When a finance executive at a British engineering firm, Arup, joined a routine Skype meeting in Jan 2024, familiar faces filled the screen. The company leaders he had interacted with countless times spoke with authority, blinked naturally, and smiled reassuringly. Except the people on screen were not real, and the AI-crafted doubles duped the firm, which worked on the Statue of Unity and India’s rail projects, of nearly $25 million. The deepfakes were so flawless that 15 transfers from the Hong Kong office went through before anyone realised the con.Months later, a deepfake surfaced closer to home. At the India office of a global chipmaker, a man used artificial intelligence (AI) to impersonate a real job candidate in an online interview. “He synced facial movements and tone quite well but we detected the use of deepfake tech, and he was out,” says Naveen Sharma, co-founder of Kroop AI, which built Vizmantiz, a detection tool for synthetic videos and audio.
Deception turns domestic
India’s deepfake saga began in 2020, when deepfake videos of politician Manoj Tiwari “speaking fluent Haryanvi” to appeal to voters went viral ahead of Delhi assembly polls. By mid-2023, the menace had turned personal: in Kerala, a 73-year-old man lost Rs 40,000 following a WhatsApp deepfake call seemingly from his friend, pleading for urgent help from Dubai. India recorded a staggering 280% year-on-year increase in deepfake incidents in Q1 2024, particularly in the leadup to national elections, reported Sumsub, a global identity verification provider. A McAfee survey in Nov 2024 found 75% of Indians had seen deepfake content in the past year, and 45%reported that they knew someone who had been duped by a deepfake fraud.“The term ‘deepfake’ covers both synthetic content created from scratch and manipulated content that alters existing videos,” said Sharma. “Both forms distort truth: one invents it, the other rewrites it.”
Tracing AI fingerprints
Forensic experts are now learning to read what AI cannot hide. “Deepfake audio is often too clean, lacking normal background noise,” says Dr Surbhi Mathur, head of Centre of Excellence in Multimedia Forensics at National Forensic Sciences University (NFSU). “AI faces also lack natural light variations and photo response non-uniformity (PRNU) — the ‘fingerprint’ left by camera sensors. The facial micro-expressions often lacks natural blinking patterns, or the way someone moves their face or their hands near the face.”Sandeep Shukla, director of International Institute of Information Technology, Hyderabad, says, “There are tools that claim over 90% accuracy in deepfake detection. However, as these rely on neural network-based deep learning tech, there’s no guarantee that every form of media manipulation will be detected.” Police and judges must be trained in detection and its limits, he urges. “And, when guilt is proven, the punishment must be high enough to act as a deterrent.”
Faces you know, scams you know
The surge in deepfake content is driven by mass appeal, says Mathur. “Scammers exploit trusted faces for massive financial gain.” Since 2023, the deepfakes analysis unit (DAU), established under the Misinformation Combat Alliance with support from the corporate affairs and IT ministries, has tracked hundreds of AI-created scams. Fabricated “endorsements” for investment schemes and gaming apps have used the faces of Ratan Tata, N R Narayana Murthy, Rahul Gandhi, Nirmala Sitharaman, Virat Kohli, and even doctors Naresh Trehan and Devi Shetty for bogus health cures. A viral Ratan Tata “investment video” was found to be 83.8% AI-generated. Tata later publicly debunked it on Instagram.In Nov 2023, a deepfake of actor Rashmika Mandanna was created by morphing influencer Zara Patel’s face. Experts warn that ‘DeepNude’ or AI-generated explicit content created without consent pose grave risks to women’s privacy.
Tangles in law and deterrence
With no dedicated deepfake law, several Bollywood actors are taking legal recourse. Delhi HC has issued ‘John Doe’ orders that allow legal action against anonymous offenders. It introduced ‘dynamic+ injunctions’, which compel platforms to remove the original deepfake as well as proactively monitor and take down re-uploads.
Protecting the future
Sharma identified the “most dangerous” use of deepfakes in the banking, financial services, and insurance (BFSI) sector. He says that during online know your customer (KYC) processes,“I can clone a voice or face, and the system may verify that person as genuine”.To tackle this, the “first mandate” of the upcoming deepfake policy will require “BFSI to use deepfake detectors for KYC and document verification,” Sharma reveals. Amendments to the IT Rules, 2021, will enforce content labelling over 10% of the screen or audio, and faster takedowns — within 36 hours, or three hours during elections. “A person circulating or a platform hosting deepfake content, even unknowingly, will face legal action,” he says.However, experts say India must move faster. “Technology changes every 15-20 days and scammers evolve just as quickly,” warns Sharma.“We need laws, detection, and awareness that move at the same speed.”
Defending against deepfakes
KERALA’S CYBERDOME uses a public-private model for deepfake detection and digital forensics TELANGANA WOMEN SAFETY WING CYBER MODULE tackles image-based abuse, deepfake threatsNATIONAL SAHYOG PORTAL allows anonymous reporting of deepfake, cyber abuse incidents to aid state police forces
Creating tools to tackle menace
MeitY is funding three strategic deepfake defence projects through India’s IITs:PROJECT SAAKSHYA (IIT JODHPUR/ MADRAS) focuses on real-time deepfake detection for images, video, audioAI VISHLESHAK (IIT Mandi & Himachal Forensics) develops Explainable AI (Xai) for deepfakes, signature forgeryVOICE DEEPFAKE DETECTION SYSTEM (IIT KHARAGPUR) specifically targets AI-generated voice impersonation to prevent fraud Images : AI
How you can tell if it’s not real
Deepfake detection focuses on subtle, systemic errors that reveal AI manipulationAUDIO/VOICE PROMPTS : ‘Too clean’ audio lacking normal background noise is a red flagLACK OF NATURAL DELIVERY: Sounding scripted, robotic, lacking human cadence or characteristic pausesTONAL SHIFTS: Sudden changes in pitch, tone following awkward cut or pauseAUDIBLE ARTEFACTS: Distinct noise like human breathing, heard during pausesVIDEO/FACIAL CLUES: Lip synchronisation errors or the mouth moving during audio pausesSHADOW/LIGHT ISSUES: Deepfakes often miss the natural shadow effects and light variationsBLENDING/MERGING: Objects near the face (e.g., microphone, hands) merging unnaturally into the chin or jawlineTEETH IRREGULARITIES: Teeth appearing blurred, inconsistent, or lacking definition, are common flawsCOMPRESSION ARTEFACTS: Unique artefacts carried by the application that created the image
Key detection tools
FAKECATCHER (INTEL, US): Uses blood flow analysis to detect deepfakes by checking facial colouration caused by heartbeatsHIYA AI SOLUTIONS (US): Used to check if an audio is AI-generatedDEEPFAKE-O-METER (MEDIA FORENSICS LAB, US): The open platform detects synthetic contentVALIDIA (US): Proprietary software for voice and audio authenticity checkDEEPFAKE DETECTOR (OPENAI, US): Detects AI images, especially from models like DALL·E 3 Go to Source
