Layerwise Convergence Fingerprints for Runtime Misbehavior Detection in Large Language Models
大型语言模型在运行时可能表现出各种异常行为，例如训练时注入的后门在触发词下被激活、越狱攻击绕过安全对齐、提示注入覆盖部署者指令。现有的运行时防御通常针对单一威胁，且依赖干净参考模型、触发知识或可编辑权重，这些假设对不透明的第三方模型往往不成
#cs.AI
共收录 1693 条相关情报。
← 返回所有主题GAMMAF: A Common Framework for Graph-Based Anomaly Monitoring Benchmarking in LLM Multi-Agent Systems
该论文针对大型语言模型（LLM）在多智能体系统（MAS）中集成所带来的安全挑战，提出了一种名为Gammaf（Graph-based Anomaly Monitoring for LLM Multi-Agent systems Framewo
Unveiling the Backdoor Mechanism Hidden Behind Catastrophic Overfitting in Fast Adversarial Training
本文旨在解释快速对抗训练（Fast Adversarial Training, FAT）中出现的灾难性过拟合（Catastrophic Overfitting, CO）现象。FAT能高效提升神经网络对对抗样本的鲁棒性，但容易发生CO，即模型
X-NegoBox: An Explainable Privacy-Budget Negotiation Framework for Secure Peer-to-Peer Energy Data Exchange
论文提出 X-NegoBox，一个用于点对点能源数据安全交换的透明隐私预算协商框架。现代能源系统正从集中式向分布式转变，消费者转变为产消者，持续与聚合器、对等节点和市场运营商交换数据。虽然这些数据对点对点交易、需求响应和分布式预测至关重要，
Agentic Witnessing: Pragmatic and Scalable TEE-Enabled Privacy-Preserving Auditing
该论文针对审计专有数据语义属性时的隐私与透明矛盾，提出了一种名为“Agentic Witnessing”的框架。传统方法如零知识证明（ZKP）适用于精确代数约束，但难以验证定性、非结构化属性（如代码库中的逻辑）。该框架将验证从可证明执行扩展
Defusing the Trigger: Plug-and-Play Defense for Backdoored LLMs via Tail-Risk Intrinsic Geometric Smoothing
Defending against backdoor attacks in large language models remains a critical practical challenge. Existing defenses mi
The Kerimov-Alekberli Model: An Information-Geometric Framework for Real-Time System Stability
This study introduces the Kerimov-Alekberli model, a novel information-geometric framework that redefines AI safety by f
Jailbreaking Frontier Foundation Models Through Intention Deception
Large (vision-)language models exhibit remarkable capability but remain highly susceptible to jailbreaking. Existing saf
An Information-Geometric Framework for Stability Analysis of Large Language Models under Entropic Stress
As large language models (LLMs) are increasingly deployed in high-stakes and operational settings, evaluation strategies
ArmSSL: Adversarial Robust Black-Box Watermarking for Self-Supervised Learning Pre-trained Encoders
提出一种名为ArmSSL的框架，用于对自监督学习预训练编码器进行黑盒可验证且对抗鲁棒的水印保护，在不影响主任务效用的前提下实现知识产权防护。
SSG: Logit-Balanced Vocabulary Partitioning for LLM Watermarking
提出SSG方法，通过对数几率平衡的词汇分区提升LLM水印在低熵场景（代码生成、数学推理）下的检测能力。
PrivSTRUCT: Untangling Data Purpose Compliance of Privacy Policies in Google Play Store
提出PrivSTRUCT框架，通过编码器-解码器结构解析隐私政策中的数据结构，揭示Google Play商店应用中数据目的声明与实际数据流之间的透明度差距。
Transient Turn Injection: Exposing Stateless Multi-Turn Vulnerabilities in Large Language Models
提出一种名为Transient Turn Injection (TTI)的新型多轮攻击技术，通过跨隔离交互分布恶意意图来利用大语言模型的无状态审核漏洞。