TL;DR
This paper reviews the state-of-the-art in continuous authentication (CA) and user profiling (UP) using sensor data from devices like smartphones. It highlights the growing importance of these techniques for securing personal information and emphasizes the critical need for privacy-preserving methods to protect sensitive biometric data, a gap that remains largely unaddressed in current research.
What Problem Does This Paper Address?
In today's digital world, our devices store vast amounts of sensitive personal information, making them prime targets for cyberattacks. Traditional authentication methods (passwords, PINs) are often weak and only applied at the start of a session, leaving data vulnerable if the device is compromised later. Continuous Authentication (CA) aims to solve this by constantly verifying the user's identity, while User Profiling (UP) can categorize users for initial access control. However, the biometric data used for these methods, collected from various sensors, is highly sensitive. If this data is compromised, especially when processed by external servers, it can lead to identity theft and other malicious activities. This paper addresses the challenge of developing effective CA and UP systems that also rigorously protect user privacy.
Methods
The paper provides a comprehensive review of existing research, categorizing methods based on the type of sensor data used. It covers:
- Device Interaction Sensors: Data like battery consumption, network activity, and ambient light.
- Motion Sensors: Accelerometer, gyroscope, and magnetometer data to track user movement and activities.
- Touch Dynamics: Analyzing user interaction with touchscreens (pressure, speed, position).
- Voice Sensors: Using voice commands and characteristics for authentication.
- Facial Recognition: Employing camera data to identify users.
- Ocular Recognition: Utilizing eye-related features (iris, pupil).
- Other Bodyparts Recognition: Including fingerprints, hand geometry, and even odor.
- Physiological Features: Data from wearable devices like ECG, PPG, EEG, reflecting health status.
The review also delves into privacy-preserving techniques, categorizing them into:
- Template Protection: Methods to secure stored biometric templates, including cancelable biometrics (salting, non-invertible transformations) and biometric cryptosystems (key binding, key generation).
- Data Outsourcing: Techniques for securely sharing data with external servers, such as hash functions, symmetric and asymmetric encryption, and homomorphic encryption.
Datasets
The paper lists several publicly available datasets used in CA and UP research. Key behavioral datasets include:
- SherLock: A comprehensive dataset from Samsung Galaxy S5 smartphones, collecting a wide variety of sensor data and simulating malicious activities.
- Device Analyzer: Collected general mobile sensor data from over 30,000 Android users.
- LiveLab Project: Gathered wireless network and behavioral data from 34 iPhones.
- LDCC: Behavioral attributes reflecting user social interaction from 170 users over two years.
- Social Evolution: Mobile behavioral characteristics from 80 students.
- Reality Mining: Behavioral data from 100 Nokia 6600 smartphones.
Biological datasets mentioned include:
- MOBIO: Face and voice data from 150 English-speaking subjects.
- CSIP: Iris and periocular information from 50 individuals.
- FTV: Face, teeth, and voice data from 50 people.
- MobBIO: Face, iris, and voice data from 105 volunteers.
- UMDAA: Face and various behavioral sensor data from 48 volunteers.
- MobiBits: Voice, face, iris, hand image, and handwritten signature data.
- SWAN Multimodal Biometric Dataset: Face, periocular, and voice data from 150 subjects.
Key Insights and Ideas
- Sensor Fusion is Key: Combining data from multiple sensors (e.g., motion and touch dynamics) significantly improves authentication accuracy and robustness.
- Passive CA is the Goal: The trend is towards "passive" or "implicit" CA, where user verification happens seamlessly in the background without user intervention, enhancing usability.
- Privacy is a Major Hurdle: While CA and UP offer enhanced security, the sensitive nature of biometric data collected poses significant privacy risks, especially when processed remotely.
- UP for Initial Access: User Profiling, while not direct authentication, can serve as a valuable first-level filter by categorizing users into groups.
- Limited Privacy-Preserving UP Research: The review highlights a notable lack of research specifically addressing privacy-preserving techniques for User Profiling using sensor data.
- Wearables and Medical Devices are Next Frontiers: Physiological features from wearables and implantable medical devices are emerging as promising, yet largely unexplored, sources for CA.
Key Results
The paper is a review, so it doesn't present new experimental results. However, it synthesizes findings from numerous studies, indicating:
- High Accuracy Achieved: Many CA systems report high accuracy rates, often exceeding 95% for various sensor modalities. For example, some motion-based CA achieved accuracies of 99.81%, and touch dynamics systems reported Equal Error Rates (EER) as low as 0-4%.
- Privacy-Preserving Methods Show Promise: Studies on privacy-preserving CA demonstrate that techniques like cancelable biometrics and format-preserving encryption can maintain high accuracy (e.g., only a 5.18% reduction in accuracy with FPE) while significantly enhancing security.
- Challenges Remain: Despite high accuracies, issues like distinguishing between real users and sophisticated attacks (e.g., presentation attacks for face recognition) and handling behavioral variations due to user state or device usage are ongoing challenges.
Limitations and Future Work
- Privacy Gap in UP: The authors explicitly state that no investigations focused on User Profiling and privacy-preserving based on sensor use were found, indicating a significant research gap.
- Limited Privacy in CA: While privacy-preserving techniques are being applied to CA, many existing works focus on simple distance-based methods rather than advanced AI. The authors suggest exploring more sensors and combining them with AI tools.
- Resource Constraints: Current smartphones have limitations that make complex CA protocols challenging to implement solely on-device. Outsourcing to external servers is common but introduces its own privacy risks.
- Unexplored Sensors: The paper suggests exploring more sensors for privacy-preserving CA, particularly biological features like facial attributes and physiological signals from wearables and implantable medical devices.
- Need for Standardized Benchmarks: The review points out the lack of detailed information on authentication frequency and resource consumption in many studies, hindering direct comparison and practical implementation.
- Future Directions: The authors advocate for the use of advanced cryptographic protocols like Shamir's Secret Sharing (SSS), Secure Multi-party Computation, and Zero-Knowledge Proofs (ZKPs) to enhance privacy in outsourced CA systems. They also emphasize the potential of combining machine learning with these cryptographic techniques.
Summary updated