Journal of Electronic Science and Technology, Volume. 23, Issue 1, 100301(2025)
On large language models safety, security, and privacy: A survey
Figures & Tables(4)
Fig. 3. Overview of safety, security, and privacy issues and their defense methods.
Table 1. Classification of safety, security, and privacy issues in LLMs.
View tableView in ArticleTable 1. Classification of safety, security, and privacy issues in LLMs.
Category Feature Example Safety Toxicity and bias Toxicity and bias refer to the inappropriate or prejudiced content that may be generated or amplified by LLMs due to biased training data. This can result in harmful or discriminatory outcomes. [ 12 ,13 ]Hallucination Hallucination refers to the generation of text that is nonsensical, irrelevant, or factually incorrect, typically arising from the model’s inability to accurately understand or process the context of the input data. [ 14 ,15 ]Jailbreak Jailbreak attack refers to the exploitation of vulnerabilities to bypass the model’s intended constraints and generate content that violates its operational guidelines or ethical safeguards. [ 16 ,17 ]Security Backdoor & poisoning attacks Backdoor and poisoning attacks refer to the malicious insertion of hidden triggers or corrupted data during the training process, which can cause the model to produce harmful or targeted outputs when prompted with specific inputs. [ 6 ,18 ]Adversarial attack Adversarial attack is a method where carefully crafted inputs are used to deceive the model into making errors or generating unintended outputs, often by exploiting the model’s weaknesses or vulnerabilities. [ 19 ,20 ]Privacy Privacy leakage Privacy leakage occurs when sensitive or personal information is inadvertently disclosed through the model’s responses, due to the model’s exposure to or training on data containing such information. [ 7 ,21 ]Inference attack Inference attacks involve the exploitation of model responses to deduce sensitive information about the training data or the underlying algorithms, potentially compromising privacy or security. [ 22 ,23 ]Extraction attack Extraction attacks are attempts to reverse-engineer or illicitly obtain proprietary information, such as training data or model parameters, by interacting with the model’s outputs. [ 24 ,25 ]
Tools
Get Citation
Copy Citation Text
Ran Zhang, Hong-Wei Li, Xin-Yuan Qian, Wen-Bo Jiang, Han-Xiao Chen. On large language models safety, security, and privacy: A survey[J]. Journal of Electronic Science and Technology, 2025, 23(1): 100301
Paper Information
Category:
Received: Sep. 25, 2024
Accepted: Jan. 9, 2025
Published Online: Apr. 7, 2025
The Author Email: Hong-Wei Li (hongweili@uestc.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20