Abstract
Identity tracing is a foundational technology in social security, using collected attributes to discover an object's true identity. However, traditional single-attribute biometric systems often struggle with low data quality and a lack of interpretability. This paper proposes the Trustworthy Identity Tracing (TIT) task and a framework based on multi-attribute synergistic identification. By modeling identity through information theory—specifically identity entropy—the framework provides an interpretable path to identification by optimizing the "core identification set" and reducing search space uncertainty.
Introduction: Beyond the Biometric Black Box
In forensic science and public security, identifying a suspect is rarely as straightforward as a high-resolution face match. Real-world data is often messy, incomplete, or low-quality. Traditional deep learning models for biometrics often act as "black boxes," providing an identity without explaining the underlying logic.
Researchers from Wuhan and Xidian Universities propose a shift toward Trustworthy Identity Tracing (TIT). This framework mimics human investigative intuition by layering multiple, human-understandable attributes to systematically narrow down a suspect pool.
The TIT Framework: Identity as Entropy
The core of the Trustworthy Identity Tracing Framework (TITF) is built on the mathematical concept of identity entropy. The uncertainty of an identity is defined as $H(U_x) = \log_2 n$. As more attributes (clues) are collected, the entropy drops. The ultimate goal of identity tracing is to drive this entropy value to zero, representing a unique identification.
Key Components of TITF:
- Core Identification Set: The smallest set of attributes required to uniquely identify a person within a specific group.
- Attribute Discriminability: Not all clues are equal. The framework calculates which attributes (e.g., age, occupation, or language) narrow down the suspect pool most efficiently in a given context.
- Search Space Reduction: The system reorders attribute checks to provide a step-by-step "tracing path" that a human can follow and verify.
Experimental Validation
To evaluate the framework, the authors constructed the Identity Tracing Dataset (ITD) using 1990 US Census data. The dataset includes 1,000 unique individuals with 20 labeled identity attributes to simulate real-world tracing scenarios.
Key Results:
- Efficiency Gain: TITF reduced the time required to identify a target by an average of 41.44% compared to random search baselines.
- Robustness: Even with 80% of attribute data missing, the TITF significantly outperformed traditional methods by selecting the most "information-dense" attributes first.
- Human-Machine Synergy: Because the attributes used are semantic (e.g., job, marital status), the system provides a transparent logic chain that law enforcement can trust and verify.
Conclusion
The TIT framework successfully formalizes qualitative investigative intuition into a quantitative mathematical relationship. By treating identity as a convergence of entropy, it offers a robust, interpretable, and efficient alternative to traditional black-box identification systems, particularly in scenarios where data is vague or incomplete.
Paper Citation: Decheng Liu, Jiahao Yu, Ruimin Hu, and Wenbin Feng. "Toward Trustworthy Identity Tracing via Multi-attribute Synergistic Identification." arXiv:2311.02703v1 [cs.IR], 5 Nov 2023.
Summary updated