Detecting sensor level spoof attacks using joint encoding of temporal and spatial features

Automated detection of sensor level spoof attacks is critical for success of surveillance technologies and the detection of crimes. This paper presents a new approach to more accurately detect such spoof face presentation attempts during the surveillance. The image data from the spoof samples not only illustrates the subtle texture difference in the spatial domain but is also accompanied by temporal differences as compared to those from the live/real human samples. Therefore, we investigate a new approach to encode the sparsity of these two different categories while combining these two cues from the acquired image sequences. The spatial and temporal information in the acquired image data and the sparse dictionary approach are used to encode such information in effectively separating the spoof and the live categories. This approach does not require large number of training samples, such as for the deep learning based methods, while achieving very good performance. The experimental results presented in this paper on publicly available database achieve outperforming results and illustrate the effectiveness of the proposed approach.