pipeline optimizes an intermediate representation but deploys multi-layer weight updates whose joint effect through the true forward pass is never itself optimized: regardless of how anchors are set or propagated, each update comes from a local solve, so propagation-induced attenuation and distortion go uncorrected, leaving a closure gap between anchor targets and realized edits. We propose DOW-KE, an anchor-free method built on a single principle: what is optimized must be exactly what is deployed. DOW-KE backpropagates the final editing objective through the complete model, jointly optimizing the updates of all edited layers so cross-layer propagation and coupling enter every gradient step. The same principle dictates where preservation resides: embedding the preservation projection in the update parameterization, inside the computation graph, makes every gradient act on the deployed update; post-hoc constraints would reopen the gap, and the constrained search keeps edits clear of protected knowledge. In large-scale sequential editing on two datasets and three models, DOW-KE achieves the highest overall Score and neighborhood Specificity in five of six model-dataset settings among the evaluated baselines.

cs.LG↗

arXiv2026

DataShield: Safety-degrading Data Filtering for LLM Benign Instruction Fine-Tuning

Large language models (LLMs) suffer from degraded safety capabilities even when fine-tuned with benign datasets. However, existing methods for identifying safety-degrading samples in benign datasets suffer from high computational costs and significant noise issues. In this paper, we propose DataShield to efficiently and effectively identify potential safety-degrading samples. Our key intuition is based on the observation that benign fine-tuning increases the overall response compliance of LLMs. DataShield's key technical insight is to quantify each sample's contribution to the model's compliance behavior as its safety degradation score. DataShield consists of three core components: (1) Compliance Vector Extraction, which captures the LLM's compliance behavior tendency; (2) a novel Compliance-Aware Score (CAS), which automatically identifies the optimal safety-critical layer; and (3) Safety-degrading Sample Filtering, which quantifies the projection shift of training data along the compliance direction. Extensive experimental evaluation on Llama3-8B, Llama3.1-8B, and Qwen2.5-7B using the Alpaca and Dolly benign datasets validates our method's effectiveness in identifying high-risk and low-risk data subsets. We also observe that open-ended question answering is more likely to trigger safety degradation, and corresponding responses tend to be longer. We hope this work can provide new insights into data-centric defense methods. The source code is available at: https://github.com/ZJunBo/DataShield.

cs.CR↗

arXiv2026

Feature Entanglement-based Quantum Multimodal Fusion Neural Network

Multimodal learning aims to enhance perceptual and decision-making capabilities by integrating information from diverse sources. However, classical deep learning approaches face a critical trade-off between the high accuracy of black-box feature-level fusion and the interpretability of less outstanding decision-level fusion, alongside the challenges of parameter explosion and complexity. This paper discusses the accuracy-interpretablity-complexity dilemma under the quantum computation framework and propose a feature entanglement-based quantum multimodal fusion neural network. The model is composed of three core components: a classical feed-forward module for unimodal processing, an interpretable quantum fusion block, and a quantum convolutional neural network (QCNN) for deep feature extraction. By leveraging the strong expressive power of quantum, we have reduced the complexity of multimodal fusion and post-processing to linear, and the fusion process also possesses the interpretability of decision-level fusion. The simulation results demonstrate that our model achieves classification accuracy comparable to classical networks with dozens of times of parameters, exhibiting notable stability and performance across multimodal image datasets.

quant-ph↗

arXiv2025

Understanding and Mitigating Over-refusal for Large Language Models via Representation Intervention

Large language models (LLMs) demonstrate powerful capabilities across various natural language processing tasks,yet their inherent safety vulnerabilities undermine the reliable application of LLMs in real-world scenarios. To enhance LLM safety, various jailbreak defense methods have been proposed to guard against harmful outputs. However, improvements in model safety often come at the cost of severe over-refusal, failing to strike a good balance between safety and usability. This phenomenon is a critical reliability degradation issue in LLM intelligent systems, failing to strike a good balance between safety defense effectiveness and system usability reliability. In this paper, we first analyze the causes of over-refusal from a representation perspective, revealing that LLMs are unable to effectively distinguish between over-refusal samples and malicious samples. Based on this, we propose to mitigate overrefusal by intervening in the safety representation space of LLMs. Our method incorporates two core strategies: (1) OverlapAware Loss Weighting, which determines the erasure weight for malicious samples by quantifying their similarity to overrefusal samples in the representation space, and (2) ContextAware Augmentation, which supplements the necessary context for rejection decisions by adding harmful prefixes before rejection responses. Experiments demonstrate that our method achieves a better trade-off between mitigating over-refusal and maintaining safety, compared with existing approaches. This paper also aims to encourage researchers to consider the reliability of defending methods against jailbreak attacks from both the perspectives of safety and over-refusal.

cs.CR↗

arXiv2024

Sample-agnostic Adversarial Perturbation for Vision-Language Pre-training Models

Recent studies on AI security have highlighted the vulnerability of Vision-Language Pre-training (VLP) models to subtle yet intentionally designed perturbations in images and texts. Investigating multimodal systems' robustness via adversarial attacks is crucial in this field. Most multimodal attacks are sample-specific, generating a unique perturbation for each sample to construct adversarial samples. To the best of our knowledge, it is the first work through multimodal decision boundaries to explore the creation of a universal, sample-agnostic perturbation that applies to any image. Initially, we explore strategies to move sample points beyond the decision boundaries of linear classifiers, refining the algorithm to ensure successful attacks under the top $k$ accuracy metric. Based on this foundation, in visual-language tasks, we treat visual and textual modalities as reciprocal sample points and decision hyperplanes, guiding image embeddings to traverse text-constructed decision boundaries, and vice versa. This iterative process consistently refines a universal perturbation, ultimately identifying a singular direction within the input space which is exploitable to impair the retrieval performance of VLP models. The proposed algorithms support the creation of global perturbations or adversarial patches. Comprehensive experiments validate the effectiveness of our method, showcasing its data, task, and model transferability across various VLP models and datasets. Code: https://github.com/LibertazZ/MUAP

cs.CV↗

arXiv2024

A Unified Understanding of Adversarial Vulnerability Regarding Unimodal Models and Vision-Language Pre-training Models

With Vision-Language Pre-training (VLP) models demonstrating powerful multimodal interaction capabilities, the application scenarios of neural networks are no longer confined to unimodal domains but have expanded to more complex multimodal V+L downstream tasks. The security vulnerabilities of unimodal models have been extensively examined, whereas those of VLP models remain challenging. We note that in CV models, the understanding of images comes from annotated information, while VLP models are designed to learn image representations directly from raw text. Motivated by this discrepancy, we developed the Feature Guidance Attack (FGA), a novel method that uses text representations to direct the perturbation of clean images, resulting in the generation of adversarial images. FGA is orthogonal to many advanced attack strategies in the unimodal domain, facilitating the direct application of rich research findings from the unimodal to the multimodal scenario. By appropriately introducing text attack into FGA, we construct Feature Guidance with Text Attack (FGA-T). Through the interaction of attacking two modalities, FGA-T achieves superior attack effects against VLP models. Moreover, incorporating data augmentation and momentum mechanisms significantly improves the black-box transferability of FGA-T. Our method demonstrates stable and effective attack capabilities across various datasets, downstream tasks, and both black-box and white-box settings, offering a unified baseline for exploring the robustness of VLP models.

cs.CV↗

arXiv2020

Basic concepts, definitions, and methods in D number theory

As a generalization of Dempster-Shafer theory, D number theory (DNT) aims to provide a framework to deal with uncertain information with non-exclusiveness and incompleteness. Although there are some advances on DNT in previous studies, however, they lack of systematicness, and many important issues have not yet been solved. In this paper, several crucial aspects in constructing a perfect and systematic framework of DNT are considered. At first the non-exclusiveness in DNT is formally defined and discussed. Secondly, a method to combine multiple D numbers is proposed by extending previous exclusive conflict redistribution (ECR) rule. Thirdly, a new pair of belief and plausibility measures for D numbers are defined and many desirable properties are satisfied by the proposed measures. Fourthly, the combination of information-incomplete D numbers is studied specially to show how to deal with the incompleteness of information in DNT. In this paper, we mainly give relative math definitions, properties, and theorems, concrete examples and applications will be considered in the future study.

cs.AI↗

arXiv2019

Belief and plausibility measures for D numbers

As a generalization of Dempster-Shafer theory, D number theory provides a framework to deal with uncertain information with non-exclusiveness and incompleteness. However, some basic concepts in D number theory are not well defined. In this note, the belief and plausibility measures for D numbers have been proposed, and basic properties of these measures have been revealed as well.

cs.AI↗

arXiv2019

On the negation of a Dempster-Shafer belief structure based on maximum uncertainty allocation

Probability theory and Dempster-Shafer theory are two germane theories to represent and handle uncertain information. Recent study suggested a transformation to obtain the negation of a probability distribution based on the maximum entropy. Correspondingly, determining the negation of a belief structure, however, is still an open issue in Dempster-Shafer theory, which is very important in theoretical research and practical applications. In this paper, a negation transformation for belief structures is proposed based on maximum uncertainty allocation, and several important properties satisfied by the transformation have been studied. The proposed negation transformation is more general and could totally compatible with existing transformation for probability distributions.

cs.AI↗

arXiv2017

A total uncertainty measure for D numbers based on belief intervals

As a generalization of Dempster-Shafer theory, the theory of D numbers is a new theoretical framework for uncertainty reasoning. Measuring the uncertainty of knowledge or information represented by D numbers is an unsolved issue in that theory. In this paper, inspired by distance based uncertainty measures for Dempster-Shafer theory, a total uncertainty measure for a D number is proposed based on its belief intervals. The proposed total uncertainty measure can simultaneously capture the discord, and non-specificity, and non-exclusiveness involved in D numbers. And some basic properties of this total uncertainty measure, including range, monotonicity, generalized set consistency, are also presented.

cs.AI↗

arXiv2017

D numbers theory based game-theoretic framework in adversarial decision making under fuzzy environment

Adversarial decision making is a particular type of decision making problem where the gain a decision maker obtains as a result of his decisions is affected by the actions taken by others. Representation of alternatives' evaluations and methods to find the optimal alternative are two important aspects in the adversarial decision making. The aim of this study is to develop a general framework for solving the adversarial decision making issue under uncertain environment. By combining fuzzy set theory, game theory and D numbers theory (DNT), a DNT based game-theoretic framework for adversarial decision making under fuzzy environment is presented. Within the proposed framework or model, fuzzy set theory is used to model the uncertain evaluations of decision makers to alternatives, the non-exclusiveness among fuzzy evaluations are taken into consideration by using DNT, and the conflict of interests among decision makers is considered in a two-person non-constant sum game theory perspective. An illustrative application is given to demonstrate the effectiveness of the proposed model. This work, on one hand, has developed an effective framework for adversarial decision making under fuzzy environment; One the other hand, it has further improved the basis of DNT as a generalization of Dempster-Shafer theory for uncertainty reasoning.

cs.AI↗

arXiv2017

Exploring the Combination Rules of D Numbers From a Perspective of Conflict Redistribution

Dempster-Shafer theory of evidence is widely applied to uncertainty modelling and knowledge reasoning because of its advantages in dealing with uncertain information. But some conditions or requirements, such as exclusiveness hypothesis and completeness constraint, limit the development and application of that theory to a large extend. To overcome the shortcomings and enhance its capability of representing the uncertainty, a novel model, called D numbers, has been proposed recently. However, many key issues, for example how to implement the combination of D numbers, remain unsolved. In the paper, we have explored the combination of D Numbers from a perspective of conflict redistribution, and proposed two combination rules being suitable for different situations for the fusion of two D numbers. The proposed combination rules can reduce to the classical Dempster's rule in Dempster-Shafer theory under a certain conditions. Numerical examples and discussion about the proposed rules are also given in the paper.

cs.AI↗

arXiv2015

Impact of Preference and Equivocators on Opinion Dynamics with Evolutionary Game Framework

Opinion dynamics, aiming to understand the evolution of collective behavior through various interaction mechanisms of opinions, represents one of the most challenges in natural and social science. To elucidate this issue clearly, binary opinion model becomes a useful framework, where agents can take an independent opinion. Inspired by the realistic observations, here we propose two basic interaction mechanisms of binary opinion model: one is the so-called BSO model in which players benefit from holding the same opinion; the other is called BDO model in which players benefit from taking different opinions. In terms of these two basic models, the synthetical effect of opinion preference and equivocators on the evolution of binary opinion is studied under the framework of evolutionary game theory (EGT), where the replicator equation (RE) is employed to mimick the evolution of opinions. By means of numerous simulations, we show the theoretical equilibrium states of binary opinion dynamics, and mathematically analyze the stability of each equilibrium state as well.

cs.GT↗

arXiv2015

A quantum extension to inspection game

Quantum game theory is a new interdisciplinary field between game theory and physical research. In this paper, we extend the classical inspection game into a quantum game version by quantizing the strategy space and importing entanglement between players. Our result shows that the quantum inspection game has various Nash equilibrium depending on the initial quantum state of the game. It is also shown that quantization can respectively help each player to increase his own payoff, yet fails to bring Pareto improvement for the collective payoff in the quantum inspection game.

cs.GT↗

arXiv2015

Quantum games of opinion formation based on the Marinatto-Weber quantum game scheme

Quantization becomes a new way to study classical game theory since quantum strategies and quantum games have been proposed. In previous studies, many typical game models, such as prisoner's dilemma, battle of the sexes, Hawk-Dove game, have been investigated by using quantization approaches. In this paper, several game models of opinion formations have been quantized based on the Marinatto-Weber quantum game scheme, a frequently used scheme to convert classical games to quantum versions. Our results show that the quantization can change fascinatingly the properties of some classical opinion formation game models so as to generate win-win outcomes.

cs.GT↗

arXiv2015

Transformation of basic probability assignments to probabilities based on a new entropy measure

Dempster-Shafer evidence theory is an efficient mathematical tool to deal with uncertain information. In that theory, basic probability assignment (BPA) is the basic element for the expression and inference of uncertainty. Decision-making based on BPA is still an open issue in Dempster-Shafer evidence theory. In this paper, a novel approach of transforming basic probability assignments to probabilities is proposed based on Deng entropy which is a new measure for the uncertainty of BPA. The principle of the proposed method is to minimize the difference of uncertainties involving in the given BPA and obtained probability distribution. Numerical examples are given to show the proposed approach.

cs.AI↗