Performance Analysis of Motion-Sensor Behavior for User Authentication on Smartphones

Chao Shen, Tianwen Yu, Sheng Yuan, Yunpeng Li, Xiaohong Guan

TL;DR

This paper explores using smartphone motion sensor data (accelerometer and gyroscope) to authenticate users during passcode entry. By analyzing subtle motion patterns, they achieve promising authentication accuracy, suggesting it could serve as a valuable second factor for smartphone security.

What Problem Does This Paper Address?

Smartphones are treasure troves of personal information, making secure authentication crucial. While PINs and patterns are common, they are often weak and easily compromised. Fingerprint scanners are better but require dedicated hardware and aren't universally adopted. This paper tackles the challenge of enhancing smartphone security by leveraging the inherent, passive behavior of users when they interact with their devices, specifically during passcode input, without requiring additional hardware.

Methods

The core of this research lies in analyzing the subtle movements of a smartphone as a user inputs their passcode. The authors extract a rich set of features from accelerometer and gyroscope data, categorizing them into "descriptive" (e.g., mean, variance) and "intensive" (e.g., energy, entropy) metrics. These features are then used to train one-class classifiers (One-Class SVM, Neural Network, and k-Nearest Neighbor). The key idea is to build a model of a legitimate user's typical motion behavior during passcode entry and then flag deviations as potential impostor activity. They also employ preprocessing steps like Kalman filtering for gravity removal and wavelet denoising to clean the sensor data.

Datasets

The authors collected a new dataset for this study, comprising 129,621 passcode samples from 48 participants. This dataset was collected across three operational scenarios (hand-hold-input, table-hold-input, and hand-hold-walk) and three different smartphone models (Huawei Mate7, Samsung N7100, and HongMi 1s). The dataset is made publicly available to encourage further research.

Key Insights and Ideas

  • Passive Authentication: The most compelling idea is using motion sensor data as a passive authentication mechanism. Users don't need to do anything extra; their natural interaction with the phone is analyzed.
  • Subtle Differentiators: Even seemingly simple actions like typing a passcode generate unique motion patterns that can distinguish individuals. The "energy" and "entropy" of accelerometer data in the y and z axes were found to be particularly informative.
  • Complementary Security: The authors position this as a complementary authentication method, not a replacement. It's ideal for enhancing existing PIN or pattern security, making it a two-factor authentication system.
  • Usability vs. Security Trade-off: Longer passcodes generally lead to better authentication accuracy but increase the time it takes to unlock the phone. This highlights a crucial trade-off that needs to be managed in practical applications.

Key Results

The proposed approach demonstrates promising results:

  • Overall Performance: The best performance achieved was a False Rejection Rate (FRR) of 6.85% and a False Acceptance Rate (FAR) of 5.01% using the One-Class SVM classifier in the hand-hold-input scenario.
  • Scenario Performance: The "hand-hold-input" scenario yielded the best results, while "hand-hold-walk" scenarios showed slightly worse performance, likely due to increased noise from movement.
  • Passcode Length: Authentication accuracy improves with longer passcodes. A 5-digit passcode achieved a FAR of 5.01% and FRR of 6.85%, while a 6-digit passcode reduced these to 3.92% and 4.97%, respectively, at the cost of increased authentication time.
  • Training Data Size: Authentication accuracy improves with more training data, but performance saturates around 50 samples, suggesting a balance is needed to avoid excessive training time.
  • Screen Size: Larger smartphone screens generally led to better authentication accuracy, possibly because larger input areas allow for more distinct motion patterns.

Limitations and Future Work

  • Not a Standalone Solution: The authors acknowledge that current accuracy levels are not sufficient for standalone authentication and suggest it's best suited as a second factor.
  • Privacy Concerns: While motion sensor data is less sensitive than keystroke dynamics, users should still be aware of the observation and have control over their data.
  • Environmental Factors: The study notes that factors like screen size can influence performance, and more research is needed to understand these interactions fully.
  • Future Directions: The authors suggest exploring more advanced pattern classifiers, improving noise reduction techniques, and developing methods to handle variability in user behavior. They also emphasize the need for secure profile storage and a robust initial user verification process.

Summary updated

BibTeX
@inproceedings{shen2016performance,
  author = {Chao Shen and Tianwen Yu and Sheng Yuan and Yunpeng Li and Xiaohong Guan},
  title = {{Performance Analysis of Motion-Sensor Behavior for User Authentication on Smartphones}},
  booktitle = {Sensors},
  year = {2016},
  url = {https://openreview.net/forum?id=nAAOY36edK},
}
Figure