TL;DR
This paper introduces MQID (Multiple-user Quickest Intruder Detection), a framework designed to protect mobile devices shared among multiple legitimate users (like a family or a medical team). By extending Quickest Change Detection (QCD) mathematics, the authors developed an algorithm that catches intruders with minimal delay while ensuring that switching between authorized users doesn't trigger a false alarm.
What Problem Does This Paper Address?
Most Active Authentication (AA) research assumes a "one device, one user" model. In reality, we often share our phones or tablets with family members or colleagues. Standard AA systems struggle here: if they are too sensitive, they flag a spouse or teammate as an intruder; if they are too relaxed, a real thief can access data for a long time before being caught.
The challenge is to balance three competing needs:
- Accuracy: Distinguishing between multiple "enrolled" users and unknown intruders.
- Latency: Detecting an intruder as quickly as possible (low "Average Detection Delay").
- Usability: Avoiding "false detections" when the device is handed from one legitimate user to another.
Methods
The authors propose two main algorithms based on the Quickest Change Detection (QCD) framework:
- MQID (Multiple-user Quickest Intruder Detection): 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." 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.
- DEMQID (Data-Efficient MQID): This version 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.
Datasets
The researchers evaluated their method using two prominent face-based AA datasets:
- UMDAA01: Contains 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 their experiments, they simulated scenarios with 1 to 7 enrolled users and used the remaining subjects as "intruders" to test the system's limits.
Key Insights and Ideas
- The "Minimum" Strategy: The core innovation is using the minimum likelihood ratio across the set of enrolled users. This mathematically ensures that the system is only "surprised" (and thus increases its intrusion score) when the current user doesn't match the profile of any authorized person.
- Robustness through Sparsity: Surprisingly, the data-efficient version (DEMQID) sometimes performed better than the full version when the data was noisy. By sampling more sparsely, the algorithm effectively "filtered out" momentary noise or bad frames that might have otherwise caused a false alarm.
- Scalability: The system scales gracefully. Adding more authorized users only slightly increases the time it takes to catch an intruder, proving that multi-user AA is a computationally and practically viable security model.
Key Results
- Superior Latency: MQID consistently outperformed baseline methods (like Single Score-based Authentication) at low False Detection rates.
- Minimal "Handover" Cost: On the UMDAA01 dataset, the system only required 2.35 additional samples on average to maintain a 2% false detection rate when the number of authorized users increased from 1 to 7.
- Challenging Environments: 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 real-world conditions.
- Efficiency: The DEMQID algorithm achieved these results while only looking at roughly 30.4% of the available data frames.
Limitations and Future Work
- Modality Constraints: The study focused exclusively on the face modality. While the math should apply to other sensors (like gait or touch patterns), those were not tested here.
- Simple Classifiers: The authors used a relatively simple template-matching classifier (nearest neighbor with ResNet18 features) to demonstrate the framework's effectiveness. Integrating more sophisticated, state-of-the-art classifiers could further improve performance.
- User Capacity: While the paper tested up to 7 users (typical for a family), it is unclear how the latency would scale for much larger groups, such as an entire corporate department sharing a pool of devices.
Summary updated