SLAPBench Benchmarks MLLMs for Fingerprint Verification

Four-finger SLAP fingerprints are critical for identity verification in sensitive applications like border control and law enforcement. Despite the rise of multimodal large language models (MLLMs), their capability for this specific type of biometric verification had not been systematically assessed. To address this gap, researchers introduced SLAPBench, the first benchmark designed to evaluate MLLMs on four-finger SLAP fingerprint verification tasks. The benchmark, built using NIST SD302b data, tested several open-source MLLMs (InternVL3-8B, Qwen2.5-VL-7B, Qwen3-VL-8B, Gemma-3-12B) and the proprietary Claude Opus 4.8. A key finding was the profound impact of prompting strategies on MLLM behavior. Simple "task-description" prompts often led open-source models to collapse, resulting in near-100% False Accept Rates. However, "similarity-scoring" prompts mitigated this collapse, revealing significant capability disparities among the models. Claude Opus 4.8 demonstrated the best performance, resisting collapse and achieving high discrimination (AUC = 0.953). Among open-source models, Gemma-3-12B showed reasonable performance, while others struggled or even inverted their discrimination. The study also highlighted potential issues with the benchmark data, such as near-duplicate detection, and suggested that fairness disparities might increase as discrimination weakens. SLAPBench establishes a crucial baseline and underscores that effective prompting is paramount for MLLM performance in biometric verification.