Abstract
In this paper, we investigate how to detect intruders with low latency for Active Authentication (AA) systems with multiple-users. We extend the Quickest Change Detection (QCD) framework to the multiple-user case and formulate the Multiple-user Quickest Intruder Detection (MQID) algorithm. Furthermore, we extend the algorithm to the data-efficient scenario where intruder detection is carried out with fewer observation samples. We evaluate the effectiveness of the proposed method on two publicly available AA datasets on the face modality.
Introduction
Most Active Authentication (AA) research assumes a "one device, one user" model. In reality, mobile devices are often shared among family members or colleagues. Standard AA systems struggle in these scenarios: if they are too sensitive, they flag a legitimate co-user as an intruder; if they are too relaxed, a real thief can access data for a long period before being caught.
The challenge is to balance three competing needs: accuracy in distinguishing between multiple enrolled users and unknown intruders, low latency in detecting an intruder (Average Detection Delay), and high usability by avoiding false detections when the device is handed from one legitimate user to another.
Methodology: The MQID Framework
The authors propose two main algorithms based on the Quickest Change Detection (QCD) framework. The core innovation is the MQID (Multiple-user Quickest Intruder Detection) algorithm.
During enrollment, the system builds a classifier and two probability distributions (matched and non-matched scores) for every authorized user. During operation, the system calculates a "likelihood ratio" for the current observation against all enrolled users. It specifically uses the minimum likelihood ratio among all users to update its internal "intrusion score." This strategy ensures that the system is only "surprised" (increasing its intrusion score) when the current user doesn't match the profile of any authorized person. If the observation looks like any of the authorized users, the score stays low. If it looks like no one, the score climbs until it hits a threshold and declares an intrusion.
A second variant, DEMQID (Data-Efficient MQID), reduces the number of samples processed to save battery and computation. It uses a "control rule" to decide whether to skip the next observation based on the current confidence level.
Experimental Evaluation
The researchers evaluated their method using two prominent face-based AA datasets:
- UMDAA01: Images from 50 individuals captured via an iPhone 5S across three sessions with varying lighting.
- UMDAA02: A more challenging, unconstrained multimodal dataset with 44 subjects recorded over two months using a Nexus 5.
In experiments simulating 1 to 7 enrolled users, the system demonstrated high scalability. Adding more authorized users only slightly increased the time it takes to catch an intruder.
Key Results and Performance
MQID consistently outperformed baseline methods at low False Detection rates. On the UMDAA01 dataset, the system required only 2.35 additional samples on average to maintain a 2% false detection rate when the number of authorized users increased from 1 to 7.
Key insights from the results include:
- Robustness through Sparsity: Surprisingly, the data-efficient DEMQID sometimes performed better than the full version when data was noisy. By sampling more sparsely, the algorithm effectively filtered out momentary noise or bad frames.
- Efficiency: DEMQID achieved competitive results while only looking at roughly 30.4% of the available data frames.
- Real-world Robustness: On the more difficult UMDAA02 dataset, the system required about 4.33 more samples to maintain the same security level as users were added, showing it remains robust even in unconstrained conditions.
Conclusion
The MQID framework demonstrates that multi-user active authentication is both computationally and practically viable. By utilizing a "minimum" likelihood strategy, the system ensures security without sacrificing usability during user handovers. Future work may involve testing other modalities like gait or touch patterns and integrating more advanced classifiers.
Summary updated