Implicit Sensor-based Authentication of Smartphone Users with Smartwatch

Wei-Han Lee, Ruby B. LeeHasp@isca 2016 · DOI

TL;DR

This paper introduces iAuth, a novel implicit authentication system that continuously verifies smartphone users by analyzing sensor data from both their smartphone and smartwatch. By leveraging machine learning on time and frequency domain sensor features, iAuth achieves high accuracy (92.1%) with minimal overhead, offering a more convenient and secure alternative to traditional authentication methods.

What Problem Does This Paper Address?

Smartphones have become central to accessing sensitive cloud-based services and storing personal data. While initial login provides a security layer, users often remain logged in, leaving their devices vulnerable if stolen or accessed by an unauthorized individual. Traditional re-authentication methods (like passwords or biometrics) are inconvenient for continuous use and can be bypassed by attackers who have already gained access. This paper addresses the critical need for an automatic, non-intrusive, and continuous authentication system that can detect and prevent unauthorized access to smartphones and their associated cloud services without requiring explicit user interaction.

Methods

iAuth employs a two-pronged approach to implicit authentication:

  • Sensor Fusion: It combines sensor data from both the smartphone (accelerometer, gyroscope) and a wearable device like a smartwatch. This fusion is crucial because different sensors, even of the same type but on different body parts, capture distinct behavioral patterns.
  • Time and Frequency Domain Analysis: Raw sensor data is segmented into time windows. Within each window, features are extracted in both the time domain (mean, variance, max, min) and the frequency domain (peak amplitudes and frequencies). This dual-domain analysis captures a richer representation of user behavior.
  • Machine Learning Classification: A lightweight machine learning algorithm, Kernel Ridge Regression (KRR), is used to build and deploy the authentication model. KRR is chosen for its computational efficiency, making it suitable for on-device processing.
  • Continuous Authentication: The system operates as a background service, continuously monitoring sensor data and performing authentication checks. This ensures that if a device is compromised, the attacker is quickly detected.
  • Cloud-based Training: While authentication happens on the smartphone, the training of the authentication models is performed on a trusted cloud server. This allows for the use of a larger dataset, including data from other enrolled users (anonymized), to build more robust models.

Datasets

The study utilized data collected from 20 users over one to two weeks. Participants used their own Nexus 5 smartphones and Moto 360 smartwatches in their daily lives. Sensor data (accelerometer and gyroscope) was collected at a sampling rate of 50 Hz. The paper does not mention the creation of a new public benchmark dataset.

Key Insights and Ideas

  • Smartwatch as a Security Enhancer: The integration of smartwatch sensor data significantly improves authentication accuracy compared to using the smartphone alone. This highlights the value of multi-device sensing for security.
  • Time-Frequency Features are Key: Combining time and frequency domain features from sensor data provides a more discriminative representation of user behavior, leading to higher authentication accuracy.
  • Implicit Authentication is Feasible and Practical: The system demonstrates that continuous, implicit authentication is achievable with minimal impact on user experience and device resources.
  • Lightweight ML for On-Device Security: The use of KRR showcases how efficient machine learning models can be deployed on resource-constrained mobile devices for real-time security applications.
  • Behavioral Drift Detection: The system is designed to adapt to behavioral drift over time, with the potential for online retraining.

Key Results

  • High Authentication Accuracy: iAuth achieved an authentication accuracy of 92.1% when combining smartphone and smartwatch data. Using only the smartphone, the accuracy was 83.2%.
  • Low False Acceptance Rate (FAR): The system demonstrated a FAR of 7.5% when using both devices, indicating a low probability of an attacker being misidentified as a legitimate user.
  • Fast Authentication: The authentication process takes less than 13 milliseconds per check, and with a 6-second window size, the implicit authentication is effectively performed every 6 seconds.
  • Minimal System Overhead: CPU utilization was around 4-6% on average, and memory utilization was approximately 3 MB.
  • Low Battery Consumption: iAuth consumes an additional 1.8% of battery power per hour when the phone is locked and running the service, and 2% per hour when the phone is in use.
  • Effective Against Masquerading Attacks: iAuth detected 90% of attackers attempting masquerading attacks within an average of 18 seconds, and all adversaries within 24 seconds.

Limitations and Future Work

  • User Enrollment Time: The paper mentions that about 800 measurements are needed for initial training, which might be a considerable amount of data collection for some users.
  • Privacy of Training Data: While the paper states that user data is anonymized for cloud-based training, the exact anonymization techniques and their robustness are not detailed.
  • Dependence on Trusted Cloud Server: The training module relies on a trusted cloud server, which could be a single point of failure or a target for attacks.
  • Future Work: The authors suggest exploring other wearable devices and sensors, as well as investigating more sophisticated machine learning techniques for even higher accuracy and robustness. They also implicitly suggest the need for more robust anonymization techniques for training data.

Summary updated

BibTeX
@inproceedings{lee2016implicit,
  author = {Wei-Han Lee and Ruby B. Lee},
  title = {{Implicit Sensor-based Authentication of Smartphone Users with Smartwatch}},
  booktitle = {Hasp@isca},
  year = {2016},
  doi = {10.1145/2948618.2948627},
}
Figure