COMBATING DEEPFAKES
By Akshay Govind Hande, Intern, Seth Associates
Keywords: Deepfakes, Cyberlaw, cybersecurity, ItAct, AI
What is deepfake ?
A deepfake is an artificial image, video, sound, text or series of them created using a special kind of machine learning called deep learning.In simple words, deepfake is similar to any kind of machine learning where an algorithm is provided with examples and it learns to produce similar output that resembles the example provided.
Evolution of Deep fake
19th century& 20th Century saw the development of photo manipulation, applied to motion pictures further in 20th century, digital video was introduced. In 1990, researchers in computer graphics (CGI) were trying to create realistic human image. In 1997, Christopher Bregler published a research paper titled Video Rewrite: Driving visual speech and audio. These could synthesize new facial animation based on audio input which movie studios did with complex technique.
In 2010, large amount of data for training AI was available and deeplearning a subfield of machine learning made significant progress. in 2014, Generative Adversarial Network was introduced by Ian Goodfellow and his colleague’s in their research paper. In 2017, the term deepfake emerged when an anonymous Reddit user named deepfakes posted manipulated videos on platform. These videos used machine learning algorithms to superimpose celebrity faces onto adult film actor’s bodies. 2018 – 2020 during this period variant of GAN such as styleGAN enhanced the quality of generated images and videos these models could create highly realistic human faces. User friendly deepfake creation tools and apps like Faceapp, DeepFaceLab, Zao emerged making the technology accessible to broader audience. Presently researchers are developing AI based detection tools to identify deepfakes by analysing inconsistencies in visual and audio data, companies like Facebook and Microsoft invested in deepfake detection technology.
How are deepfakes created ?
GAN (Generative Adversarial Network):
The process of producing complex deepfakes involves two algorithms, one algorithm called the Generator is trained to produce the best fake replicas of real image and the other algorithm called the Discriminator is trained to detect when an image is fake and when it’s not. These both models are put back and forth each getting better at their respective task, once the models are pitched together they create an extremely realistic fake image which is so adept that in fact humans can’t tell if the output is fake or real. The combination of these generator and discriminator algorithm creates a GAN generative adversarial network, an example of this network would be imagine a two children playing a game where one child tries to draw a picture of an object and other child judges it, the game continues till the drawer gets a realistic picture of the object and the judge gets better at spotting mistakes. Similarly, the generator algorithm creates an image based on reality studying it from different angles and the discriminator examines it whether the image is fake or real this process is repeated multiple times till the creator looks realistic.
Autoencoders:
Thousands of face shots are processed of the two people through an AI algorithm called an Encoder thus encoder finds and learns similarities between these two faces and compresses them down to shared common feature images. A second AI algorithm called Decoder are two machines, one is trained to take compressed image and reconstruct first persons face and the other machine is trained to reconstruct the second person’s face. To face swap, the compressed face image of first person is placed into decoder trained on another person, this decoder reconstructs second persons face but with the expression and orientation of first person. For a convincing face swap video this process is applied to every frame of the video. Each frames face is encoded, swapped and decoded to maintain consistent expressions and orientations across the entire video.
Forms of Deepfakes
Face-Swapping Deepfakes: The most common where AI replaces a person’s face in a video with another person’s face, which can be used to make it look like a celebrity who is supporting or encouraging a product they never endorsed.
Voice Synthesis Deepfakes: These target audio wherein the AI model is trained on a person’s voice recordings to learn their speech its patterns and their vocal characteristics. This helps in generating the synthetic speech which mimics exactly like the targeted person thus in a way it enables the creation of fake audio messages or impersonation.
Lip Sync Deepfakes: These change the lip movements of a person in a video to match a different audio track, which creates the illusion showing that they are saying something they did not. It is widely used in creating the natural-looking dubbed foreign language films or mostly changing the speech of a politician to mislead the public opinion.
Gesture and Body Movement Manipulation: Deep learning technique can make alteration in body movements, gestures, and expressions in a video. It usually involves showing of an individual who is nodding in an agreement but in fact they disagreed.
Text-Based Deepfakes: AI is used to generate text including articles, social media posts and email which imitates the writing style of a specific person. This technology is misused to create misleading content which appears to be created by a popular individual or organization.
Object Manipulation Deepfakes: Deep learning can also manipulate the objects within a video.It involves changing the appearance of objects in a scene for example the alteration of a brand logo. It also includes alteration of objects behaviour, for example, making a car appear to drift on a road.
Hybrid Deepfakes: Hybrid deepfakes are such wherein the technology is combined with multiple manipulation techniques, for example a deepfake that involves replacement of someone’s face or alteration in their speech and manipulation of objects in the background to create a highly convincing and complex manipulated media.
Laws regulating Deepfakes in India
Information Technology Act, 2000
Under Section 66 C: Deepfake created to exploit an individual’s unique identification features such as their face, voice, body language and the manner in which they behave or deepfakes that involve electronic signatures or passwords. Under Section 66 D, creating deepfakes to use person’s image or voice for replicating them for fraudulent purposes. Under Section 66 E: criminals creating and distributing deepfakes which gains the unauthorized use of someone’s images or videos of private areas. Under Section 66F: Deepfake created resulting in widespread panic and social unrest such could be considered as cyberterrorism if it meets the requirements of this section. If Deepfakes are used to commit cheating or extortion, relevant provisions of IPC,1860 , now BNS Act,2024 would apply
Global Regulation on AI
Bletchley Declaration: 29 Nations including the US, Canada, Australia, China, Germany, and India, as well as the European Union, joined forces to prevent the catastrophic harm which is deliberate or unintentional result of stronger artificial Intelligence.
It has two Agendas , namely , What are the AI-related risks And Steps to Develop risk-based policies across nations to improve transparency by companies who are developing advanced AI technologies?
UK: To introduce National guidelines for the AI sector, currently Artificial intelligence regulation Bill is being considered.
EU: Implemented Digital Services Act, mandating social media to labelling requirements. Also strengthened its Code of Practice on Disinformation requiring social media companies like Google, Meta, and Twitter to clearly indicate deepfake content or pay fine.
South Korea: Information and Communications Network Act makes it illegal for distribution of deepfakes which harm public interests.
US: President Joe Biden signed “executive order on the safe, secure and trustworthy department and use of artificial intelligence” on October 30th 2023 to manage the AI and its risks related to national security and privacy. Section (a) of H.R.5586 of DEEP FAKES Accountability Bill, 2023 makes it mandatory for creators to label the use of artificial intelligence.
AI and its regulation are evolving set of laws which will need to be harmonised as AI has global footprint. An area which is already received consensus among nations is Responsible use of AI. Likewise, issues of criminality, IP protection, corporate ownership , attribution will need to be examined aswell in the near future.
Bibliography
ORGANIZATIONAL ARTICLES:
- University of Virginia, “What the heck is a deepfake?” https://security.virginia.edu/deepfakes#:~:text=Introduction%20to%20Deepfakes&text=Deep%20learning%20is%20similar%20to,edible%20and%20what%20isn’t.
- Homeland Security, “Increasing Threat of DeepFake Identities” https://www.dhs.gov/sites/default/files/publications/increasing_threats_of_deepfake_identities_0.pdf
- TechTarget, “what is deepfake AI?” by Nick Barney < https://www.techtarget.com/whatis/definition/deepfake>
WEB ARTICLES:
- Ian Sample, “What are deepfakes – and how can you spot them?” published on 13th January, 2020 https://www.theguardian.com/technology/2020/jan/13/what-are-deepfakes-and-how-can-you-spot-them
- Aaratrika Bhaumik, “Regulating deepfakes and generative AI in India” published on 04th December, 2023 https://www.thehindu.com/news/national/regulating-deepfakes-generative-ai-in-india-explained/article67591640.ece
- Vikrant rana, Anuradha Gandhi and Rachita Thakur, “Deepfakes and Breach of Personal Data – A Bigger Picture” published on 24th November, 2023 https://www.livelaw.in/law-firms/law-firm-articles-/deepfakes-personal-data-artificial-intelligence-machine-learning-ministry-of-electronics-and-information-technology-information-technology-act-242916#:~:text=Section%2066E%20of%20the%20IT,fine%20of%20INR%202%20lakh.
- Harshvardhan Mudgal, “The deepfake dilemma: Detection and decree” published on 18th November 2023 https://www.barandbench.com/columns/deepfake-dilemma-detection-and-desirability
- Pranav Dixit, “What are different types of deepfakes and how can you spot them?” published on 25th November 2023
6.Shumbham Pandey and Gaurav Jadhav, “Emerging Technologies and law: Legal Status of Tackling Crimes Relating to Deepfakes in India” published on 17th March, 2023 https://www.scconline.com/blog/post/2023/03/17/emerging-technologies-and-law-legal-status-of-tackling-crimes-relating-to-deepfakes-in-india/