Privacy and Other Legal Concerns in the Wake of Deepfake Technology: Comparative Study of India, US, and China

Introduction

The goal of this chapter is to understand the potential risks involved in the use of deepfake technology. There is an old saying - “Seeing is believing”. However, in the wake of the advancing deepfake technology, it could be dangerous for individuals as well as society to believe in what they view. Considering the fact that detection and regulatory measures concerning deepfakes are still in their nascent stage throughout the world, it is pertinent to be well aware of our technological and legal capabilities in the fight against this hazard. For the same purpose, the laws and regulatory frameworks of US and China shall also be analysed since these nations have proactively taken techno-legal measures to curb the havoc that deepfakes are capable of creating. Based on our critical understanding of these measures, we would be better equipped to look for solutions in the Indian set-up.

In order to reach solutions, we need to first understand the problem at hand. It all starts with the increasing involvement of social media platforms in the preceding decade. As a result, uploading of digital images and videos is trending worldwide. This has further led to creation of a significant number of new image/video-altering techniques and apps. In the given backdrop, deepfake origination took place.

Deepfakes have a close nexus with Digital Image Forensics Research Field. The term “deepfakes” is derived from a deep understanding of the technology that is required for creation of very realistic fake videos and images. Original faces, expressions and even voices are mimicked, altered and replaced by fake ones, thereby breaching social trust and creating privacy and other legal concerns. A deepfake is a product of AI and advanced machine learning algorithm called deep learning. It utilizes Generative Adversarial Networks (hereinafter referred to as “GAN”) for morphing original content. Simply understood, the role of GAN is to train two neural net architectures i.e., a generator (decoder) and a discriminator in an adversarial relationship (Smith & Mansted, 2020). Once the encoder has extracted the latent features of original face images, the generator (decoder) reconstructs the images and the discriminator detects whether or not the image created by generator is extremely real looking.

In this way, the generator-discriminator paired algorithms constantly compete against each other and constantly evolve. As a result, any defects in a deepfake can be instantly detected and corrected, thereby making this technology difficult to combat. Recently, multi-task Convolutional Neural Networks (hereinafter referred to as “CNN”) have been developed to increase the stability of face detection and reliability of face alignment in deepfakes. It involves image mapping wherein initially the algorithms used to work on massive data sets to extract, train and create content (Gerstner, 2020). However, with the constant advancements in this technology, now even a single image can be used to render a high quality deepfake which is difficult to detect by humans as well as machines. Another characteristic of this technology is that it requires very little skill and renders quick results in the form of extremely deceptive fake content.

Morphed images and videos are not new to this world, but extremely real looking ones are a recent development. It wasn’t until 2017, when the first public example of deepfakes surfaced that experts from the technological and legal field started viewing them as a threat. In this case, the faces of Hollywood celebrities were extracted from original images and superimposed onto faces of pornographic actresses by some anonymous users of the online platform of Reddit. Apart from that, certain non-pornographic deepfakes were also created by swapping the face of Nicholas Cage into numerous movie videos, though this was purely meant for entertainment purposes. This led to widespread circulation of deepfakes, thereby turning it into a new trend. In 2018, apps such as DeepNude and FakeApp surfaced which provided both free and paid platform to users for creating deepfakes and thereby harassing, blackmailing, bullying targets and spreading fake news (Schick, 2020). In the past 4 years, multiple apps have been created and made easily available to the public at large for creating deepfakes.

Deepfakes can be broadly divided into four categories based on the method of creation employed, namely – (1) Image deepfake which employs the method of face and body swapping;