What are deep fakes and how can you spot one?

Recently, a viral deepfake video featuring a woman who closely resembled actress Rashmika Mandanna has caused distress and concern in the entertainment. The video, created through AI, depicted her entering an elevator and forced influential figures like Amitabh Bachchan and Union Minister Rajeev Chandrasekhar to express their concern on the subject. 

The incident also raised issues related to identity theft and the grave risks associated with the advanced and more accessible tech tools. In this report, we will discuss what are deep fakes how they are made, and how can you identify them using some basic caution. 

What are Deepfakes?

Deepfakes are a type of artificial intelligence (AI) technology that involves the manipulation of digital media, typically videos or images, to create convincing and often deceptive content. The term "deepfake" is made up by joining two words; "deep learning" and "fake." Deep learning refers to the use of neural networks which are used to analyse and generate data.

Deepfake technology employs deep neural networks to replace one person's face and voice with another's in a video, making it appear as though the target individual is saying or doing things they never actually did. 

This is achieved through a process known as face swapping or voice cloning, which uses a large dataset of the target person's images or voice recordings to create a convincing imitation. This data is generally readily available for celebrities in large quantities making them most susceptible to deep fakes. 

When and Why Deepfakes became popular?

Deepfakes first got mass attention in the late 2010s, gaining popularity and widespread attention for several interconnected reasons. These digital manipulations of media, be it videos or images, became known as "deepfakes" in 2017 and have since become both a subject of fascination and concern. The rapid advancement in AI technology and democratisation of AI models made deep fakes much easier to generate. 

Deepfake technology has not only been a source of controversy. It has also been employed for creative and entertainment purposes. Some individuals used it to insert actors into classic movie scenes, generate humorous content, or create fictional scenarios. 

Governments, tech companies, and researchers have launched initiatives to address the risks associated with deepfakes. These efforts focus on developing detection methods, establishing regulations, and implementing preventive measures to mitigate the negative impacts of deepfake technology. 

How Deepfakes are made?

Deepfakes are created using a combination of artificial intelligence, machine learning, and computer vision techniques. The process of making a deep fake typically involves collecting a significant volume of data for both the target person and the source person (providing facial expressions or voice samples). This data may encompass a variety of media, such as images, videos, or audio recordings.

Then the collected data undergoes preprocessing following which it is typically processed through a generative adversarial network (GAN) or an autoencoder. GANs consist of a generator, responsible for crafting fake content, and a discriminator, tasked with evaluating the authenticity of the content. These two components work in opposition, with the generator striving to produce convincing content and the discriminator attempting to distinguish between real and fake.

After this, the facial expressions or voice characteristics of the target person are extracted from their data. This is followed by using trained neural networks for face-swapping or superimposing the target person’s face on the source person’s body. 

In post-processing adjustments in lighting, color, and other visual elements are made to match the original media. Once all the necessary adjustments and enhancements are made, the final deepfake content, be it a video, image, or audio recording, is generated.

How do you spot a deepfake?

Detecting deepfakes can be challenging because they are designed to look and sound convincing. However, there are several methods and techniques that can help you spot a deep fake:

• Deepfakes generally have inconsistencies in facial expressions, especially unnatural or exaggerated facial movements can be a sign of a deepfake.
• Deepfakes often struggle with realistic eye movements and blinking. If a person's eyes appear unnatural, with a lack of blinking or strange eye behaviour, it could be a red flag.
• Pay attention to the synchronisation of lip movements and speech. Deepfakes may not perfectly match the lip movements with the spoken words.
• Look for visual artefacts, such as blurriness, strange lighting, or unnatural skin textures. These can be indicative of a deepfake.
• If the audio quality is inconsistent, with changes in background noise or voice pitch, it may suggest a deepfake. 
• Consider the source of the video or image and the context in which it is presented. If the content is from an unverified or unreliable source and seems out of place, exercise caution.
• Exercise critical thinking and scepticism when encountering media content. 

Indian laws on tackling illegal deep fakes

Information Technology Act, of 2000, and the IT Intermediary Rules, have implications for addressing issues related to deepfakes in India.

Information Technology Act, 2000: Section 66D

This section deals with "Punishment for cheating by personation by using computer resources." It stipulates that individuals found guilty of impersonation using a computer resource can face imprisonment for up to 3 years and a fine of up to Rs 1 lakh.

IT intermediary rules

The IT Intermediary Rules, particularly Rule 3(1)(b)(vii) and Rule 3(2)(b), address the responsibilities of social media intermediaries in dealing with deep fake content:

Rule 3(1)(b)(vii): Social media intermediaries are required to observe due diligence, which includes ensuring that their rules and regulations, privacy policy, or user agreement explicitly inform users not to host any content that impersonates another person.

Rule 3(2)(b): Social media intermediaries are expected to respond swiftly to user complaints regarding content that involves impersonation in electronic form, including artificially morphed images of individuals. Within 24 hours of receiving such complaints, intermediaries must take measures to remove or disable access to such content.

These rules empower the government and law enforcement to take legal action against those responsible for creating and sharing malicious deepfake content. They also place responsibilities on social media intermediaries to actively combat the spread of deepfakes on their platforms by establishing clear policies and mechanisms for addressing impersonation-related complaints.