Face anti spoofing techniques

Abstract

In this new age of computer biometrics, facial recognition is becoming the second most widely deployed biometric authentication method in the world in terms of market quota right after fingerprints. It's used in smartphones, payment methods, and biometric authentication. Each day more and more manufacturers are including face recognition in their products, such as Apple with its Face-ID technology, the banks with the implementation of eKYC solutions for the onboarding process. Due to this widespread use of face biometrics , it has become more vulnerable to spoofing attacks. Hackers and adversaries try to bypass face biometric systems using certain methods. For the safety of users and to maintain their confidence in online biometric authentication systems, it is today’s need to develop such systems which protect biometrics and don't let hackers bypass security systems. The methods which come under this are called face anti spoofing methods. These methods prevent false facial verification by using a photo, video, mask or a different substitute for an authorized person’s face. A few advancements have been done in this field, like pixel wise supervision, use of vision transformers etc. Apart from software improvements and deep learning algorithms, specialized hardware is also used to prevent face anti spoofing, like 3d cameras, infrared cameras, 3-D dot projector (used in Apple’s most popular device i.e iPhones), and they have proved to be effective at times. But still there is more work to be done, as with the advancements of technology, hackers and adversaries are getting more advanced, and there are several cases when deep fake has been used to bypass Face security systems, posing great threat to users privacy. In this research project, we have tried to improve upon pixel wise supervision by implementing new neural network architectures such BiFPN and DenseNet. Moreover in a totally separate experiment we have tried to utilize the famous siamese network in differentiating fake images from real ones. As the project progressed we observed how using BiFPN along with pixel wise supervision provided minor improvements over the vanilla Pixel Wise Supervision. Using our experiments it is also proved how the siamese network can be used as a lightweight method for anti-spoofing where not much data is provided for learning and the task of differentiating fake from real needs to be computationally in-expensive.