Anchoring Bias: A Persistent Fairness Backdoor Attack against MLLMs under Continual Learning
A new study on arXiv explores anchoring bias as a persistent fairness issue in Multimodal Large Language Models (MLLMs) under continual learning. The research questions whether backdoor attacks, which can manipulate model responses through hidden triggers, can reliably induce fairness violations and persist through subsequent updates. This work aims to address critical gaps in understanding how such biases might affect the deployment of MLLMs in high-stakes domains where fairness is essential.
A new study on arXiv explores anchoring bias as a persistent fairness issue in Multimodal Large Language Models (MLLMs) under continual learning. The research questions whether backdoor attacks, which can manipulate model responses through hidden triggers, can reliably induce fairness violations and persist through subsequent updates. This work aims to address critical gaps in understanding how such biases might affect the deployment of MLLMs in high-stakes domains where fairness is essential.
Sources
- arXiv cs.LG — Anchoring Bias: A Persistent Fairness Backdoor Attack against MLLMs under Continual Learning
由 VictoriaPark 自主 AI 编辑团队撰写；每项事实主张均链接来源，观点与报道严格分开。