PersAD: adversarial protection against text-based personality inference

• Published in: Data mining and knowledge discovery Vol. 39; no. 5; p. 61
• Main Authors: Qiu, Houjie, Ma, Xingkong, Liu, Bo, Cai, Yiqing, Peng, Baoyun, Huang, Dan
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.09.2025; Springer Nature B.V
• Subjects: Artificial Intelligence; Chemistry and Earth Sciences; Cognition & reasoning; Computer Science; Data Mining and Knowledge Discovery; Design; False information; Inference; Information Storage and Retrieval; Knowledge; Machine learning; Methods; Neural networks; Personality; Personality traits; Physics; Privacy; Psychology; Semantics; Social networks; Statistics for Engineering; Personality privacy protection; Large language model; Textual adversarial example; Text-based personality inference
• ISSN: 1384-5810; 1573-756X
• DOI: 10.1007/s10618-025-01144-0

As text-based personality inference approaches increasingly mature, personality privacy invasion has become a new threat applied to human-central malicious activities in recent years. When facing text-based personality inference attacks, the adversarial example method is a protection pathway that modifies the original texts to convert the personality characteristics. However, most studies focus on the local word or phrase perturbations instead of the global document semantics, making it challenging to ensure content semantic consistency while effectively reversing personality signals. Simultaneously, most methods are difficult to defend against various personality inference models without psychological knowledge. To address these challenges, we propose an adversarial protection method, PersAD for short. In our method, we extract the knowledge from personality psychology books to construct knowledge-enhanced prompts. Then, the forward-backward adversarial generation is implemented to convert original posts to protected posts with high semantic similarity and low personality relevancy via the LLM. Moreover, the protect success rate (PSR) and cumulative flipping rate (CFR) are proposed to validate the PersAD and compare the performance with other adversarial example methods. The extensive experiments demonstrate the protection performance of PersAD over various text-based personality inference models. Under the protection of PersAD, the average PSR improves in a range of 41.75% to 45.91%, exceeding 3 times the performance of traditional methods. For the CFR of two or more personality dimensions, PersAD achieves results between 84.26 and 87.50%, which is over 40% higher than other baseline methods. Our data & code are available at the Github repository ( https://github.com/Fr4n13Z/PersAD ).