Kırılgan damgalama ve yapay zeka kullanan çok modlu biyometrik şablonlarda kurcalama tespiti

Significant attention has been brought towards biometric authentication in recent years, according to the rapidly growing demand for secure and reliable authentication methods. To improve the accuracy of biometric authentication systems, multiple templates are collected from each candidate, for different biometrics. Accordingly, any false positives or false negatives that may occur when matching one of the templates are rectified when matching the second one, and vice versa. Moreover, to reduce the overhead in the size of the data required to represent these templates, as well as secure these templates, recent studies use digital watermarking techniques to embed one of the templates, or its features, into the other. Fragile watermarking is widely used for this purpose, according to the ability of detecting any tampering when the watermark information is lost. However, this fragility denies the ability to further reduce the size of the data by compressing the image. Thus, a new watermarking approach is proposed in this study, in which the features extracted from the iris template are embedded in the face template of the candidate. The features of the iris are extracted using an autoencoder Artificial Neural Network (ANN), then, embedded in the DCT coefficients of the face image, after being quantized and before being compressed using Huffman Coding, based on the Joint Photographic Experts Group (JPEG) standard. The watermark information is embedded using the Least-Significant-Bit (LSB) watermarking method, where each value is calculated based on the corresponding value in the iris template and the DCT coefficient in the face template, and encrypted using Arnold Transform (AT). Accordingly, the watermark information is lost when the compressed watermarked image is attacked. Additionally, as the watermark information does contain similar patterns, but never identical even for the same candidate, an ANN is used to investigate the existence of these patterns in the extracted watermark, instead of comparing it to a static array. The proposed method has achieved 100% tamper detection rate with 0.05% reduction in iris recognition accuracy while maintaining the face recognition accuracy intact.