In this paper, we introduce a novel class of fast, beam search-based adversarial attack (BEAST) for Language Models (LMs). BEAST employs interpretable parameters, enabling attackers to balance between attack speed, su...
详细信息
In this paper, we introduce a novel class of fast, beam search-based adversarial attack (BEAST) for Language Models (LMs). BEAST employs interpretable parameters, enabling attackers to balance between attack speed, success rate, and the readability of adversarial prompts. The computational efficiency of BEAST facilitates us to investigate its applications on LMs for jailbreaking, eliciting hallucinations, and privacy attacks. Our gradient-free targeted attack can jailbreak aligned LMs with high attack success rates within one minute. For instance, BEAST can jailbreak Vicuna-7B-v1.5 under one minute with a success rate of 89% when compared to a gradient-based baseline that takes over an hour to achieve 70% success rate using a single Nvidia RTX A6000 48GB GPU. BEAST can also generate adversarial suffixes for successful jailbreaks that can transfer to unseen prompts and unseen models such as GPT-4-Turbo. Additionally, we discover a unique outcome wherein our untargeted attack induces hallucinations in LM chatbots. Through human evaluations, we find that our untargeted attack causes Vicuna-7B-v1.5 to produce ∼15% more incorrect outputs when compared to LM outputs in the absence of our attack. We also learn that 22% of the time, BEAST causes Vicuna to generate outputs that are not relevant to the original prompt. Further, we use BEAST to generate adversarial prompts in a few seconds that can boost the performance of existing membership inference attacks for LMs. We believe that our fast attack, BEAST, has the potential to accelerate research in LM security and privacy. Copyright 2024 by the author(s)
Human activity recognition (HAR) from sensory data is a crucial task for a wide variety of applications. The in-built inertial sensor facilities of commercial smartphones have made the data collection process easier. ...
详细信息
The drug traceability model is used for ensuring drug quality and its safety for customers in the medical supply chain. The healthcare supply chain is a complex network, which is susceptible to failures and leakage of...
详细信息
The dialects of a language hold a significant place in speech processing (SP) applications. The objective of dialect identification is to categorize speech sample data into a specific dialect of a speaker's spoken...
详细信息
The complexity, intricacy and volatility of vast financial systems across the world make quantitative research of equities and cryptocurrencies a very challenging undertaking. Because of this, our research project use...
详细信息
This study focuses on enhancing Natural Language Processing (NLP) in generative AI chatbots through the utilization of advanced pre-trained models. We assessed five distinct Large Language Models (LLMs): TRANSFORMER M...
详细信息
Blockchain is a magnificent technology that could decentralize data and information management, distribution, and storage. With the implementation of blockchain technology, Current corporate operations should be impro...
详细信息
Both unit and integration testing are incredibly crucial for almost any software application because each of them operates a distinct process to examine the *** to resource constraints,when software is subjected to mo...
详细信息
Both unit and integration testing are incredibly crucial for almost any software application because each of them operates a distinct process to examine the *** to resource constraints,when software is subjected to modifications,the drastic increase in the count of test cases forces the testers to opt for a test optimization *** such strategy is test case prioritization(TCP).Existing works have propounded various methodologies that re-order the system-level test cases intending to boost either the fault detection capabilities or the coverage efficacy at the ***,singularity in objective functions and the lack of dissimilitude among the re-ordered test sequences have degraded the cogency of their *** such gaps and scenarios when the meteoric and continuous updations in the software make the intensive unit and integration testing process more fragile,this study has introduced a memetics-inspired methodology for *** proposed structure is first embedded with diverse parameters,and then traditional steps of the shuffled-frog-leaping approach(SFLA)are followed to prioritize the test cases at unit and integration *** 5 standard test functions,a comparative analysis is conducted between the established algorithms and the proposed approach,where the latter enhances the coverage rate and fault detection of re-ordered test *** results related to the mean average percentage of fault detection(APFD)confirmed that the proposed approach exceeds the memetic,basic multi-walk,PSO,and optimized multi-walk by 21.7%,13.99%,12.24%,and 11.51%,respectively.
The early detection of oral malignancy by physicians is a strenuous task. The analysis of histopathological oral malignancy images using image processing and deep learning techniques can be an add-on facility for doct...
详细信息
Image inpainting consists of filling holes or missing parts of an image. Inpainting face images with symmetric characteristics is more challenging than inpainting a natural scene. None of the powerful existing models ...
详细信息
暂无评论