Security of multimodal biometric systems against spoof attacks

摘要: A biometric system is essentially a pattern recognition being used in ad-versarial environment. Since, like any conventional security exposed to malicious adversaries, who can manipulate data make the ineffective by compromising its integrity. Current theory and de- sign methods of systems do not take into account vulnerability such adversary attacks. Therefore, evaluation classical design an open problem investigate whether they lead secure systems. In order it necessary understand evalu-ate threats thus develop effective countermeasures robust designs, both technical procedural, if necessary. Accordingly, extension of mandatory safeguard reliability adversarial environments. In this thesis, we provide some contributions towards this direction. Among all potential attacks discussed literature, spoof are one main against for identity recognition. Multimodal commonly believed be in-trinsically more than based on single biomet-ric trait, as combine information coming from different traits. However, recent works have question belief shown that multimodal misled attacker (impostor) even spoofing only traits. first detailed review state-of-the-art The scope ofstate-of-the-art results very limited, since were obtained under very restrictive “worst-case” hypothesis, where assumed able fabricate perfect replica trait whose matching score distribu-tion identical genuine Thus, argue validity hypothesis using large set real empirical evidence scenario representa- ixtive attacks: suitability may depend specific algorithm, techniques counterfeit spoofed Then, propose methodology applications, does require fabricating fake traits, allows designer possible qualities traits attackers, exploits impostor samples which col- lected training system. Our evaluates performances simulated attack model distribution takes explicitly degrees quality particular, two models match factors affect particular biometric, sensor, algorithm computation, technique construct biometrics, skills attacker. All these summarized parameter, call “attack strength”. Further, extension our method rank several biometric score fusion rules according their relative robustness This choose most rule prediction. We then present analysis, sets face fingerprints including show proposed good approximation traits’ our method providing adequate estimation security1 also use how evaluate publicly available benchmark data without experimental strongly depends rule, strength eventually evidence, considering fingerprint score fusion capable correct ranking