Biblio

A breakthrough in the emerging use of machine learning and deep learning is the concept of autoencoders and GAN (Generative Adversarial Networks), architectures that can generate believable synthetic content called deepfakes. The threat lies when these low-tech doctored images, videos, and audios blur the line between fake and genuine content and are used as weapons to cause damage to an unprecedented degree. This paper presents a survey of the underlying technology of deepfakes and methods proposed for their detection. Based on a detailed study of all the proposed models of detection, this paper presents SSTNet as the best model to date, that uses spatial, temporal, and steganalysis for detection. The threat posed by document and signature forgery, which is yet to be explored by researchers, has also been highlighted in this paper. This paper concludes with the discussion of research directions in this field and the development of more robust techniques to deal with the increasing threats surrounding deepfake technology.

One of the primary motivations for using ultrasound reflectometry for fingerprint imaging is the promise of increased spoof resistance over conventional optical or capacitive sensing approaches due to the ability for ultrasound to determine the elastic impedance of the imaged material. A fake 3D printed plastic finger can therefore be easily distinguished from a real finger. However, ultrasonic sensors are still vulnerable to materials that are similar in impedance to tissue, such as water or rubber. Previously we demonstrated an ultrasonic fingerprint reader operating with 1.3GHz ultrasound based on pulse echo impedance imaging on the backside silicon interface. In this work, we utilize the large bandwidth of these sensors to differentiate between a finger and materials with similar impedances using the frequency response of elastic impedance obtained by transducer excitation with a wideband RF chirp signal. The reflected signal is a strong function of impedance mismatch and absorption [Hoople 2015].