Breaking the Structural Identity: Personalized Federated LoRA Fine-tuning under Rank Heterogeneity
本文提出FedRoRA框架,通过分解共享全局方向和个人化秩幅度解决联邦学习中资源和数据异质性问题,提高个性化适应。
本文提出FedRoRA框架,通过分解共享全局方向和个人化秩幅度解决联邦学习中资源和数据异质性问题,提高个性化适应。
本文提出了一种基于图的生成对抗网络(GraphGAN)来检测DDoS攻击,通过生成对抗样本解决类别不平衡问题,并使用图卷积网络进行分类,提高了检测精度。
This work addresses the challenge of detecting backdoors in fine-tuned large language models under black-box settings—particularly when training data, clean reference models, or known trigger tokens are unavailable. The authors propose a “self-feeding” detection method that iteratively feeds a model’s own output back as its next input, progressively steering generated text toward the distribution of the fine-tuning data and thereby revealing latent backdoor behaviors. This approach is the first to systematically exploit an output-to-input feedback mechanism for efficient, internal-information-free triggering. Evaluated across six open-source models (3B–15B parameters) and eleven backdoor attack variants, the method achieves a model-level detection precision of 92.0%, maintains 100% precision within just four inference steps, and reduces query counts by 60%, demonstrating its effectiveness even under low single-prompt recall conditions.
This study addresses the critical security gap in large language model (LLM)-based autonomous agents endowed with real-world operational capabilities, where reasoning components remain highly vulnerable to attacks that can trigger unauthorized access, irreversible state changes, or cascading failures. Conducting a systematic literature review of 85 studies from 2023–2025 following PRISMA 2020 guidelines, this work reveals a pronounced imbalance between attack and defense research (3.9:1) and introduces the first four-layer vulnerability taxonomy for agent LLMs—encompassing perception, cognition, action, and interaction—identifying 13 vulnerability types and seven open challenges centered on isolation. Notably, perception-layer vulnerabilities account for 66% of reported issues, whereas high-risk action-layer threats such as tool misuse and sandbox escape constitute only 4.7%, highlighting a severe misalignment between research focus and actual risk. The study further identifies cross-layer vulnerability propagation due to architectural coupling as the root cause of systemic security weaknesses.
This study investigates the use of physiological signals—specifically electrodermal activity, heart rate, and skin temperature collected during examinations—to predict students’ academic performance. We systematically evaluate a range of machine learning and deep learning models, including logistic regression, random forest, support vector machines (SVM), LSTM, GRU, and Transformer architectures. Notably, this work presents the first application of the Transformer model to numerical physiological data, demonstrating predictive performance on par with LSTM and GRU. Intriguingly, under certain conditions, the comparatively simple random forest model outperforms more complex deep learning approaches. Our findings establish a quantifiable link between physiological stress responses and academic outcomes, while also offering new evidence supporting the efficacy of lightweight models in educational neuroscience applications.
本文提出FedRoRA框架,通过分解共享全局方向和个人化秩幅度解决联邦学习中资源和数据异质性问题,提高个性化适应。
本文提出了一种基于图的生成对抗网络(GraphGAN)来检测DDoS攻击,通过生成对抗样本解决类别不平衡问题,并使用图卷积网络进行分类,提高了检测精度。
This work addresses the challenge of detecting backdoors in fine-tuned large language models under black-box settings—particularly when training data, clean reference models, or known trigger tokens are unavailable. The authors propose a “self-feeding” detection method that iteratively feeds a model’s own output back as its next input, progressively steering generated text toward the distribution of the fine-tuning data and thereby revealing latent backdoor behaviors. This approach is the first to systematically exploit an output-to-input feedback mechanism for efficient, internal-information-free triggering. Evaluated across six open-source models (3B–15B parameters) and eleven backdoor attack variants, the method achieves a model-level detection precision of 92.0%, maintains 100% precision within just four inference steps, and reduces query counts by 60%, demonstrating its effectiveness even under low single-prompt recall conditions.
This study addresses the critical security gap in large language model (LLM)-based autonomous agents endowed with real-world operational capabilities, where reasoning components remain highly vulnerable to attacks that can trigger unauthorized access, irreversible state changes, or cascading failures. Conducting a systematic literature review of 85 studies from 2023–2025 following PRISMA 2020 guidelines, this work reveals a pronounced imbalance between attack and defense research (3.9:1) and introduces the first four-layer vulnerability taxonomy for agent LLMs—encompassing perception, cognition, action, and interaction—identifying 13 vulnerability types and seven open challenges centered on isolation. Notably, perception-layer vulnerabilities account for 66% of reported issues, whereas high-risk action-layer threats such as tool misuse and sandbox escape constitute only 4.7%, highlighting a severe misalignment between research focus and actual risk. The study further identifies cross-layer vulnerability propagation due to architectural coupling as the root cause of systemic security weaknesses.
This study investigates the use of physiological signals—specifically electrodermal activity, heart rate, and skin temperature collected during examinations—to predict students’ academic performance. We systematically evaluate a range of machine learning and deep learning models, including logistic regression, random forest, support vector machines (SVM), LSTM, GRU, and Transformer architectures. Notably, this work presents the first application of the Transformer model to numerical physiological data, demonstrating predictive performance on par with LSTM and GRU. Intriguingly, under certain conditions, the comparatively simple random forest model outperforms more complex deep learning approaches. Our findings establish a quantifiable link between physiological stress responses and academic outcomes, while also offering new evidence supporting the efficacy of lightweight models in educational neuroscience applications.