Evaluating Vulnerabilities and Countermeasures for Face, Voice, and Fingerprint Systems

Abstract

Biometric authentication systems are increasingly targeted by AI-generated spoofing attacks, yet comparative evaluations across modalities using standardized frameworks remain limited. This study evaluates the vulnerability of facial recognition, voice authentication, and fingerprint verification models to both traditional and AI-driven presentation attacks. Utilizing publicly available benchmark datasets—specifically Face Forensics++ (FF++), ASV spoof 2019, and LivDet 2021—we conducted a controlled evaluation comprising 12,000 authentication trials across three open-source biometric architectures (Arc Face, Wav2Vec 2.0, and Source AFIS). Traditional attacks included printed photos, audio replays, and physical replicas, while AI-driven attacks included deep fakes, neural voice clones, and synthetic fingerprints. Performance was measured using Spoofing Success Rate (SSR), False Acceptance Rate (FAR), and Liveness Detection Bypass Rate (LDBR). Results indicate that AI driven attacks significantly outperformed traditional methods. Voice authentication emerged as the most vulnerable modality, exhibiting an 85% SSR for AI-cloned voices (n = 2,550/3,000 attempts). Facial recognition showed a 78% SSR for deep fakes, while fingerprint systems exhibited a 61% SSR for synthetic prints. In contrast, traditional attacks yielded significantly lower success rates (27 47%). These findings demonstrate that current liveness detection mechanisms are insufficient against accessible AI tools, highlighting the urgent need for multimodal authentication and enhanced physiological liveness checks.