Two-Dimensional Dynamic Fusion for Continuous Authentication

arXiv:2309.04128 · cs.CR · Submitted 2023-09-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: We're looking at the paper: "Two-Dimensional Dynamic Fusion for Continuous Authentication".

Elias: Here's the gist of the paper. Continuous authentication has been widely studied to provide high security and usability for mobile devices by continuously monitoring and authenticating users <ref:2309.04128#pg2>.

Nadia: In this paper, we propose a new approach to multibiometric continuous authentication: two-dimensional dynamic fusion <ref:2309.04128#pg2>.

Elias: Our key insight is that multibiometric continuous authentication calculates two-dimensional matching scores over classifiers and over time <ref:2309.04128#pg2>.

Nadia: Based on this, we dynamically select a set of classifiers based on the context in which authentication is taking place, and fuse matching scores by multi-classifier fusion and multi-sample fusion <ref:2309.04128#pg2>.

Paper summary: Nadia: So we're looking at this paper called Two-Dimensional Dynamic Fusion for Continuous Authentication, and the core idea they're pushing here is moving away from constantly running every single biometric classifier all the time.

Elias: Exactly. Their main thesis is that you should stop calculating all those scores continuously and instead focus on calculating two dimensional matching scores over classifiers and over time instead <ref:2309.04128#pg2>.

Priya: But what does that mean practically for a user who just wants to use their phone seamlessly? Like, if the input quality drops because the user is in a dark room or moving around, how does this help them stay logged in?

Nadia: Well, they propose that the system dynamically selects which classifiers to calculate based on the current context. It's about minimizing resource usage by only activating a set of classifiers that they predict will have a high success probability and low cost <ref:2309.04128#pg5>.

Elias: It sounds like they're using the environment—the context—as a filter to decide which biometric checks are necessary right now, instead of running everything simultaneously <ref:2309.04128#pg3>.

Priya: If you only run a few classifiers, are you missing some important failure modes where the input might be weird? I'm worried that cutting things out dynamically could introduce new kinds of errors <ref:2309.04128#pg6>.

Nadia: They address that by using a classifier scheduling algorithm. This algorithm dynamically activates classifiers based on probability estimates for successful authentication, which is the true acceptance rate for each classifier in that specific context <ref:2309.04128#pg5>.

Elias: And they select the candidate set with the lowest resource cost if their combined a priori probability of successful authentication stays above some threshold, which helps manage those trade-offs <ref:2309.04128#pg5>.

Priya: So, when you look at the results, are we talking about less accuracy overall because they're cutting things out? I need to know what that means for trust in the security of the authentication process <ref:2309.04128#pg7>.

Nadia: The experimental evaluation shows they achieve almost the same accuracy as existing methods with fewer score calculations, but they also show higher accuracy when using the same number of score calculations by adopting multi-sample fusion <ref:2309.04128#pg9>.

Elias: That multi-sample fusion part is interesting because it seems to balance things out, giving a better trade-off between resource usage and accuracy compared to the existing fusion methods <ref:2309.04128#pg7>.

Priya: So, if I'm just listening to this show, what does this mean for my daily life? Am I getting a better lock on my phone without having to wait longer or risk being locked out?

Nadia: It means that because they are using multi-sample fusion, they're actually getting higher accuracy than other methods even when you keep the number of score calculations the same <ref:2309.04128#pg9>.

Elias: The paper is titled Two-Dimensional Dynamic Fusion for Continuous Authentication, and it’s about restructuring how we handle continuous authentication on smartphones by making it context aware <ref:2309.04128#pg1>.

Nadia: It suggests that the future of this kind of authentication isn't just about having more biometrics, but about being smart enough to know which biometrics matter at any given moment <ref:2309.04128#pg2>.

Priya: So, where does this leave us regarding the actual security? If you're running fewer calculations, are you losing some kind of security guarantee that we need for continuous monitoring <ref:2309.04128#pg6>?

Elias: The system is built around an authentication window called a time window, delta t window, which is the period a user must be successfully authenticated to keep using the device <ref:2309.04128#pg6>.

Nadia: They derive the overall confidence score by first fusing individual classifier scores using a sample fusion rule and then deriving an overall score with a classifier fusion rule <ref:2309.04128#pg6>.

Priya: That sounds like they are trying to manage uncertainty by combining different ways of scoring, which is good for data quality, but I need to know what their limitations are on the real world <ref:two thousand three hundred nine point zero four one two eight#pg6.

Elias: One limitation they mention is that the system relies on an authentication window, meaning if you slip out of that window without re-authenticating, the continuous monitoring stops working for you <ref:two thousand three hundred nine point zero four one two eight#pg6.

Nadia: That's a practical constraint—you have to be within that time frame—but the main thing is they found a way to get better accuracy while cutting down on the calculations needed for that continuous monitoring <ref:two thousand three hundred nine point zero four one two eight#pg7.

Priya: So, in short, it’s about using context to smartly choose your authentication checks and then fusing those results intelligently so you don't compromise security or usability too much <ref:two thousand three hundred nine point zero four one two eight#pg9.

Elias: That's the essence of Two-Dimensional Dynamic Fusion for Continuous Authentication, showing that by being dynamic with your classifiers and fusion, you can get a better balance between what you need for security and what the device can handle <ref:two thousand three hundred nine point zero four one two eight#pg1.

Conclusion: Nadia: So we’ve been walking through how this paper tries to fix continuous authentication by using dynamic fusion, and now we need to wrap up what all this means for the people who actually care about security.

Elias: We have to look at the title itself, "Two-Dimensional Dynamic Fusion for Continuous Authentication," because that’s the core mechanism they’re proposing. It sounds a bit heavy, but it points directly at how they handle multiple biometrics simultaneously over time.

Nadia: Yeah, and what it really means is that instead of running every single check constantly, you only run the checks that actually make sense for your current situation. It’s about making the system smart enough to know which biometric input is most likely to succeed right now.

Priya: So you’re saying the system doesn't just keep a static list of things to check; it changes its whole strategy based on where you are and what’s happening around you? That sounds like it could really improve usability because it cuts down on unnecessary processing.

Elias: Exactly. They do this by calculating two dimensional matching scores, which means they look at the different biometric classifiers and also look at how that matching score changes over time. It’s a two dimensional picture of authentication success.

Nadia: And the dynamic part is that they use your context—like whether you're in a noisy place or a dark room—to decide which specific checks to activate. That minimizes resource usage while still keeping things accurate, which is what we always want to see in applied security.

Priya: I’m still thinking about the data side, though. They claim better accuracy with the same number of score calculations by using something called multi-sample fusion and it looks pretty promising on paper. But what does that mean for real-world privacy? Is this dynamic selection process leaking more info somewhere else?

Elias: That’s a fair question about the trade-off. They do mention how they fuse those individual classifier scores, first using a sample fusion rule and then combining them with a classifier fusion rule to get an overall confidence score. It’s a specific way they manage that data flow.

Nadia: The main thing is that their experimental results show this approach gives you a better balance between keeping things accurate and not bogging down your device with calculations, especially when running at high frequencies. It’s about practical performance under stress.

Priya: So the conclusion is that they’ve found a way to be more efficient without sacrificing the accuracy we need for continuous monitoring, but I still want to know what the authors themselves say about where this can actually go next.

Elias: They conclude that by dynamically activating classifiers based on context, they achieve an optimized mix of multi classifier and multi sample fusion. It’s a practical solution for building more robust continuous authentication systems.

Nadia: So it boils down to using the environment to guide the security process, which is a smarter way to design these kinds of mobile defenses than just running everything at maximum power all the time.

Nuttapong Attrapadung, Goichiro Hanaoka, Haochen M. Kotoi-Xie, Takahiro Matsuda, Takumi Moriyama, Takao Murakami, Hidenori Nakamura, Jacob C. N. Schuldt, Masaaki Tokuyama, Jing Zhang

National Institute of Advanced Industrial Science and Technology (AIST) · AnchorZ Inc. · Institute of Statistical Mathematics (ISM)

cs.CR

Submitted: 2023-09-08

Updated: 2023-09-08

Comments: Accepted to IJCB'23

Journal ref: 2023 IEEE International Joint Conference on Biometrics (IJCB)

DOI: 10.1109/IJCB57857.2023.10448866

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 90/100

The gist: Continuous authentication has been widely studied to provide high security and usability for mobile devices by continuously monitoring and authenticating users <ref:2309.04128#pg2>.

Key concepts

Two-Dimensional Dynamic Fusion
This is the core idea of viewing continuous authentication as combining two types of scores: matching scores calculated across different biometric classifiers and matching scores calculated over time. The system dynamically chooses which classifiers to use based on the environment (context) to minimize resource use while maximizing success probability.
Classifier Scheduling Algorithm
This algorithm is responsible for deciding which biometric classifiers should be activated at any given moment. It estimates the true acceptance rate for each classifier in a specific context and selects a set of classifiers that meets a minimum success threshold while having the lowest combined resource cost.
Multi-Sample Fusion
This technique is used to combine individual classifier scores into an overall confidence score. The system uses the average of scores derived from both sample fusion rules and classifier fusion rules, which simplifies the process and helps achieve higher accuracy than traditional methods with the same number of calculations.

Terminology

Summary

Continuous authentication has been widely studied to provide high security and usability for mobile devices by continuously monitoring and authenticating users <ref:2309. p. 004128#pg2>. In this paper, we propose a new approach to multibiometric continuous authentication: two-dimensional dynamic fusion <ref:2309. p. 004128#pg2>. Our key insight is that multibiometric continuous authentication calculates two-dimensional matching scores over classifiers and over time <ref:2309. p. 004128#pg2>. Based on this, we dynamically select a set of classifiers based on the context in which authentication is taking place, and fuse matching scores by multi-classifier fusion and multi-sample fusion <ref:2309. p. 004128#pg2>. Through experimental evaluation, we show that our approach provides a better balance between resource usage and accuracy than the existing fusion methods <ref:2309. p. 004128#pg2>. In particular, we show that our approach provides higher accuracy than the existing methods with the same number of score calculations by adopting multi-sample fusion <ref:2309. p. 004128#pg2>.

Pruning redundant reasoning tokens

The core idea of two-dimensional dynamic fusion is viewing multibiometric continuous authentication as a fusion of two-dimensional matching scores over classifiers and over time <ref:2309.04128#pg2>. This framework involves dynamically determining which scores to calculate in the two-dimensional space based on the context (i. p. 1e., the environment) <ref:2309. p. 004128#pg2>. The context is used because it is a good predictor of the classifiers most likely to successfully authenticate the user, such as using facial recognition in a noisy environment <ref:2309. p. 104128#pg2>. This contextual selection minimizes resource usage by only activating a set of classifiers that has a sufficiently high combined expected success probability and the lowest recourse cost <ref:2309. p. 104128#pg2>.

How the training signal is built

The system operates through a classifier scheduling algorithm and a corresponding fusion algorithm that implement context-aware dynamic multibiometric fusion <ref:2309.04128#pg2>. The classifier scheduling algorithm dynamically activates classifiers based on probability estimates of each classifier successful authentication (true acceptance rate) for each classifier in a given context <ref:2309. p. 304128#pg5>. A set S is considered a valid candidate set if the combined a priori probability of successful authentication is above the given threshold thp, and then the algorithm selects the candidate set with the combined lowest resource cost as the set of classifiers to be activated <ref:2309. p. 304128#pg5>.

Where the gains come from

The fusion algorithm is based on an authentication window ∆twindow within which a user must be successfully authenticated to be allowed continued use of the mobile device <ref:2309.04128#pg6>. The overall confidence score β is derived by first fusing individual classifier scores via a sample fusion rule and then deriving an overall score via a classifier fusion rule <ref:2309. p. 404128#pg6>. The fusion algorithm uses the average of the available scores for both sample and classifier fusion rules, which simplifies the process <ref:2309. p. 404128#pg6>.

Experimental Evaluation

Experimental evaluation demonstrates that our approach can achieve almost the same accuracy as existing methods with fewer score calculations and higher accuracy with the same number of score calculations <ref:2309.04128#pg7>. Specifically, in a multi-modal experiment, our approach achieves almost the same accuracy as existing approaches with fewer score calculations <ref:2309. p. 604128#pg9>. The results show a clear advantage when running at a high frequency (3x) in all considered contexts <ref:2309. p. 604128#pg9>. This illustrates the better accuracy and performance trade-off offered by our approach with multi-sample fusion <ref:2309. p. 704128#pg9>.

Conclusion

In this paper, we have proposed a new approach to multi-biometric continuous authentication based on twodimensional dynamic fusion. In contrast to prior work, our approach dynamically activates classifiers based on context, which enables an optimized multi-classifier and multisample fusion. p. 7 Our fusion approach achieves a better trade-off between accuracy and the number of score calculations compared to existing approaches. p. 7

References

[1] EnterpriseAppsToday, Digital Banking Statistics.

[2] Fox News, Americans check their smartphones 96 times a day, survey says <ref:2309. p. 804128#pg2>.

[3] Lookout, Phone theft in America – Breaking down the phone theft epidemic <ref:2309. p. 804128#pg2>.

[4] The Extended M2VTS (XM2VTS) Database <ref:2309. p. 504128#pg8>.

[5] S. Bharadwaj, H. p. 8 S. Bhatt, R. Singh, M. Vatsa, and A. p. 8 Noore <ref:2309. p. 004128#pg6>.

[6] D. Crouse, H. Han, D. Chandra, B. p. 8 Barbello, and A. p. 8 K. Jain <ref:2309. p. 004128#pg2>.

[7] A. Jain, K. Nandakumar, and A Ross <ref:2309. p. 004128#pg6>.

[8] G Kalantzis, G Papakostas, T Karanikiotis, M Papamichail, and A L Symeonidis <ref:2309. p. 804128#pg8>.

[9] R Kumar, V V Phoha, and A Serwadda <ref:2309. p. 004128#pg6>.

[10] S Rasnayaka and T Sim <ref:2309. p. 004128#pg2>.

[11] A Riener <ref:2309. p. 004128#pg4>.

[12] A Ross, K Nandakumar, and A K Jain <ref:2309. p. 004128#pg6>.

[13] D Ruta and B Gabrys <ref:2309. p. 004128#pg6>.

[14] D Sivasankaran, M Ragab, T Sim, and Y Zick <ref:2309. p. 804128#pg8>.

[15] M Smith-Creasey and M Rajarajan <ref:2309. p. 204128#pg6>.

[16] SRI International. p. 8 The Speakers in the Wild (SITW) Speaker Recognition Challenge <ref:2309. p. 604128#pg8>.

[17] M Vatsa, R Singh, A Noore, and A Ross <ref:2309. p. 004128#pg6>.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Context-aware classifier scheduling will dynamically select a set of classifiers based on probability estimates of successful authentication in a given context to maximize accuracy with limited resource usage. This enables the system to dynamically determine which scores to calculate in the two-dimensional space based on the context (i.e., the environment).

  2. The system can implement an optimized combination of multi-classifier and multi-sample fusion by selecting the appropriate fusion type implicitly through classifier scheduling. For example, in a noisy environment, the fusion of two face samples (e.g., multi-sample) could provide higher accuracy than the fusion of face and voice samples obtained at the same time (e.g., multi-classifier).

  3. The continuous authentication loop can adapt its security policy based on environmental conditions by deriving a context-dependent authentication window. This allows for flexible security, such as letting the window length depend on whether the device is in a low-risk environment such as a company officer or personal home, or in a high-risk environment such as a public place.

  4. The system can achieve higher accuracy with the same number of score calculations by adopting multi-sample fusion when some classifiers provide poor performance. This demonstrates that our approach provides high accuracy by adopting multi-sample fusion in a context where some classifiers provide poor performance.

Related papers