arXiv subjects

Yisheng Zhong

Publications and source records attributed to Yisheng Zhong.

18 recordsLinked to original sources

arXiv2026

CALIBURN: Self-Calibrated LLM Unlearning Alignment

LLM unlearning aims to remove the influence of undesirable knowledge from pretrained language models, which offers a practical mechanism for addressing safety and privacy concerns. Existing unlearning approaches, such as Gradient Ascent, are prone to catastrophic forgetting. Alignment-based approaches provide an alternative direction, yet their effectiveness is limited by the quality of the reference model. In realistic settings, both methods still require large retention datasets to preserve general knowledge. We propose a principled method that quantifies the target LLM's confidence in undesirable knowledge and uses it to calibrate the model's unlearning gradient updates more precisely. It enables fine-grained control over forgetting while better preserving model utility, thus reducing the dependence on retention data or prohibitive unlearning training data. Extensive evaluations on multiple benchmarks, including MUSE and WMDP, show that our method achieves effective unlearning and improves the trade-off between knowledge removal and utility preservation compared with state-of-the-art methods.

cs.CL↗

arXiv2026

Harmonizing Multi-Objective LLM Unlearning via Unified Domain Representation and Bidirectional Logit Distillation

Large Language Models (LLMs) unlearning is crucial for removing hazardous or privacy-leaking information from the model. Practical LLM unlearning demands satisfying multiple challenging objectives simultaneously: removing undesirable knowledge, preserving general utility, avoiding over-refusal of neighboring concepts, and, crucially, ensuring robustness against adversarial probing attacks. However, existing unlearning methods primarily focus on a limited subset of these goals, typically unlearning efficacy and utility preservation while overlooking robustness and boundary behaviors. Naively extending these methods to multi-objective settings may lead to unlearning task interference. We propose a novel multi-objective unlearning framework that harmonizes multiple unlearning objectives through a data and optimization co-design: We standardize training corpora into a unified data representation to reduce the domain gap, and then introduce a bidirectional distillation method that simultaneously elicits desired behavior from a context-instructed teacher while suppressing undesirable behavior in the student model. Theoretical and empirical analyses show that our method aligns domain distributions and converts seemingly irrelevant unlearning tasks into cooperative optimization. Evaluation demonstrates state-of-the-art performance, which enables balanced and reliable unlearning across diverse, challenging requirements.

cs.LG↗

arXiv2026

DUET: Distilled LLM Unlearning from an Efficiently Contextualized Teacher

LLM unlearning is a technique to remove the impacts of undesirable knowledge from the model without retraining from scratch, which is indispensable towards trustworthy AI. Existing unlearning methods face significant limitations: conventional tuning-based unlearning is computationally heavy and prone to catastrophic forgetting. In contrast, in-contextualized unlearning is lightweight for precise unlearning but vulnerable to prompt removal or reverse engineering attacks. In response, we propose Distilled Unlearning from an Efficient Teacher (DUET), a novel distillation-based unlearning method that combines the merits of these two lines of work. It learns a student model to imitate the behavior of a prompt-steered teacher that effectively refuses undesirable knowledge generation while preserving general domain knowledge. Extensive evaluations on existing benchmarks with our enriched evaluation protocols demonstrate that DUET achieves higher performance in both forgetting and utility preservation, while being orders of magnitude more data-efficient than state-of-the-art unlearning methods.

cs.LG↗

arXiv2026

Reach-avoid games for players with damped double integrator dynamics

This paper investigates a reach-avoid game between two players with damped double integrator dynamics. An optimal state-feedback strategy is derived using a differential game framework combined with geometric analysis. To facilitate the analysis, we introduce the concept of multiple reachable region by characterizing the motion of players with damped double integrator dynamics. Based on this, a new type of the attackers dominance region is introduced. We show that distinct strategies are required depending on the location of the terminal position within different areas of the attacker's dominance region. Furthermore, we prove that the proposed strategies satisfy the necessary condition for optimality. Numerical simulations are provided to illustrate the conclusions.

eess.SY↗

arXiv2025

Hierarchical Federated Unlearning for Large Language Models

Large Language Models (LLMs) are increasingly integrated into real-world applications, raising concerns about privacy, security and the need to remove undesirable knowledge. Machine Unlearning has emerged as a promising solution, yet faces two key challenges: (1) practical unlearning needs are often continuous and heterogeneous, and (2) they involve decentralized, sensitive data with asymmetric access. These factors result in inter-domain and intra-domain interference, which further amplifies the dilemma of unbalanced forgetting and retaining performance. In response, we propose a federated unlearning approach for LLMs that is scalable and privacy preserving. Our method decouples unlearning and retention via task-specific adapter learning and employs a hierarchical merging strategy to mitigate conflicting objectives and enables robust, adaptable unlearning updates. Comprehensive experiments on benchmarks of WMDP, MUSE, and TOFU showed that our approach effectively handles heterogeneous unlearning requests while maintaining strong LLM utility compared with baseline methods.

cs.LG↗

arXiv2025

Web Intellectual Property at Risk: Preventing Unauthorized Real-Time Retrieval by Large Language Models

The protection of cyber Intellectual Property (IP) such as web content is an increasingly critical concern. The rise of large language models (LLMs) with online retrieval capabilities enables convenient access to information but often undermines the rights of original content creators. As users increasingly rely on LLM-generated responses, they gradually diminish direct engagement with original information sources, which will significantly reduce the incentives for IP creators to contribute, and lead to a saturating cyberspace with more AI-generated content. In response, we propose a novel defense framework that empowers web content creators to safeguard their web-based IP from unauthorized LLM real-time extraction and redistribution by leveraging the semantic understanding capability of LLMs themselves. Our method follows principled motivations and effectively addresses an intractable black-box optimization problem. Real-world experiments demonstrated that our methods improve defense success rates from 2.5% to 88.6% on different LLMs, outperforming traditional defenses such as configuration-based restrictions.

cs.CR↗

arXiv2023

PROFL: A Privacy-Preserving Federated Learning Method with Stringent Defense Against Poisoning Attacks

Federated Learning (FL) faces two major issues: privacy leakage and poisoning attacks, which may seriously undermine the reliability and security of the system. Overcoming them simultaneously poses a great challenge. This is because privacy protection policies prohibit access to users' local gradients to avoid privacy leakage, while Byzantine-robust methods necessitate access to these gradients to defend against poisoning attacks. To address these problems, we propose a novel privacy-preserving Byzantine-robust FL framework PROFL. PROFL is based on the two-trapdoor additional homomorphic encryption algorithm and blinding techniques to ensure the data privacy of the entire FL process. During the defense process, PROFL first utilize secure Multi-Krum algorithm to remove malicious gradients at the user level. Then, according to the Pauta criterion, we innovatively propose a statistic-based privacy-preserving defense algorithm to eliminate outlier interference at the feature level and resist impersonation poisoning attacks with stronger concealment. Detailed theoretical analysis proves the security and efficiency of the proposed method. We conducted extensive experiments on two benchmark datasets, and PROFL improved accuracy by 39% to 75% across different attack settings compared to similar privacy-preserving robust methods, demonstrating its significant advantage in robustness.

cs.CR↗

arXiv2022

Evaluation and Learning in Two-Player Symmetric Games via Best and Better Responses

Artificial intelligence and robotic competitions are accompanied by a class of game paradigms in which each player privately commits a strategy to a game system which simulates the game using the collected joint strategy and then returns payoffs to players. This paper considers the strategy commitment for two-player symmetric games in which the players' strategy spaces are identical and their payoffs are symmetric. First, we introduce two digraph-based metrics at a meta-level for strategy evaluation in two-agent reinforcement learning, grounded on sink equilibrium. The metrics rank the strategies of a single player and determine the set of strategies which are preferred for the private commitment. Then, in order to find the preferred strategies under the metrics, we propose two variants of the classical learning algorithm self-play, called strictly best-response and weakly better-response self-plays. By modeling learning processes as walks over joint-strategy response digraphs, we prove that the learnt strategies by two variants are preferred under two metrics, respectively. The preferred strategies under both two metrics are identified and adjacency matrices induced by one metric and one variant are connected. Finally, simulations are provided to illustrate the results.

cs.GT↗

arXiv2021

Multiplayer Homicidal Chauffeur Reach-Avoid Games via Guaranteed Winning Strategies

This paper studies a planar multiplayer Homicidal Chauffeur reach-avoid differential game, where each pursuer is a Dubins car and each evader has simple motion. The pursuers aim to protect a goal region cooperatively from the evaders. Due to the high-dimensional strategy space among pursuers, we decompose the whole game into multiple one-pursuer-one-evader subgames, each of which is solved in an analytical approach instead of solving Hamilton-Jacobi-Isaacs equations. For each subgame, an evasion region (ER) is introduced, based on which a pursuit strategy guaranteeing the winning of a simple-motion pursuer under specific conditions is proposed. Motivated by the simple-motion pursuer, a strategy for a Dubins-car pursuer is proposed when the pursuer-evader configuration satisfies separation condition (SC) and interception orientation (IO). The necessary and sufficient condition on capture radius, minimum turning radius and speed ratio to guarantee the pursuit winning is derived. When the IO is not satisfied (Non-IO), a heading adjustment pursuit strategy is proposed, and the condition to achieve IO within a finite time, is given. Then, a two-step pursuit strategy is proposed for the SC and Non-IO case. A non-convex optimization problem is introduced to give a condition guaranteeing the winning of the pursuer. A polynomial equation gives a lower bound of the non-convex problem, providing a sufficient and efficient pursuit winning condition. Finally, these pairwise outcomes are collected for the pursuer-evader matching. Simulations are provided to illustrate the theoretical results.

cs.GT↗

arXiv2020

Policy Evaluation and Seeking for Multi-Agent Reinforcement Learning via Best Response

This paper introduces two metrics (cycle-based and memory-based metrics), grounded on a dynamical game-theoretic solution concept called sink equilibrium, for the evaluation, ranking, and computation of policies in multi-agent learning. We adopt strict best response dynamics (SBRD) to model selfish behaviors at a meta-level for multi-agent reinforcement learning. Our approach can deal with dynamical cyclical behaviors (unlike approaches based on Nash equilibria and Elo ratings), and is more compatible with single-agent reinforcement learning than alpha-rank which relies on weakly better responses. We first consider settings where the difference between largest and second largest underlying metric has a known lower bound. With this knowledge we propose a class of perturbed SBRD with the following property: only policies with maximum metric are observed with nonzero probability for a broad class of stochastic games with finite memory. We then consider settings where the lower bound for the difference is unknown. For this setting, we propose a class of perturbed SBRD such that the metrics of the policies observed with nonzero probability differ from the optimal by any given tolerance. The proposed perturbed SBRD addresses the opponent-induced non-stationarity by fixing the strategies of others for the learning agent, and uses empirical game-theoretic analysis to estimate payoffs for each strategy profile obtained due to the perturbation.

cs.GT↗

arXiv2020

Matching-Based Capture Strategies for 3D Heterogeneous Multiplayer Reach-Avoid Differential Games

This paper studies a 3D multiplayer reach-avoid differential game with a goal region and a play region. Multiple pursuers defend the goal region by consecutively capturing multiple evaders in the play region. The players have heterogeneous moving speeds and the pursuers have heterogeneous capture radii. Since this game is hard to analyze directly, we decompose the whole game as many subgames which involve multiple pursuers and only one evader. Then, these subgames are used as a building block for the pursuer-evader matching. First, for multiple pursuers and one evader, we introduce an evasion space (ES) method characterized by a potential function to construct a guaranteed pursuer winning strategy. Then, based on this strategy, we develop conditions to determine whether a pursuit team can guard the goal region against one evader. It is shown that in 3D, if a pursuit team is able to defend the goal region against an evader, then at most three pursuers in the team are necessarily needed. We also compute the value function of the Hamilton-Jacobi-Isaacs (HJI) equation for a special subgame of degree. To capture the maximum number of evaders in the open-loop sense, we formulate a maximum bipartite matching problem with conflict graph (MBMC). We show that the MBMC is NP-hard and design a polynomial-time constant-factor approximation algorithm to solve it. Finally, we propose a receding horizon strategy for the pursuit team where in each horizon an MBMC is solved and the strategies of the pursuers are given. We also extend our results to the case of a bounded convex play region where the evaders escape through an exit. Two numerical examples are provided to demonstrate the obtained results.

math.OC↗

arXiv2019

Guarding a Subspace in High-Dimensional Space with Two Defenders and One Attacker

This paper considers a subspace guarding game in high-dimensional space which consists of a play subspace and a target subspace. Two faster defenders cooperate to protect the target subspace by capturing an attacker which strives to enter the target subspace from the play subspace without being captured. A closed-form solution is provided from the perspectives of kind and degree. Contributions of the work include the use of the attack subspace (AS) method to construct the barrier, by which the game winner can be perfectly predicted before the game starts. In addition to this inclusion, with the priori information about the game result, a critical payoff function is designed when the defenders can win the game. Then, the optimal strategy for each player is explicitly reformulated as a saddle-point equilibrium. Finally, we apply these theoretical results to a half-space guarding game in three-dimensional space. Since the whole achieved developments are analytical, they require a little memory without computational burden and allow for real-time updates, beyond the capacity of traditional Hamilton-Jacobi-Isaacs method. It is worth noting that this is the first time in the current work to consider the target guarding games for arbitrary high-dimensional space, and in a fully analytical form.

eess.SY↗

arXiv2019

Construction of the Barrier for Reach-Avoid Differential Games in Three-Dimensional Space with Four Equal-Speed Players

This paper considers a reach-avoid differential game in three-dimensional space with four equal-speed players. A plane divides the game space into a play subspace and a goal subspace. The evader aims at entering the goal subspace while three pursuers cooperate to prevent that by capturing the evader. A complete, closed-form barrier for this differential game is provided, by which the game winner can be perfectly predicted before the game starts. All possible cooperations among three pursuers are considered and thus the guaranteed winning for each team is a prior. Furthermore, an algorithm is designed to compute the barrier for multiple pursuers of any numbers and any initial configurations. More realistically, since the whole achieved developments are analytical, they require a little memory without computational burden and allow for real-time updates, beyond the capacity of traditional Hamilton-Jacobi-Isaacs method.

math.OC↗

arXiv2019

Task Assignment for Multiplayer Reach-Avoid Games in Convex Domains via Analytical Barriers

This work considers a multiplayer reach-avoid game between two adversarial teams in a general convex domain which consists of a target region and a play region. The evasion team, initially lying in the play region, aims to send as many its team members into the target region as possible, while the pursuit team with its team members initially distributed in both play region and target region, strives to prevent that by capturing the evaders. We aim at investigating a task assignment about the pursuer-evader matching, which can maximize the number of the evaders who can be captured before reaching the target region safely when both teams play optimally. To address this, two winning regions for a group of pursuers to intercept an evader are determined by constructing an analytical barrier which divides these two parts. Then, a task assignment to guarantee the most evaders intercepted is provided by solving a simplified 0-1 integer programming instead of a non-deterministic polynomial problem, easing the computation burden dramatically. It is worth noting that except the task assignment, the whole analysis is analytical. Finally, simulation results are also presented.

math.OC↗

arXiv2016

Person Re-Identification by Unsupervised Video Matching

Most existing person re-identification (ReID) methods rely only on the spatial appearance information from either one or multiple person images, whilst ignore the space-time cues readily available in video or image-sequence data. Moreover, they often assume the availability of exhaustively labelled cross-view pairwise data for every camera pair, making them non-scalable to ReID applications in real-world large scale camera networks. In this work, we introduce a novel video based person ReID method capable of accurately matching people across views from arbitrary unaligned image-sequences without any labelled pairwise data. Specifically, we introduce a new space-time person representation by encoding multiple granularities of spatio-temporal dynamics in form of time series. Moreover, a Time Shift Dynamic Time Warping (TS-DTW) model is derived for performing automatically alignment whilst achieving data selection and matching between inherently inaccurate and incomplete sequences in a unified way. We further extend the TS-DTW model for accommodating multiple feature-sequences of an image-sequence in order to fuse information from different descriptions. Crucially, this model does not require pairwise labelled training data (i.e. unsupervised) therefore readily scalable to large scale camera networks of arbitrary camera pairs without the need for exhaustive data annotation for every camera pair. We show the effectiveness and advantages of the proposed method by extensive comparisons with related state-of-the-art approaches using two benchmarking ReID datasets, PRID2011 and iLIDS-VID.

cs.CV↗

arXiv2012

Sufficient and Necessary Condition of Admissibility for Fractional-order Singular System

This paper has been withdrawn. This paper focuses on the admissibility condition for fractional-order singular system with order $α\in (0,1)$. The definitions of regularity, impulse-free and admissibility are given first, then a sufficient and necessary condition of admissibility for fractional-order singular system is established. A numerical example is included to illustrate the proposed condition.

eess.SY↗

arXiv2012

Robust stability for fractional-order systems with structured and unstructured uncertainties

The issues of robust stability for two types of uncertain fractional-order systems of order $α\in (0,1)$ are dealt with in this paper. For the polytope-type uncertainty case, a less conservative sufficient condition of robust stability is given; for the norm-bounded uncertainty case, a sufficient and necessary condition of robust stability is presented. Both of these conditions can be checked by solving sets of linear matrix inequalities. Two numerical examples are presented to confirm the proposed conditions.

eess.SY↗

arXiv2009

Markov Chain Hidden behind Power Law Mechanism of Self-organized Criticality

To describe and analyze the dynamics of Self-Organized Criticality (SOC) systems, a four-state continuous-time Markov model is proposed in this paper. Different to computer simulation or numeric experimental approaches commonly employed for explaining the power law in SOC, in this paper, based on this Markov model, using E.T.Jayness' Maximum Entropy method, we have derived a mathematical proof on the power law distribution for the size of these events. Both this Makov model and the mathematical proof on power law present a new angle on the universality of power law distributions, they also show that the scale free property exists not necessary only in SOC system, but in a class of dynamical systems which can be modelled by the proposed Markov model.

cond-mat.stat-mech↗