摘要
Abstract
With the widely used of smartphones,traditional explicit authentication methods(e.g.,passwords,fingerprints)are increasingly vulnerable due to their reliance on active user input,making the non-invasive implicit authentication a critical research focus.A transfer autoencoder-based implicit authentication framework for smartphones was proposed.Users'behavioral features were captured during pattern unlocking(e.g.,accelerometer and gyroscope data)through multimodal sensors,an autoencoder was employed to extract discriminative latent representations,and a transfer learning mechanism was incorporated for rapid model fine-tuning.The results of the experiment indicate that following the pre-training of the scheme on a 10 GB offline dataset,which is generated from the unlocking patterns of 50 users across 4 smartphones,the online authentication process can be executed in a mere 1.3 seconds,achieving an accuracy rate of 99.06%.This perfor-mance is markedly superior to that of traditional machine learning methods,such as support vector machine(SVM),which exhibits an accuracy rate of 89.19%.Furthermore,it surpasses conventional deep learning approaches,including K-nearest neighbor(KNN)with an accuracy of 95.49%,as well as existing state-of-the-art(SOTA)schemes like EspialCog,which achieves an accuracy of 98.76%.Additionally,it is noteworthy that users are required to perform the unlocking be-havior only six times prior to utilization in order to complete the model adaptation,thereby balancing considerations of se-curity with user experience.关键词
隐式身份认证/自编码器/迁移学习/图案解锁/隐私保护Key words
implicit authentication/autoencoder/transfer learning/pattern unlock/privacy protection分类
信息技术与安全科学
