Leveraging Cross-Attribute Heterogeneity and Joint Training to Detect Ever-Evolving and Era-Diverse Social Media Bots

ACL ARR 2025 February Submission2837 Authors

Abstract: Social media bots detection is a crucial task in maintaining the health of the Internet. The challenge of this task is that bots are evolving themselves by constantly stealing information from human accounts, a behavior also known as feature camouflage, to evade detection. To reduce the impact of camouflage, existing methods detect by using intra-attribute heterogeneity. However, our work reveals that intra-attribute heterogeneity is being diluted by the further stealing behavior of bots, hindering the development of these methods. To address this, we propose a novel concept called cross-attribute heterogeneity. Compared to intra-attribute heterogeneity, it is less susceptible to camouflage. Based on this superior nature, we design a framework called BCH to better detect more advanced bots through cross-attribute heterogeneity. Additionally, to enhance compatibility with bots from different eras, BCH incorporates a joint training strategy. Extensive experiments shows the superiority of BCH in detecting ever-evolving and era-diverse bots, as well as detailed analysis highlights the benefits of cross-attribute heterogeneity and the necessity of improving detection methods compatibility.

Paper Type: Long

Research Area: Computational Social Science and Cultural Analytics

Research Area Keywords: NLP tools for social analysis, quantitative analyses of news and/or social media

Contribution Types: NLP engineering experiment, Data analysis

Languages Studied: English

Submission Number: 2837