paper-with-me

Papers

Red Alarm for Pre-trained Models: Universal Vulnerability to Neuron-Level Backdoor Attacks

2021-01-18 · ICML Workshop AML 2021 7 · Zhengyan Zhang, Guangxuan Xiao, Yongwei Li, Tian Lv, Fanchao Qi, Zhiyuan Liu, Yasheng Wang, Xin Jiang, Maosong Sun

Pre-trained models (PTMs) have been widely used in various downstream tasks. The parameters of PTMs are distributed on the Internet and may suffer backdoor attacks. In this work, we demonstrate the universal vulnerability of PTMs, where fine-tuned PTMs can be easily controlled by backdoor attacks in arbitrary downstream tasks. Specifically, attackers can add a simple pre-training task, which restricts the output representations of trigger instances to pre-defined vectors, namely neuron-level backdoor attack (NeuBA). If the backdoor functionality is not eliminated during fine-tuning, the triggers can make the fine-tuned model predict fixed labels by pre-defined vectors. In the experiments of both natural language processing (NLP) and computer vision (CV), we show that NeuBA absolutely controls the predictions for trigger instances without any knowledge of downstream tasks. Finally, we apply several defense methods to NeuBA and find that model pruning is a promising direction to resist NeuBA by excluding backdoored neurons. Our findings sound a red alarm for the wide use of PTMs. Our source code and models are available at \url{https://github.com/thunlp/NeuBA}.

📄 PDF Abstract BibTeX arXiv:2101.06969

Code (1)

thunlp/NeuBA 공식 구현 pytorch

Tasks

Backdoor Attack

Methods 이 논문이 사용한 방법론

Pruning 설명 없음

Similar Papers 제목 키워드 기반

Do Hallucination Neurons Generalize? Evidence from Cross-Domain Transfer in LLMs

2026-03-27 · Snehit Vaddi, Pujith Vaddi arxiv

Recent work identifies a sparse set of "hallucination neurons" (H-neurons), less than 0.1% of feed-forward network neurons, that reliably predict when large language models will hallucinate. These neurons are identified …

Question Answering

Active Subspace of Neural Networks: Structural Analysis and Universal Attacks

2019-10-29 · Chunfeng Cui, Kaiqi Zhang, Talgat Daulbaev, Julia Gusak 외

Active subspace is a model reduction method widely used in the uncertainty quantification community. In this paper, we propose analyzing the internal structure and vulnerability and deep neural networks using active subs…

Adversarial AttackUncertainty Quantification

Universal Neurons in GPT-2: Emergence, Persistence, and Functional Impact

2025-07-28 · Advey Nandan, Cheng-Ting Chou, Amrit Kurakula, Cole Blondin 외 arxiv

We investigate the phenomenon of neuron universality in independently trained GPT-2 Small models, examining these universal neurons-neurons with consistently correlated activations across models-emerge and evolve through…

Exploring the Universal Vulnerability of Prompt-based Learning Paradigm

2022-04-11 · Findings (NAACL) 2022 7 · Lei Xu, Yangyi Chen, Ganqu Cui, Hongcheng Gao 외

Prompt-based learning paradigm bridges the gap between pre-training and fine-tuning, and works effectively under the few-shot setting. However, we find that this learning paradigm inherits the vulnerability from the pre-…

Universal Neurons in GPT2 Language Models

2024-01-22 · Wes Gurnee, Theo Horsley, Zifan Carl Guo, Tara Rezaei Kheirkhah 외

A basic question within the emerging field of mechanistic interpretability is the degree to which neural networks learn the same underlying mechanisms. In other words, are neural mechanisms universal across different mod…