Normalizing flow-based latent space mapping for implicit pattern authentication on mobile devices

• Published in: Applied soft computing 2025-01, Vol.169, p.112469, Article 112469
• Main Authors: Heo, Jaehyuk, Kim, Jeongseob, Chung, Euisuk, Kim, Subin, Kang, Pilsung, Shin, Donghwa, Shin, Jinho, Han, Daehee
• Format: Article
• Language: English
• Subjects: Anomaly detection; Latent space; Mobile user authentication; Normalizing flow; Randomized PIN pad; Touch dynamics
• ISSN: 1568-4946
• DOI: 10.1016/j.asoc.2024.112469

In numerous mobile device applications, safeguarding user privacy through authentication is of paramount importance. Traditional methods employ information-matching-based systems, wherein users authenticate themselves using preset passwords. However, such systems pose a challenge in maintaining privacy if the password is compromised. This study introduces a behavior-based authentication approach that capitalizes on implicit patterns generated by individual users’ device usage habits during password entry on mobile devices, complementing information-based authentication. To optimize this feature, we develop a user-specific latent space mapping framework utilizing a normalizing flow and train a one-class classification machine learning model based on the latent space to enhance authentication performance. Upon applying the proposed methodology to data from 1,000 mobile banking service users, we observed a significant improvement in authentication performance. The integrated error decreased from 21.42% to 9.48%, and the equal error rate reduced from 27.84% to 16.3% compared to the method solely employing a machine learning model in the data space without implementing latent space mapping. •A user authentication method using the normalizing flow-based latent space mapping is proposed.•Our proposed framework enables user authentication by learning an anomaly detection algorithm using only a few valid user data points.•Our proposed framework yields a significant improvement in authentication performance on random PIN pad.•The normalizing flow-based latent space mapping can maximize the difference in characteristics between a valid user and imposters.