Single architecture and multiple task deep neural network for altered fingerprint analysis

Oliver Giudice,Mattia Litrico,Sebastiano Battiato

DOI: https://doi.org/10.48550/arXiv.2007.04931

2020-07-10

Abstract:Fingerprints are one of the most copious evidence in a crime scene and, for this reason, they are frequently used by law enforcement for identification of individuals. But fingerprints can be altered. "Altered fingerprints", refers to intentionally damage of the friction ridge pattern and they are often used by smart criminals in hope to evade law enforcement. We use a deep neural network approach training an Inception-v3 architecture. This paper proposes a method for detection of altered fingerprints, identification of types of alterations and recognition of gender, hand and fingers. We also produce activation maps that show which part of a fingerprint the neural network has focused on, in order to detect where alterations are positioned. The proposed approach achieves an accuracy of 98.21%, 98.46%, 92.52%, 97.53% and 92,18% for the classification of fakeness, alterations, gender, hand and fingers, respectively on the <a class="link-external link-http" href="http://SO.CO.FING" rel="external noopener nofollow">this http URL</a>. dataset.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

This paper aims to solve the problem of fingerprint tampering detection. Specifically, the paper proposes a method based on deep neural networks, using the Inception - v3 architecture to detect tampered fingerprints, identify the types of tampering, and further identify gender, hand, and specific fingers. The paper mentions that fingerprint tampering is a method often used by criminals. They deliberately destroy the friction ridge pattern of fingerprints to evade legal tracking. This tampering behavior has a serious impact on the performance of Automatic Fingerprint Identification Systems (AFIS), especially when dealing with low - quality, tampered or incomplete fingerprint images.

The main contribution of the paper is the development of a single - architecture/multi - task solution that can simultaneously complete five classification tasks:
1. Distinguish between real and tampered fingerprints;
2. Detect the type of tampering (including blurring, central rotation and Z - shaped cutting);
3. Identify which finger the fingerprint comes from;
4. Determine whether the fingerprint comes from the left or right hand;
5. Distinguish whether the fingerprint belongs to a male or a female.

This method was trained and tested on the SO.CO.FING dataset. The results show that its accuracies in the above five tasks are 98.21%, 98.46%, 92.52%, 97.53% and 92.18% respectively, all exceeding the levels of existing technologies. In addition, by generating activation maps, this method can also point out which parts of the fingerprint are focused on by the neural network, thus helping to locate the specific position of tampering. This achievement is of great significance for improving the security and reliability of fingerprint recognition.