Abstract
Biometric systems are increasingly deployed for security, yet they remain vulnerable to spoofing attacks. Evaluating these vulnerabilities typically requires the physical fabrication of fake traits, a process that is both costly and difficult to scale. This paper introduces a novel statistical framework for simulating "fake" match score distributions by interpolating between known impostor and genuine distributions using a tunable "attack strength" parameter. We validate this model against real-world face and fingerprint spoofing data, demonstrating that it provides a realistic and flexible tool for system designers to assess security degradation without exhaustive physical testing.
Introduction
The rise of biometric authentication in consumer and enterprise sectors has been accompanied by the threat of "spoof" or "direct" attacks. In these scenarios, an attacker presents a synthetic or stolen trait—such as a high-resolution photograph or a silicone fingerprint mold—to the sensor. While liveness detection mechanisms exist, they often introduce a trade-off by increasing the False Rejection Rate (FRR) for legitimate users. To optimize system design, researchers require robust methods to test how systems behave under attack. However, the traditional approach of assuming a "worst-case" scenario—where a spoof perfectly mimics a genuine user—is often overly pessimistic and fails to capture the nuances of varying attack qualities.
The Simulation Model
We propose a mathematical model to simulate the score distribution produced by spoof attacks, $p(s|F)$. The core of the approach is a linear interpolation between impostor ($s_I$) and genuine ($s_G$) scores:
$$s_F = (1 - \alpha)s_I + \alpha s_G$$
where $\alpha \in [0, 1]$ represents the "Attack Strength." An $\alpha$ of 0 corresponds to a completely ineffective attack (indistinguishable from a random impostor), while an $\alpha$ of 1 represents a perfect forgery. This parameter allows designers to simulate a spectrum of attack effectiveness influenced by forgery technique, attacker skill, or sensor sensitivity.
Experimental Methodology
To validate the model, we collected a dataset from 40 individuals. Face data included genuine samples and "photo attacks" (high-resolution images displayed on a laptop). Fingerprint data included genuine samples and "fake" samples created using liquid silicon and plasticine molds. We utilized Neurotechnology’s VeriFinger 6.0 for fingerprints and an Elastic Bunch Graph Matching (EBGM) algorithm for face recognition.
Results and Validation
The model was validated by comparing simulated distributions to empirical spoof data using the L1-norm Hellinger distance. For face spoofing, the model achieved an excellent fit at $\alpha = 0.9144$ (Hellinger distance = 0.0939). For fingerprints, the fit was less precise ($\alpha = 0.0522$, Hellinger distance = 0.4397) due to the high skewness of the empirical scores. However, the model tended to overestimate the False Acceptance Rate (FAR) in the fingerprint case, which serves as a beneficial "safety margin" for security designers.
Discussion and Conclusion
Our results indicate that real-world spoofs rarely achieve "perfect" genuine scores. The proposed model offers a more accurate prediction of system performance than the traditional worst-case assumption, particularly for low-quality spoofs. By treating the fake distribution as an intermediate state, researchers can quantify the "distance" a distribution shifts from the impostor baseline. Future work will focus on expanding this validation to larger datasets and diverse biometric traits, such as iris and voice, to further refine the linear interpolation model.
Summary updated