Evaluating Privacy Leakage and Memorization Attacks on Large Language Models (LLMs) in Generative AI Applications

The recent interest in the deployment of Generative AI applications that use large language models (LLMs) has brought to the forefront significant privacy concerns, notably the leakage of Personally Identifiable Information (PII) and other confidential or protected information that may have been memorized during training, specifically during a fine-tuning or customization process. We describe different black-box attacks from potential adversaries and study their impact on the amount and type of information that may be recovered from commonly used and deployed LLMs. Our research investigates the relationship between PII leakage, memorization, and factors such as model size, architecture, and the nature of attacks employed. The study utilizes two broad categories of attacks: PII leakage-focused attacks (auto-completion and extraction attacks) and memorization-focused attacks (various membership inference attacks). The findings from these investigations are quantified using an array of evaluative metrics, providing a detailed understanding of LLM vulnerabilities and the effectiveness of different attacks.

Security Vulnerability Analyses of Large Language Models (LLMs) through Extension of the Common Vulnerability Scoring System (CVSS) Framework

Large Language Models (LLMs) have revolutionized Generative Artificial Intelligence (GenAI) tasks, becoming an integral part of various applications in society, including text generation, translation, summarization, and more. However, their widespread usage emphasizes the critical need to enhance their security posture to ensure the integrity and reliability of their outputs and minimize harmful effects. Prompt injections and training data poisoning attacks are two of the most prominent vulnerabilities in LLMs, which could potentially lead to unpredictable and undesirable behaviors, such as biased outputs, misinformation propagation, and even malicious content generation. The Common Vulnerability Scoring System (CVSS) framework provides a standardized approach to capturing the principal characteristics of vulnerabilities, facilitating a deeper understanding of their severity within the security and AI communities. By extending the current CVSS framework, we generate scores for these vulnerabilities such that organizations can prioritize mitigation efforts, allocate resources effectively, and implement targeted security measures to defend against potential risks.

Enhancing Relational Triple Extraction in Specific Domains:Semantic Enhancement and Synergy of Large Language Models and Small Pre-Trained Language Models

In the process of constructing domain-specific knowledge graphs,the task of relational triple extraction plays a critical role in transforming unstructured text into structured information.Existing relational triple extraction models facemultiple challenges when processing domain-specific data,including insufficient utilization of semantic interaction information between entities and relations,difficulties in handling challenging samples,and the scarcity of domain-specific datasets.To address these issues,our study introduces three innovative components:Relation semantic enhancement,data augmentation,and a voting strategy,all designed to significantly improve the model’s performance in tackling domain-specific relational triple extraction tasks.We first propose an innovative attention interaction module.This method significantly enhances the semantic interaction capabilities between entities and relations by integrating semantic information fromrelation labels.Second,we propose a voting strategy that effectively combines the strengths of large languagemodels(LLMs)and fine-tuned small pre-trained language models(SLMs)to reevaluate challenging samples,thereby improving the model’s adaptability in specific domains.Additionally,we explore the use of LLMs for data augmentation,aiming to generate domain-specific datasets to alleviate the scarcity of domain data.Experiments conducted on three domain-specific datasets demonstrate that our model outperforms existing comparative models in several aspects,with F1 scores exceeding the State of the Art models by 2%,1.6%,and 0.6%,respectively,validating the effectiveness and generalizability of our approach.

LKPNR: Large Language Models and Knowledge Graph for Personalized News Recommendation Framework

Accurately recommending candidate news to users is a basic challenge of personalized news recommendation systems.Traditional methods are usually difficult to learn and acquire complex semantic information in news texts,resulting in unsatisfactory recommendation results.Besides,these traditional methods are more friendly to active users with rich historical behaviors.However,they can not effectively solve the long tail problem of inactive users.To address these issues,this research presents a novel general framework that combines Large Language Models(LLM)and Knowledge Graphs(KG)into traditional methods.To learn the contextual information of news text,we use LLMs’powerful text understanding ability to generate news representations with rich semantic information,and then,the generated news representations are used to enhance the news encoding in traditional methods.In addition,multi-hops relationship of news entities is mined and the structural information of news is encoded using KG,thus alleviating the challenge of long-tail distribution.Experimental results demonstrate that compared with various traditional models,on evaluation indicators such as AUC,MRR,nDCG@5 and nDCG@10,the framework significantly improves the recommendation performance.The successful integration of LLM and KG in our framework has established a feasible way for achieving more accurate personalized news recommendation.Our code is available at https://github.com/Xuan-ZW/LKPNR.

Potential use of large language models for mitigating students’problematic social media use:ChatGPT as an example

The problematic use of social media has numerous negative impacts on individuals'daily lives,interpersonal relationships,physical and mental health,and more.Currently,there are few methods and tools to alleviate problematic social media,and their potential is yet to be fully realized.Emerging large language models(LLMs)are becoming increasingly popular for providing information and assistance to people and are being applied in many aspects of life.In mitigating problematic social media use,LLMs such as ChatGPT can play a positive role by serving as conversational partners and outlets for users,providing personalized information and resources,monitoring and intervening in problematic social media use,and more.In this process,we should recognize both the enormous potential and endless possibilities of LLMs such as ChatGPT,leveraging their advantages to better address problematic social media use,while also acknowledging the limitations and potential pitfalls of ChatGPT technology,such as errors,limitations in issue resolution,privacy and security concerns,and potential overreliance.When we leverage the advantages of LLMs to address issues in social media usage,we must adopt a cautious and ethical approach,being vigilant of the potential adverse effects that LLMs may have in addressing problematic social media use to better harness technology to serve individuals and society.

DeBERTa-GRU: Sentiment Analysis for Large Language Model

Modern technological advancements have made social media an essential component of daily life.Social media allow individuals to share thoughts,emotions,and ideas.Sentiment analysis plays the function of evaluating whether the sentiment of the text is positive,negative,neutral,or any other personal emotion to understand the sentiment context of the text.Sentiment analysis is essential in business and society because it impacts strategic decision-making.Sentiment analysis involves challenges due to lexical variation,an unlabeled dataset,and text distance correlations.The execution time increases due to the sequential processing of the sequence models.However,the calculation times for the Transformer models are reduced because of the parallel processing.This study uses a hybrid deep learning strategy to combine the strengths of the Transformer and Sequence models while ignoring their limitations.In particular,the proposed model integrates the Decoding-enhanced with Bidirectional Encoder Representations from Transformers(BERT)attention(DeBERTa)and the Gated Recurrent Unit(GRU)for sentiment analysis.Using the Decoding-enhanced BERT technique,the words are mapped into a compact,semantic word embedding space,and the Gated Recurrent Unit model can capture the distance contextual semantics correctly.The proposed hybrid model achieves F1-scores of 97%on the Twitter Large Language Model(LLM)dataset,which is much higher than the performance of new techniques.

Smaller & Smarter: Score-Driven Network Chaining of Smaller Language Models

With the continuous evolution and expanding applications of Large Language Models (LLMs), there has been a noticeable surge in the size of the emerging models. It is not solely the growth in model size, primarily measured by the number of parameters, but also the subsequent escalation in computational demands, hardware and software prerequisites for training, all culminating in a substantial financial investment as well. In this paper, we present novel techniques like supervision, parallelization, and scoring functions to get better results out of chains of smaller language models, rather than relying solely on scaling up model size. Firstly, we propose an approach to quantify the performance of a Smaller Language Models (SLM) by introducing a corresponding supervisor model that incrementally corrects the encountered errors. Secondly, we propose an approach to utilize two smaller language models (in a network) performing the same task and retrieving the best relevant output from the two, ensuring peak performance for a specific task. Experimental evaluations establish the quantitative accuracy improvements on financial reasoning and arithmetic calculation tasks from utilizing techniques like supervisor models (in a network of model scenario), threshold scoring and parallel processing over a baseline study.

基于人工智能LLM技术的虚拟患者系统构建与临床教学应用

目的构建一种基于人工智能大语言模型(large language model,LLM)技术、可用于医学教育的新型虚拟患者(virtual patient,VP)系统,评价该系统在基层医生进修学习全科医学临床思维中的应用效果。方法选取2021年1月至2024年2月在东南大学附属中大医院进修的基层社区医生为研究对象,随机分为试验组和对照组,分别采用基于LLM的VP系统教学、传统教学方法进行授课,通过临床思维理论知识考核、临床思维能力考核、课程满意度调查评估教学效果,并对结果进行相应的统计学分析。结果共纳入124名基层社区医生,其中试验组60例、对照组64例,两组在一般基线资料上差异无统计学意义,具有可比性。课程结束后,试验组临床思维理论知识考核成绩显著高于对照组(83.83±3.15 vs.79.92±4.52,P<0.01),且不及格率显著低于对照组(0.00%vs.9.38%,P<0.05);试验组在临床思维能力3个维度(批判性、系统性、循证思维)方面教学后分数均显著高于教学前,而对照组仅在批判性思维维度上教学前后差异有统计学意义;教学后试验组在系统思维、循证思维方面分数均显著高于对照组(P<0.05),但在批判性思维上两组分数差异无统计学意义。试验组对授课的总体满意度也显著高于对照组(93.33%vs.85.48%,P<0.05)。结论基于LLM的VP系统提升了学员对临床思维理论知识的掌握程度,也促进了其临床思维能力的培养,该教学方法可为其他医学教育群体提供新的教学工具和思路。

基于LLM的多粒度口令分析研究

基于口令的认证是常见的身份认证机制。然而,大规模口令泄露事件时有发生,表明口令仍面临着被猜测或者盗用等风险。由于口令可以被视作一种特殊的自然语言,近年来运用自然语言处理技术进行口令分析的研究工作逐渐展开。目前少有工作在大语言模型(LLM,large language model)上探究口令文本分词粒度对口令分析效果的影响。为此,提出了基于LLM的多粒度口令分析框架,总体上沿用预训练范式,在大量未标记数据集上自主学习口令分布先验知识。该框架由同步网络、主干网络、尾部网络3个模块构成。其中,同步网络模块实现了char-level、template-level和chunk-level这3种粒度的口令分词,并提取了口令的字符分布、结构、词块组成等特征知识;主干网络模块构建了通用的口令模型来学习口令组成规律;尾部网络模块生成了候选口令对目标库进行猜测分析。在Tianya、Twitter等8个口令库上进行大量实验,分析总结了多粒度分词下所提框架在不同语言环境中的口令分析效果。实验结果表明,在中文用户场景中,基于char-level和chunk-level分词的框架口令分析性能接近一致,且显著优于基于template-level分词的框架;在英文用户场景中,基于chunk-level分词的框架口令分析性能最佳。

基于大语言模型(LLM)的宝石知识图谱的构建

宝石学知识的来源有书籍、期刊、课程、市场等,宝石知识点繁多且在存储上处于相对孤立的状态,不利于从业者和研究者检索知识,图谱能够处理知识点之间的复杂关联,是常用的结构化数据无法实现的,构建图谱形式的宝石知识库系统可以方便学习和检索。本文介绍了传统的知识图谱构建方法并指出了其中的难点(成本高、工作量大、技术难、容易出错),提出了使用大语言模型(LLM)来完成知识图谱构建中的一些任务来改善成本和工作量的问题;构思了一种基于LLM的知识图谱构建思路(步骤包括数据清洗、知识获取和知识精炼),构建了一个能够覆盖本科阶段宝石知识的宝石知识图谱,对一些查询场景做了展示,经过内部测试评估证明了新方法的可行性和高效率,并展望了该图谱未来可能的应用方向。

GUARDIAN: A Multi-Tiered Defense Architecture for Thwarting Prompt Injection Attacks on LLMs

This paper introduces a novel multi-tiered defense architecture to protect language models from adversarial prompt attacks. We construct adversarial prompts using strategies like role emulation and manipulative assistance to simulate real threats. We introduce a comprehensive, multi-tiered defense framework named GUARDIAN (Guardrails for Upholding Ethics in Language Models) comprising a system prompt filter, pre-processing filter leveraging a toxic classifier and ethical prompt generator, and pre-display filter using the model itself for output screening. Extensive testing on Meta’s Llama-2 model demonstrates the capability to block 100% of attack prompts. The approach also auto-suggests safer prompt alternatives, thereby bolstering language model security. Quantitatively evaluated defense layers and an ethical substitution mechanism represent key innovations to counter sophisticated attacks. The integrated methodology not only fortifies smaller LLMs against emerging cyber threats but also guides the broader application of LLMs in a secure and ethical manner.

基于ChatGLM的应用程序日志分析系统

在以大语言模型为代表的人工智能技术高速发展下,自然语言处理技术发展日新月异,被广泛应用于生产生活。然而,应用程序日志随着数据量的提升变得越发复杂,难以实现高效地分析与处理。对此,文章提出了一种基于大语言模型ChatGLM的分析处理方法,同时设计了一个日志分析系统。该系统借助ChatGLM具有的高效文本建模能力,采用微服务架构设计及容器部署,达到了高效、省时、省力的目标,为大规模日志分析提供了一种有益的思路。

The impact of ChatGPT on foreign language teaching and learning: Opportunities in education and research

The revolutionary online application ChatGPT has brought immense concerns to the education field.Foreign language teachers being some of those most reliant on writing assessments were among the most anxious,exacerbated by the extensive media coverage about the much-fantasized functionality of the chatbot.Hence,the article starts by elucidating the mechanisms,functions and common misconceptions about ChatGPT.Issues and risks associated with its usage are discussed,followed by an in-depth discussion of how the chatbot can be harnessed by learners and teachers.It is argued that ChatGPT offers major opportunities for teachers and education institutes to improve second/foreign language teaching and assessments,which similarly provided researchers with an array of research opportunities,especially towards a more personalized learning experience.

大型语言模型对高等教育的影响与中国应对

大型语言模型(LLM)在短期内将从创设教学情境、创新教学模式、教学AI开发和应用、拓展高校在线学习、助力虚拟仿真课程开发等方面助力高等教育;长期来看能够助推“四新”建设、赋能基础学科拔尖人才培养,并催生高校结构性变革。同时,LLM也可能引发教育安全风险,造成新的教育智能鸿沟,并被大企业寡头垄断。为促进LLM赋能中国高等教育,需要我国加快开发迭代自己的LLM,促进其均等化可达,提升师生数字素养和能力,夯实数字化教学环境。

发展新质生产力 推动我国经济高质量发展

中华人民共和国(新中国)成立以来,从毛泽东的《论十大关系》,到邓小平的“科学技术是第一生产力”,再到习近平的“整合科技创新资源,引领发展战略性新兴产业和未来产业,加快形成新质生产力”,我党对经济工作规律性的认识,随着时代的发展而不断深化。习近平总书记在2024年1月31日召开的中央政治局第十一次集体学习会议上的重要讲话,更是把这种认识推向了全新的高度。总书记在主持学习时明确指出“必须牢记高质量发展是新时代的硬道理”,“高质量发展需要新的生产力理论来指导,而新质生产力已经在实践中形成并展示出对高质量发展的强劲推动力、支撑力,需要我们从理论上进行总结、概括,用以指导新的发展实践”,并强调“科技创新能够催生新产业、新模式、新动能,是发展新质生产力的核心要素”。为了深入学习贯彻总书记讲话精神,围绕“发展新质生产力推动我国经济高质量发展”这个新时代经济发展的核心课题,本刊邀请国内著名专家、学者,撰写一组笔谈文章,以飨读者。

Object Detection Meets LLMs: Model Fusion for Safety and Security

This paper proposes a novel model fusion approach to enhance predictive capabilities of vision and language models by strategically integrating object detection and large language models. We have named this multimodal integration approach as VOLTRON (Vision Object Linguistic Translation for Responsive Observation and Narration). VOLTRON is aimed at improving responses for self-driving vehicles in detecting small objects crossing roads and identifying merged or narrower lanes. The models are fused using a single layer to provide LLaMA2 (Large Language Model Meta AI) with object detection probabilities from YoloV8-n (You Only Look Once) translated into sentences. Experiments using specialized datasets showed accuracy improvements up to 88.16%. We provide a comprehensive exploration of the theoretical aspects that inform our model fusion approach, detailing the fundamental principles upon which it is built. Moreover, we elucidate the intricacies of the methodologies employed for merging these two disparate models, shedding light on the techniques and strategies used.

是“神马”还是“灰犀牛”:ChatGPT等大语言模型对教育的多维影响及应对之策

ChatGPT以“入侵”的方式在教育领域登陆,初露“灰犀牛”面貌,由此引发教育思想、教育体系和学校教育层面的危机。在教育思想层面,多将ChatGPT视为“神马”,在态度上选择抵制和观望,其背后既有思维惯性、惰性等历史原因,也有因ChatGPT等大语言模型管理制度缺位导致安全感无处寄放的现实依据。就教育体系而言,ChatGPT引发的教育自洽与替代焦虑加持,将从内外两个角度解构既有的教育目标体系,由此带来基于人的自由全面发展的教育目标体系的重构;ChatGPT将推动知识生产与知识学习的转向,进而以知识教育价值重估为“支点”,撬动教育整体价值重估,促使教育立足“人本”和“高阶”开展价值创造;ChatGPT还将引发学生发展与评价标准、方式的变革,渐次推动教育评价体系的全面创新。就学校教育而言,知识学习的变革必将推动学校教育时空的重组和学校生态的创新,使课堂教学由“三维”向“四维”转型,进而推动教学生态重塑、流程再造和课堂教学革命,引发教师角色、工作方式和发展方式的变革。

大语言模型在中学历史学科中的应用测评分析

大语言模型一经发布便获得广泛关注,但其在实际应用特别是教育领域的应用还存在诸多局限与挑战,因此需要对大语言模型在中文语境下的能力与风险进行测评。基于此,文章首先收集整理了一个包括10万条客观选择题与10套中学主观题测试卷的中学历史数据集,并在以ChatGPT、GPT-4和讯飞星火为代表的大语言模型上测试了该数据集中题目的回答表现。然后,文章详细分析了测试结果,发现虽然当前大语言模型的突出能力在于能够产生完整且流畅的表达,但其在中学历史知识测试中仍远低于适龄学生的平均水平,大语言模型应用于教育领域仍存在可靠性较差、可信度较低、具有偏见与歧视、推理能力不足、无法自动更新知识等问题。最后,文章针对大语言模型在中文语境下教育领域的应用提出建议,以期助力大语言模型在教育领域发挥更大的作用,为学生、教师带来更好的学习和教学体验。

LLM在工业品物料分类场景的应用

为了解决工业品分类中传统技术所面临的数据不足、泛化性能不佳和准确度不高的问题,提出了一种结合大语言模型与深度学习文本分类的新型工业品物料分类模型。通过利用大语言模型进行数据增强和分类结果的校准,优化了原有的技术框架。经过验证,与原模型相比,新模型在精准率、召回率和准确率方面均实现了显著提升。

大语言模型在英语教学中的角色

本文提炼了大语言模型在英语教学应用中扮演的三种角色,即语言顾问、语伴和语言测评专家。在语言顾问角色中,模型为师生提供语言知识,充当母语者或语言学家。在语伴角色中,模型协助用户完成语言交际任务,可以作为听说练习中的对话伙伴,也可以是读写练习中的小组讨论成员。在语言测评专家角色中,模型分析用户提供的语言材料,并对相关语言表现进行评价。本文主要展示了如何利用提示工程在听、说、读、写、译教学中发挥大语言模型的三类角色作用。