arXiv subjects

Dezhang Kong

Publications and source records attributed to Dezhang Kong.

17 recordsLinked to original sources

arXiv2026

Reasoning as a Weapon: Adaptive Dual-Path Jailbreak Attack on Large Language Models

Large language models (LLMs) have achieved remarkable progress across a wide range of applications. Despite advances in safety alignment, they remain vulnerable to jailbreak attacks. Existing methods mainly rely on input-level deception to conceal harmful intent, overlooking a more fundamental attack surface: harmful intent can emerge during reasoning even when absent from the input. To study this gap, we propose ReDPJ, a Reasoning-guided Dual-Path Jailbreak framework that shifts attacks from prompt manipulation to reasoning manipulation. ReDPJ first converts harmful intent into benign dual-modal adversarial anchors, then adaptively guides reasoning to reconstruct latent harmful intent and steer the model toward unsafe trajectories. This design reduces explicit malicious signals in the input while exploiting vulnerabilities in the model's reasoning process. Experiments on state-of-the-art LLMs demonstrate strong effectiveness, achieving ASRs of 82.1% on GPT-4o-2024-11-20, 85.1% on Llama-3-8B-Instruct, and 88.2% on DeepSeek-V3.

cs.CR↗

arXiv2026

Forecasting Trajectory-Level Safety Risks in Black-Box Multi-Turn Interactions

As large language models (LLMs) evolve from standalone assistants into autonomous agents, ensuring their safety requires shifting beyond pointwise risk assessment to understand how risks emerge and unfold over long-horizon trajectories. In multi-turn interactions, malicious intent can be decomposed across seemingly harmless turns and gradually reconstructed through interaction trajectories, eventually resulting in safety failures. Existing safeguards remain largely reactive, detecting manifested violations while lacking the ability to predict latent risk evolution and enable preemptive prevention. To address this limitation, we propose Recast, a safety risk forecasting framework that advances LLM safeguarding beyond turn-level violation detection to trajectory-level risk prediction. Recast first retrieves risk-relevant evidence from both short-term dialogue progression and long-term historical context via a dual-scale trajectory view. It then models compositional risk evolution by capturing the current risk configuration and its temporal dynamics. Finally, a causal temporal encoder learns latent risk evolution patterns and predicts the distribution of future risk emergence turns. Extensive experiments across 7 risk categories show that Recast predicts 88.3% of future safety failures with an average lead time of 2.41 turns, while maintaining a false alarm rate of 12.3%, showcasing the effectiveness of trajectory-level forecasting in identifying emerging risks before safety violations occur.

cs.LG↗

arXiv2026

Before Agents Speak: Pre-hoc Failure Risk Inference in Multi-Agent Systems

LLM-based multi-agent systems (MAS) have exhibited remarkable capabilities in collaborative reasoning and decision-making, yet their interconnected communications introduce new systemic risk: localized hallucinations can propagate along agent communication chain, amplify through interactions, and ultimately trigger cascading failures. Existing countermeasures predominantly follow a post-hoc paradigm, identifying failures only after unsafe behaviors emerge, by which time harmful effects may have already spread throughout the agent network. To tackle this problem, we investigate a complementary pre-hoc approach and propose HalluProp, a Propagation-aware Hallucination inference framework that estimates individual agent failures and emergent system-level hallucination risks before inter-agent interaction. First, we model intrinsic hallucination risks by identifying fine-grained semantic misalignment between agent roles and task queries. We then characterize inter-agent risk propagation by modeling both semantic influence and communication topology. Finally, we integrate these two risks via a differentiable Noisy-OR inference mechanism to derive a systemic diagnosis. Extensive experiments show that HalluProp accurately localizes faulty agents, achieving an average AUROC of 84.6%, while enabling sub-second diagnosis with over $65\times$ speedup over post-hoc methods. By facilitating early intervention through upstream screening, HalluProp effectively complements post-hoc methods, highlighting the potential of pre-hoc risk inference for building more reliable multi-agent systems.

cs.CR↗

arXiv2026

Policy of Thoughts: Scaling Test-Time Training for LLM Reasoning via Online Policy Evolution

Large language models (LLMs) struggle with complex, long-horizon reasoning due to instability caused by their frozen policy assumption. Current test-time scaling methods treat execution feedback merely as an external signal for filtering or rewriting trajectories, without internalizing it to improve the underlying reasoning strategy. Inspired by Popper's epistemology of "conjectures and refutations," we argue that intelligence requires real-time evolution of the model's policy through learning from failed attempts. We introduce Policy of Thoughts (PoT), a framework that recasts reasoning as a within-instance online optimization process. PoT first generates diverse candidate solutions via an efficient exploration mechanism, then uses Group Relative Policy Optimization (GRPO) to update a transient LoRA adapter based on execution feedback. This closed-loop design enables dynamic, instance-specific refinement of the model's reasoning priors. Experiments show that PoT dramatically boosts performance: a 4B model achieves 49.71% accuracy on LiveCodeBench, outperforming GPT-4o and DeepSeek-V3 despite being over 50 smaller.

cs.AI↗

arXiv2026

Copyright Protection for Large Language Models: A Survey of Methods, Challenges, and Trends

Copyright protection for large language models is of critical importance, given their substantial development costs, proprietary value, and potential for misuse. Existing surveys have predominantly focused on techniques for tracing LLM-generated content-namely, text watermarking-while a systematic exploration of methods for protecting the models themselves (i.e., model watermarking and model fingerprinting) remains absent. Moreover, the relationships and distinctions among text watermarking, model watermarking, and model fingerprinting have not been comprehensively clarified. This work presents a comprehensive survey of the current state of LLM copyright protection technologies, with a focus on model fingerprinting, covering the following aspects: (1) clarifying the conceptual connection from text watermarking to model watermarking and fingerprinting, and adopting a unified terminology that incorporates model watermarking into the broader fingerprinting framework; (2) providing an overview and comparison of diverse text watermarking techniques, highlighting cases where such methods can function as model fingerprinting; (3) systematically categorizing and comparing existing model fingerprinting approaches for LLM copyright protection; (4) presenting, for the first time, techniques for fingerprint transfer and fingerprint removal; (5) summarizing evaluation metrics for model fingerprints, including effectiveness, harmlessness, robustness, stealthiness, and reliability; and (6) discussing open challenges and future research directions. This survey aims to offer researchers a thorough understanding of both text watermarking and model fingerprinting technologies in the era of LLMs, thereby fostering further advances in protecting their intellectual property.

cs.CR↗

arXiv2026

AttnDiff: Attention-based Differential Fingerprinting for Large Language Models

Protecting the intellectual property of open-weight large language models (LLMs) requires verifying whether a suspect model is derived from a victim model despite common laundering operations such as fine-tuning (including PPO/DPO), pruning/compression, and model merging. We propose \textsc{AttnDiff}, a data-efficient white-box framework that extracts fingerprints from models via intrinsic information-routing behavior. \textsc{AttnDiff} probes minimally edited prompt pairs that induce controlled semantic conflicts, captures differential attention patterns, summarizes them with compact spectral descriptors, and compares models using CKA. Across Llama-2/3 and Qwen2.5 (3B--14B) and additional open-source families, it yields high similarity for related derivatives while separating unrelated model families (e.g., $>0.98$ vs.\ $<0.22$ with $M=60$ probes). With 5--60 multi-domain probes, it supports practical provenance verification and accountability.

cs.CR↗

arXiv2026

MalURLBench: A Benchmark Evaluating Agents' Vulnerabilities When Processing Web URLs

LLM-based web agents have become increasingly popular for their utility in daily life and work. However, they exhibit critical vulnerabilities when processing malicious URLs: accepting a disguised malicious URL enables subsequent access to unsafe webpages, which can cause severe damage to service providers and users. Despite this risk, no benchmark currently targets this emerging threat. To address this gap, we propose MalURLBench, the first benchmark for evaluating LLMs' vulnerabilities to malicious URLs. MalURLBench contains 61,845 attack instances spanning 10 real-world scenarios and 7 categories of real malicious websites. Experiments with 12 popular LLMs reveal that existing models struggle to detect elaborately disguised malicious URLs. We further identify and analyze key factors that impact attack success rates and propose URLGuard, a lightweight defense module. We believe this work will provide a foundational resource for advancing the security of web agents. Our code is available at https://github.com/JiangYingEr/MalURLBench.

cs.CR↗

arXiv2026

DNF: Dual-Layer Nested Fingerprinting for Large Language Model Intellectual Property Protection

The rapid growth of large language models raises pressing concerns about intellectual property protection under black-box deployment. Existing backdoor-based fingerprints either rely on rare tokens -- leading to high-perplexity inputs susceptible to filtering -- or use fixed trigger-response mappings that are brittle to leakage and post-hoc adaptation. We propose \textsc{Dual-Layer Nested Fingerprinting} (DNF), a black-box method that embeds a hierarchical backdoor by coupling domain-specific stylistic cues with implicit semantic triggers. Across Mistral-7B, LLaMA-3-8B-Instruct, and Falcon3-7B-Instruct, DNF achieves perfect fingerprint activation while preserving downstream utility. Compared with existing methods, it uses lower-perplexity triggers, remains undetectable under fingerprint detection attacks, and is relatively robust to incremental fine-tuning and model merging. These results position DNF as a practical, stealthy, and resilient solution for LLM ownership verification and intellectual property protection.

cs.CR↗

arXiv2026

ICPO: Illocution-Calibrated Policy Optimization for Multi-Turn Conversation

Large Language Models (LLMs) in multi-turn conversations often suffer from a ``lost-in-conversation'' phenomenon, where they struggle to recover from early incorrect assumptions, particularly when users provide ambiguous initial instructions. We find that standard post-training techniques like Reinforcement Learning with Verifiable Rewards (RLVR) exacerbate this issue by rewarding confident, direct answers, thereby inducing overconfidence and discouraging the model from seeking clarification. To address this, we propose Illocution-Calibrated Policy Optimization (ICPO), a novel training framework that sensitizes the model to instruction ambiguity. ICPO augments the training corpus with underspecified prompts and conditions the reward signal on the user's illocutionary intent, rewarding the model for expressing uncertainty or asking for clarification when faced with ambiguity. Experiments demonstrate that ICPO fosters appropriate humility, yielding a substantial average improvement of 75\% in multi-turn conversation, while preserving robust performance on single-turn benchmarks. Our work presents a practical path toward more robust and collaborative conversational AI that can better navigate the nuances of human interaction.

cs.CL↗

arXiv2026

Web Fraud Attacks Against LLM-Driven Multi-Agent Systems

With the proliferation of LLM-driven multi-agent systems (MAS), the security of Web links has become a critical concern. Once MAS is induced to trust a malicious link, attackers can use it as a springboard to expand the attack surface. In this paper, we propose Web Fraud Attacks, a novel type of attack manipulating unique structures of web links to deceive MAS. We design 12 representative attack variants that encompass various methods, such as homoglyph deception, sub-directory nesting, and parameter obfuscation. Through extensive experiments on these attack vectors, we demonstrate that Web fraud attacks not only exhibit significant destructive potential across different MAS architectures but also possess a distinct advantage in evasion: they circumvent the need for complex input design, lowering the threshold for attacks significantly. These results underscore the importance of addressing Web fraud attacks, providing new insights into MAS safety. Our code is available at https://github.com/JiangYingEr/Web-Fraud-Attack-in-MAS.

cs.CR↗

arXiv2025

Scalpel-SAM: A Semi-Supervised Paradigm for Adapting SAM to Infrared Small Object Detection

Infrared small object detection urgently requires semi-supervised paradigms due to the high cost of annotation. However, existing methods like SAM face significant challenges of domain gaps, inability of encoding physical priors, and inherent architectural complexity. To address this, we designed a Hierarchical MoE Adapter consisting of four white-box neural operators. Building upon this core component, we propose a two-stage paradigm for knowledge distillation and transfer: (1) Prior-Guided Knowledge Distillation, where we use our MoE adapter and 10% of available fully supervised data to distill SAM into an expert teacher (Scalpel-SAM); and (2) Deployment-Oriented Knowledge Transfer, where we use Scalpel-SAM to generate pseudo labels for training lightweight and efficient downstream models. Experiments demonstrate that with minimal annotations, our paradigm enables downstream models to achieve performance comparable to, or even surpassing, their fully supervised counterparts. To our knowledge, this is the first semi-supervised paradigm that systematically addresses the data scarcity issue in IR-SOT using SAM as the teacher model.

cs.CV↗

arXiv2025

A Survey of LLM-Driven AI Agent Communication: Protocols, Security Risks, and Defense Countermeasures

In recent years, Large-Language-Model-driven AI agents have exhibited unprecedented intelligence and adaptability. Nowadays, agents are undergoing a new round of evolution. They no longer act as an isolated island like LLMs. Instead, they start to communicate with diverse external entities, such as other agents and tools, to perform complex tasks. Under this trend, agent communication is regarded as a foundational pillar of the next communication era, and many organizations have intensively begun to design related communication protocols (e.g., Anthropic's MCP and Google's A2A) within the past year. However, this new field exposes significant security hazards, which can cause severe damage to real-world scenarios. To help researchers quickly figure out this promising topic and benefit the future agent communication development, this paper presents a comprehensive survey of agent communication security. More precisely, we present the first clear definition of agent communication. Besides, we propose a framework that categorizes agent communication into three classes and uses a three-layered communication architecture to illustrate how each class works. Next, for each communication class, we dissect related communication protocols and analyze the security risks, illustrating which communication layer the risks arise from. Then, we provide an outlook on the possible defense countermeasures for each risk. In addition, we conduct experiments using MCP and A2A to help readers better understand the novel vulnerabilities brought by agent communication. Finally, we discuss open issues and future directions in this promising research field. We also publish a repository that maintains a list of related papers on https://github.com/theshi-1128/awesome-agent-communication-security.

cs.CR↗

arXiv2025

Fingerprint Vector: Enabling Scalable and Efficient Model Fingerprint Transfer via Vector Addition

Backdoor-based fingerprinting has emerged as an effective technique for tracing the ownership of large language models. However, in real-world deployment scenarios, developers often instantiate multiple downstream models from a shared base model, and applying fingerprinting to each variant individually incurs prohibitive computational overhead. While inheritance-based approaches -- where fingerprints are embedded into the base model and expected to persist through fine-tuning -- appear attractive, they suffer from three key limitations: late-stage fingerprinting, fingerprint instability, and interference with downstream adaptation. To address these challenges, we propose a novel mechanism called the Fingerprint Vector. Our method first embeds a fingerprint into the base model via backdoor-based fine-tuning, then extracts a task-specific parameter delta as a fingerprint vector by computing the difference between the fingerprinted and clean models. This vector can be directly added to any structurally compatible downstream model, allowing the fingerprint to be transferred post hoc without additional fine-tuning. Extensive experiments show that Fingerprint Vector achieves comparable or superior performance to direct injection across key desiderata. It maintains strong effectiveness across diverse model architectures as well as mainstream downstream variants within the same family. It also preserves harmlessness and robustness in most cases. Even when slight robustness degradation is observed, the impact remains within acceptable bounds and is outweighed by the scalability benefits of our approach.

cs.CR↗

arXiv2025

NeuRel-Attack: Neuron Relearning for Safety Disalignment in Large Language Models

Safety alignment in large language models (LLMs) is achieved through fine-tuning mechanisms that regulate neuron activations to suppress harmful content. In this work, we propose a novel approach to induce disalignment by identifying and modifying the neurons responsible for safety constraints. Our method consists of three key steps: Neuron Activation Analysis, where we examine activation patterns in response to harmful and harmless prompts to detect neurons that are critical for distinguishing between harmful and harmless inputs; Similarity-Based Neuron Identification, which systematically locates the neurons responsible for safe alignment; and Neuron Relearning for Safety Removal, where we fine-tune these selected neurons to restore the model's ability to generate previously restricted responses. Experimental results demonstrate that our method effectively removes safety constraints with minimal fine-tuning, highlighting a critical vulnerability in current alignment techniques. Our findings underscore the need for robust defenses against adversarial fine-tuning attacks on LLMs.

cs.LG↗

arXiv2023

Packet-Level Adversarial Network Traffic Crafting using Sequence Generative Adversarial Networks

The surge in the internet of things (IoT) devices seriously threatens the current IoT security landscape, which requires a robust network intrusion detection system (NIDS). Despite superior detection accuracy, existing machine learning or deep learning based NIDS are vulnerable to adversarial examples. Recently, generative adversarial networks (GANs) have become a prevailing method in adversarial examples crafting. However, the nature of discrete network traffic at the packet level makes it hard for GAN to craft adversarial traffic as GAN is efficient in generating continuous data like image synthesis. Unlike previous methods that convert discrete network traffic into a grayscale image, this paper gains inspiration from SeqGAN in sequence generation with policy gradient. Based on the structure of SeqGAN, we propose Attack-GAN to generate adversarial network traffic at packet level that complies with domain constraints. Specifically, the adversarial packet generation is formulated into a sequential decision making process. In this case, each byte in a packet is regarded as a token in a sequence. The objective of the generator is to select a token to maximize its expected end reward. To bypass the detection of NIDS, the generated network traffic and benign traffic are classified by a black-box NIDS. The prediction results returned by the NIDS are fed into the discriminator to guide the update of the generator. We generate malicious adversarial traffic based on a real public available dataset with attack functionality unchanged. The experimental results validate that the generated adversarial samples are able to deceive many existing black-box NIDS.

cs.CR↗

arXiv2021

STEP: Spatial-Temporal Network Security Event Prediction

Network security events prediction helps network operators to take response strategies from a proactive perspective, and reduce the cost caused by network attacks, which is of great significance for maintaining the security of the entire network. Most of the existing event prediction methods rely on temporal characteristics and are dedicated to exploring time series predictions, but ignoring the spatial relationship between hosts. This paper combines the temporal and spatial characteristics of security events and proposes a spatial-temporal event prediction model, named STEP. In particular, STEP formulates the security events prediction into a spatial-temporal sequence prediction. STEP utilizes graph convolution operation to capture the spatial characteristics of hosts in the network, and adopts the long short term memory (LSTM) to capture the dynamic temporal dependency of events. This paper verifies the proposed STEP scheme on two public data sets. The experimental results show that the prediction accuracy of security events under STEP is higher than that of benchmark models such as LSTM, ConvLSTM. Besides, STEP achieves high prediction accuracy when we predict events from different lengths of sequence.

cs.CR↗

arXiv2021

Machine Learning based Malicious Payload Identification in Software-Defined Networking

Deep packet inspection (DPI) has been extensively investigated in software-defined networking (SDN) as complicated attacks may intractably inject malicious payloads in the packets. Existing proprietary pattern-based or port-based third-party DPI tools can suffer from limitations in efficiently processing a large volume of data traffic. In this paper, a novel OpenFlow-enabled deep packet inspection (OFDPI) approach is proposed based on the SDN paradigm to provide adaptive and efficient packet inspection. First, OFDPI prescribes an early detection at the flow-level granularity by checking the IP addresses of each new flow via OpenFlow protocols. Then, OFDPI allows for deep packet inspection at the packet-level granularity: (i) for unencrypted packets, OFDPI extracts the features of accessible payloads, including tri-gram frequency based on Term Frequency and Inverted Document Frequency (TF-IDF) and linguistic features. These features are concatenated into a sparse matrix representation and are then applied to train a binary classifier with logistic regression rather than matching with specific pattern combinations. In order to balance the detection accuracy and performance bottleneck of the SDN controller, OFDPI introduces an adaptive packet sampling window based on the linear prediction; and (ii) for encrypted packets, OFDPI extracts notable features of packets and then trains a binary classifier with a decision tree, instead of decrypting the encrypted traffic to weaken user privacy. A prototype of OFDPI is implemented on the Ryu SDN controller and the Mininet platform. The performance and the overhead of the proposed sulotion are assessed using the real-world datasets through experiments. The numerical results indicate that OFDPI can provide a significant improvement in detection accuracy with acceptable overheads.

cs.NI↗