This page is still being polished. If you have thoughts, please share them via the feedback form.
Data on this page is preliminary and may change. Please do not share or cite these figures publicly.
Changes to the model's learned parameters, architecture, or training process, including modifications to training data that affect what the model learns.
Also in AI System
This section delves into mitigating risks associated with models, encompassing privacy preservation, detoxification and debiasing, mitigation of hallucinations, and defenses against model attacks.
Reasoning
Section describes multiple model mitigations (detoxification, debiasing, hallucination mitigation, privacy preservation) without specifying particular technical approach.
Privacy Preserving
Privacy leakage is a crucial risk of LLMs, since the powerful memorization and association capabilities of LLMs raise the risk of revealing private information within the training data. Researchers are devoted to designing privacypreserving frameworks in LLMs [226], [227], aiming to safeguard sensitive PII from possible disclosure during humanmachine conservation
1.1 ModelDetoxifying and Debiasing
To reduce the toxicity and bias of LLMs, prior efforts mainly focus on enhancing the quality of training data and conducting safety training.
1.1 ModelHallucination Mitigation
Hallucinations, one of the key challenges associated with LLMs, have received extensive studies.
1.2.9 OtherDefending Against Model Attacks
Recognizing the significant threats posed by various model attacks, earlier studies [144], [288] have proposed a variety of countermeasures for conventional deep learning models. Despite the advancements in the scale of parameters and training data seen in LLMs, they still exhibit vulnerabilities similar to their predecessors. Leveraging insights from previous defense strategies applied to earlier language models, it is plausible to employ existing defenses against extraction attacks, inference attacks, poisoning attacks, evasion attacks, and overhead attacks on LLMs.
1.2.4 Security InfrastructureMitigation in Input Modules
Mitigating the threat posed by the input module presents a significant challenge for LLM developers due to the diversity of the harmful inputs and adversarial prompts [209], [210].
1.2.1 Guardrails & FilteringMitigation in Input Modules > Defensive Prompt Design
Directly modifying the input prompts is a viable approach to steer the behavior of the model and foster the generation of responsible outputs. This method integrates contextual information or constraints in the prompts to provide background knowledge and guidelines while generating the output [22].
1.2.1 Guardrails & FilteringMitigation in Input Modules > Malicious Prompt Detection
Different from the methods of designing defensive prompts to preprocess the input, the malicious prompt detection method aims to detect and filter out the harmful prompts through the input safeguard
1.2.1 Guardrails & FilteringMitigation in Toolchain Modules
Existing studies have designed methods to alleviate the security issues of tools in the lifecycle of LLMs. In this section, we summarize the mitigations of those issues according to the categories of tools
2.4 Engineering & DevelopmentMitigation in Toolchain Modules > Defenses for Software Development Tools
Most existing vulnerabilities in programming languages, deep learning frameworks, and pre-processing tools, aim to hijack control flows. Therefore, control-flow integrity (CFI), which ensures that the control flows follow a predefined set of rules, can prevent the exploitation of these vulnerabilities. However, CFI solutions incur high overheads when applied to large-scale software such as LLMs [320], [321]. To tackle this issue, a low-precision version of CFI was proposed to reduce overheads [322]. Hardware optimizations are proposed to improve the efficiency of CFI [323]. In addition, it is critical to analyze and prevent security accidents in the environments of LLMs developing and deploying. We argue that data provenance analysis tools can be leveraged to forensic security issues [324]–[327] and detect attacks against LLM actively [328]–[330]. The key concept of data provenance revolves around the provenance graph, which is constructed based on audit systems. Specifically, the vertices in the graph represent file descriptors, e.g., files, sockets, and devices. Meanwhile, the edges depict the relationships between these file descriptors, such as system calls.
1.2.4 Security InfrastructureMitigation in Toolchain Modules > Defenses for LLM Hardware Systems
For memory attacks, many existing defenses against manipulating DNN inferences via memory corruption are based on error correction [160], [167], whereas incurring high overheads [168]. In contrast, some studies aim to revise DNN architectures, making it hard for attackers to launch memory-based attacks, e.g., Aegis [169]. For network-based attacks, which disrupt the communication between GPU machines, existing traffic detection systems can identify these attacks.
1.2 Non-ModelRisk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems
Cui, Tianyu; Wang, Yanling; Fu, Chuanpu; Xiao, Yong; Li, Sijia; Deng, Xinhao; Liu, Yunpeng; Zhang, Qinglin; Qiu, Ziyi; Li, Peiyang; Tan, Zhixing; Xiong, Junwu; Kong, Xinyu; Wen, Zujie; Xu, Ke; Li, Qi (2024)
Despite their impressive capabilities, large lan- guage models (LLMs) have been observed to generate responses that include inaccurate or fabricated information, a phenomenon com- monly known as “hallucination”. In this work, we propose a simple Induce-then-Contrast De- coding (ICD) strategy to alleviate hallucina- tions. We first construct a factually weak LLM by inducing hallucinations from the original LLMs. Then, we penalize these induced hallu- cinations during decoding to enhance the fac- tuality of the generated content. Concretely, we determine the final next-token predictions by amplifying the predictions from the orig- inal model and downplaying the induced un- truthful predictions via contrastive decoding. Experimental results on both discrimination- based and generation-based hallucination eval- uation benchmarks, such as TruthfulQA and FACTSCORE, demonstrate that our proposed ICD methods can effectively enhance the factu- ality of LLMs across various model sizes and families. For example, when equipped with ICD, Llama2-7B-Chat and Mistral-7B-Instruct achieve performance comparable to ChatGPT and GPT4 on TruthfulQA, respectively.
Build and Use Model
Training, fine-tuning, and integrating the AI model
Developer
Entity that creates, trains, or modifies the AI system
Manage
Prioritising, responding to, and mitigating AI risks
Other
