Continuous Authentication against Collusion Attacks

Pin Lyu, Wandong Cai, Yao Wang

TL;DR

This paper introduces a novel "collusion attack" where attackers and victims intentionally mimic each other's gaits to fool smartphone-based authentication systems. To counter this, they propose a multi-cycle defense model that leverages acceleration direction changes, achieving a low Equal Error Rate (EER) and demonstrating robustness against such sophisticated attacks.

What Problem Does This Paper Address?

Gait authentication, a convenient and unobtrusive method for continuous user verification on smartphones, relies on the unique way individuals walk. While effective against casual observation, these systems are vulnerable to more sophisticated attacks. Existing research has explored "zero-effort" attacks (where attackers naturally have similar gaits) and basic imitation attacks. However, this paper identifies a gap in understanding and defending against a more deliberate and coordinated attack: the "collusion attack." In this scenario, an attacker and a legitimate user actively collude to make their gaits as similar as possible, aiming to overwhelm the authentication system's ability to distinguish them. This poses a significant threat to the security and reliability of smartphone-based gait authentication.

Methods

The paper proposes a novel "collusion attack" model. Instead of a single attacker trying to mimic a victim, this model involves training multiple participants (honor guards, chosen for their uniform gait) to adopt a specific, quantified gait specification. This training, conducted over three months, aims to create a highly similar gait across different individuals, making it difficult for the authentication system to differentiate them.

For defense, the authors introduce a "multi-cycle defense model" that focuses on analyzing acceleration direction changes. This approach avoids the need for additional sensors and instead leverages the existing accelerometer data. The defense model preprocesses the data, extracts gait cycles, and then uses a similarity comparison based on the direction of acceleration vectors within these cycles. A key aspect is the use of quaternions for robust coordinate transformations, allowing for more stable analysis of gait data regardless of the phone's orientation. The multi-cycle aspect involves analyzing multiple consecutive gait cycles to improve the system's confidence and reduce false positives/negatives.

Datasets

The paper utilizes the OU-ISIR dataset for baseline performance evaluation of existing methods. For their attack and defense experiments, they collected their own dataset from 20 participants (honor guards) performing both their designed "training style" gait and their "goose step" gait. This custom dataset allowed them to specifically test their collusion attack and defense model under controlled conditions.

Key Insights and Ideas

  • The power of coordinated imitation: The research highlights that deliberate, long-term training can make gaits of different individuals remarkably similar, significantly increasing the false match rate for gait authentication systems.
  • Acceleration direction as a robust feature: Focusing on the direction of acceleration, rather than just its magnitude, proves to be a more resilient feature against sophisticated imitation attacks. This is especially true when combined with quaternion-based coordinate transformations for stability.
  • Multi-cycle analysis for enhanced confidence: Analyzing multiple consecutive gait cycles, rather than a single one, significantly improves the reliability of the authentication system, reducing both false positives and false negatives.
  • Honor guards as a challenging testbed: The choice of honor guards for training highlights the difficulty of achieving highly similar gaits, and the success of their attack model in this context underscores its effectiveness.

Key Results

  • The proposed collusion attack model achieved a false match rate (FMR) of 39% on a goose-step dataset using a DTW-based baseline, significantly outperforming simple imitation attacks.
  • Their defense model achieved an Equal Error Rate (EER) of 7.32% in the defense scenario, which is better than their previous work and comparable to state-of-the-art multi-sensor systems.
  • The multi-cycle defense model showed optimal performance with three cycles, achieving the lowest EER.
  • In their evaluation, the defense model successfully resisted collusion attacks, with no attacker reaching the 50% confidence threshold in best-performing attacker-victim pairs.

Limitations and Future Work

  • Limited dataset size for attack scenarios: The authors acknowledge that the data used in the attack scenario contained only a few topics, suggesting a need for more extensive data collection for attack analysis.
  • Focus on specific gait types: The current work primarily focuses on specific trained gaits. Future work should explore a wider range of gait patterns and features.
  • Understanding "mimicable" vs. "unmimicable" gaits: The authors express interest in determining the boundaries of what constitutes a "mimicable" gait, which could inform future attack and defense strategies.
  • Exploring new features and models: The paper suggests investigating new features and models to further enhance the robustness and efficiency of gait authentication systems.

Summary updated

Abstract

As mobile devices become more and more popular, users gain many conveniences. It has also made smartphone makers install new software and prebuilt hardware on their products, including many kinds of sensors. With improved storage and computing power, users also become accustomed to storing and interacting with personally sensitive information. Due to convenience and efficiency, mobile devices use gait authentication widely. In recent years, protecting the information security of mobile devices has become increasingly important. It has become a hot research area because smartphones are vulnerable to theft or unauthorized access. This paper proposes a novel attack model called a collusion attack. Firstly, we study the imitation attack in the general state and its results and propose and verify the feasibility of our attack. We propose a collusion attack model and train participants with quantified action specifications. The results demonstrate that our attack increases the attacker’s false match rate only using an acceleration sensor in some systems sensor. Furthermore, we propose a multi-cycle defense model based on acceleration direction changes to improve the robustness of smartphone-based gait authentication methods against such attacks. Experimental results show that our defense model can significantly reduce the attacker’s success rate.
BibTeX
@article{lyu2022continuous,
  author = {Pin Lyu and Wandong Cai and Yao Wang},
  title = {{Continuous Authentication against Collusion Attacks}},
  journal = {Sensors},
  year = {2022},
  doi = {10.3390/s22134711},
}
Figure