跳到正文
arXiv cs.AI· Gabriela Cioboata, Teodor-Stefan Bildea·· 4 小时前AI 评分34

研究揭示 FinBERT 金融情感分析存在语义不敏感漏洞:ASI 攻击成功率 90.85%

Evaluating Semantic Insensitivity in Financial Sentiment Analysis: Implications for Sustainable Financial Decision-Making

AI 导读

研究者提出 Agent-driven Semantic Inversion(ASI)对抗攻击框架,利用 LLM 多智能体流水线对金融新闻句子做定向词汇替换,在 FinBERT 上取得 90.85% 的攻击成功率。

正文 · 原文

View PDF

Abstract:Financial sentiment analysis is critical for automated trading, portfolio management, and risk assessment, heavily relying on models like FinBERT. Despite high benchmark accuracy, these models remain vulnerable to subtle adversarial text modifications, threatening the integrity of algorithmic decision-making. In this paper, we propose Agent-driven Semantic Inversion (ASI), a novel adversarial attack framework targeting financial sentiment models through domain-specific semantic manipulation. Utilizing an LLM-powered multi-agent pipeline, our approach identifies eligible financial news sentences and applies targeted lexical substitutions. We empirically evaluate ASI against FinBERT, one of the most widely deployed financial sentiment models. Our results reveal a striking vulnerability: the attack successfully deceived FinBERT in 90.85% of eligible cases, with the model frequently maintaining its original sentiment prediction despite the fundamental inversion of financial performance indicators. Through qualitative analysis and token-level investigation, we demonstrate that FinBERT critically over-relies on contextual keywords and syntactic patterns rather than true logical comprehension of financial semantics. Our findings highlight the urgent need for logic-aware, robust financial NLP models to ensure the reliability of automated decision-making systems.
Comments: 17 pages, 1 figure, Presented and accepted for publication at the Symposium "Rethinking Economic Models: Sustainability, Finance, and Accounting in Times of Transformation" (Bucharest, 2026)
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Cite as: arXiv:2611.21653 [cs.CL]
  (or arXiv:2611.21653v1 [cs.CL] for this version)
  https://doi.org/10.48550/arXiv.2611.21653

arXiv-issued DOI via DataCite (pending registration)

Submission history

From: Gabriela Cioboata [view email]
[v1] Sun, 18 Oct 2026 16:43:13 UTC (427 KB)

来源:arXiv cs.AI · arxiv.org