A deepfake is video, audio or an image created or altered with AI so that it shows a real person doing or saying something they never did. The word mixes 'deep learning' and 'fake'. Early deepfakes were clumsy; today a convincing face swap or cloned voice can be made with free apps in minutes.
Deepfakes fall into three groups: face-swap video, voice cloning and fully synthetic media (people or scenes that never existed). Criminals use them for fraud, impersonation and misinformation, while filmmakers and educators use the same tech legitimately.
The good news: every generation method leaves traces. Unnatural spectral patterns in audio, missing blinks and over-smooth pixels give fakes away, and that is exactly what VigilAI checks.
How AI Voice Cloning Scams Work (and How to Stop Them)
In a voice cloning scam, a fraudster copies a few seconds of someone's voice from an Instagram reel, YouTube video or voice note, then feeds it to an AI model. The cloned voice calls a parent or friend and begs for urgent money after a fake accident or arrest.
These calls work because they trigger panic. The voice sounds right, the caller ID may be spoofed and the scammer demands secrecy and speed.
How to protect yourself: hang up and call back on the saved number, agree a family safe word, never send money under pressure, and record the call audio and scan it with VigilAI's voice detector. Real voices show natural spectral variation; cloned ones often look unnaturally uniform.
Scammers now join Zoom, Google Meet and WhatsApp calls as a fake boss, banker or relative. In a well-known 2024 case, a Hong Kong finance employee at engineering firm Arup was tricked by a deepfake video meeting into transferring about US$25 million.
Warning signs: blurry edges around the face or hair, lips out of sync with speech, flat lighting that does not match the room, rare or odd blinking, and refusal to turn the head sideways. Ask the person to wave a hand across their face or turn side-on; face-swap models often glitch.
For stronger checks, run VigilAI's webcam liveness scan alongside the call, and always verify large payment requests through a second channel.
MFCC stands for Mel-Frequency Cepstral Coefficients. They describe the short-term power spectrum of sound on a scale that mimics human hearing, and are a standard feature in speech recognition and speaker identification.
A real voice changes constantly: breath, throat tension and tiny timing irregularities create rich variation in MFCC values. Many synthetic voices are smoother and more uniform, so their coefficient variance is suspiciously low.
VigilAI extracts 40 MFCC coefficients with LibROSA and measures their variance, then combines it with other signals. Because it analyses acoustics rather than words, it works on Hindi, Tamil, Bengali, English and any other language.
People blink roughly every 4 to 6 seconds, and each blink follows a natural pattern. Early deepfake models were trained mostly on open-eyed photos, so fakes blinked too little or oddly.
Tools like Google's MediaPipe map 468 facial landmarks every frame. From the eye landmarks we compute the eye aspect ratio: it drops sharply during a blink. A face that never dips, or dips in a mechanical rhythm, is a red flag.
Modern fakes have improved, so blink analysis is never used alone. VigilAI combines it with jitter tracking on the nose bridge and pixel analysis to keep accuracy high.
Fourier Analysis: Finding the 'Too Smooth' Look of AI Faces
Real camera footage carries natural high-frequency detail: sensor grain, skin texture, fine hair. Generated faces and swapped masks are often statistically smoother because the model blends and upsamples pixels.
A 2D Fast Fourier Transform converts an image into its frequency components. By inspecting how spectral energy is distributed, a detector can see missing high-frequency data or odd repeating patterns left by upscaling.
This is one of three layers in VigilAI, which cross-checks voice, face and pixel evidence so a single weak signal does not decide the verdict.
Deepfake and AI Scam Calls in India: How to Protect Your Family
India has seen a rise in fake police, courier and 'relative in trouble' calls, and AI makes them more believable. Elderly family members and small business owners are common targets.
Simple family rules: set a secret safe word; never act on a call that demands secrecy or instant UPI transfers; hang up and call back on a known number; report fraud on the national cybercrime portal (cybercrime.gov.in) or helpline 1930.
Share VigilAI with relatives. It is free, needs no account and supports any Indian language since it analyses sound patterns, not words.
Video KYC and remote onboarding made banking convenient, and gave fraudsters a new target. Attackers can inject a face-swapped or pre-recorded feed to open mule accounts or take over existing ones.
Strong defences combine liveness detection, voice verification, device and camera-injection checks, and human review for high-risk cases. No single check is enough because attackers adapt.
VigilAI is exploring API access so fintechs and banks can add multi-signal deepfake checks to onboarding. Interested? Email abhishek@vigilai.online.
10 Visual Clues to Spot AI-Generated Images and Videos
No clue is proof alone, but several together are strong evidence. Check: 1) hands and fingers; 2) teeth and ears; 3) jewellery and glasses that merge into skin; 4) garbled text on signs; 5) inconsistent shadows; 6) backgrounds that warp.
Also look at 7) skin that is plastic-smooth; 8) hair strands that blur or vanish; 9) lip-sync drifting from audio; 10) unnatural blinking or a face that stays oddly still.
Then verify the source: reverse-image search, check the original uploader and look for coverage by trusted outlets. Upload suspicious clips to VigilAI for a signal-based second opinion.
Generators and detectors improve together. As fakes get cleaner, detectors must look beyond single artifacts to combined signals: audio, face, pixels, metadata and behaviour.
Expect more content provenance (such as C2PA signatures proving where media came from), watermarking by AI providers, and platform-level labels. Still, provenance only helps when creators cooperate, so detection stays essential.
VigilAI's roadmap follows this path: more weighted signals, better mobile experience and developer APIs, while keeping the core scanner free for everyone.