Fusion Methods for Face Presentation Attack Detection

Abstract

Face presentation attacks (PA) are a serious threat to face recognition (FR) applications. These attacks are easy to execute and difficult to detect. An attack can be carried out simply by presenting a video, photo, or mask to the camera. The literature shows that both modern, pre-trained, deep learning-based methods, and traditional hand-crafted, feature-engineered methods have been effective in detecting PAs. However, the question remains as to whether features learned in existing, deep neural networks sufficiently encompass traditional, low-level features in order to achieve optimal performance on PA detection tasks. In this paper, we present a simple feature-fusion method that integrates features extracted by using pre-trained, deep learning models with more traditional colour and texture features. Extensive experiments clearly show the benefit of enriching the feature space to improve detection rates by using three common public datasets, namely CASIA, Replay Attack, and SiW. This work opens future research to improve face presentation attack detection by exploring new characterizing features and fusion strategies.

Enhanced Face Presentation Attack Prevention Employing Feature Fusion of Pre-trained Deep Convolutional Neural Network Model and Thepade's Sorted Block Truncation Coding

“…It is difficult to compare performance because each existing approach uses a different testing environment, performance measurements, and datasets. The HTER and testing accuracy noticed in the suggested work provided here using VGG19; however, are better than the regarded existing equivalent attempts from the literature [20,23,24,29] when one compares the explorations carried out utilising datasets NUAA and Replay-Attack (as presented in Table 9 for comparison of existing methods with the proposed approach).…”

“…The fusion of features formed by using pre-trained DCNN models and conventional content features is presented by Abdullakutty et al [29]. Colour LBP was extracted in YCbCr and HSV colour spaces for content features.…”

Unmasking the Imposters: Task-specific feature learning for face presentation attack detection

“…Hybrid models, which combine hand-crafted features with deep features, have demonstrated improved PA detection recently [6], [21]. Authors of [22] proposed a hybrid method that utilized Discriminant Correlation Analysis (DCA), Canonical Correlation Analysis (CCA) and intensity distribution control using image contrast adjustment along with transfer learning and HOG features.…”

“…Fusion models using pre-trained classification models and colour texture features [6] exhibited improved intra-dataset detection performance compared to transfer learning models. Hence, fusion models were formed using fine-tuned ResNet-50 features and hand-crafted features including colour texture (CLBP), Difference of Gaussian (DoG), Histogram of Oriented Gradients (HOG) and Fast Fourier Transform (FFT).…”

“…So, it may be more effective to use a model that can extract task-specific features from RGB datasets rather than using multi-modal data. The other research [6] has shown that fusion models using deep pre-trained models and hand-crafted methods improved PA detection in intra-dataset evaluations. Thus transfer learning has been explored extensively in face anti-spoofing.…”