How to Detect Deepfakes: Signs, Tools, and Protection

Deepfakes—AI-generated videos, images, and audio that convincingly mimic real people—have become a significant security threat. Scammers clone voices to impersonate family members in distress, attackers fabricate video calls with executives to authorize fraudulent wire transfers, and disinformation campaigns deploy synthetic media to manipulate public opinion. These attacks often combine with traditional phishing techniques to create highly convincing social engineering campaigns. The NSA, FBI, and CISA released guidance warning that "threats from synthetic media, such as deepfakes, have exponentially increased" between 2021 and 2022.

Detection is getting harder, not easier. While AI tools exist to flag suspicious content, they're locked in an arms race they're losing—each new generative model is specifically designed to defeat existing detection algorithms. The good news: deepfakes still fail at the edges of human behavior and physics, struggling with the tiny, unconscious things we do without thinking.

TL;DR

What they are: AI-generated videos, audio, and images that convincingly impersonate real people
Why they matter: Used for financial fraud, phishing attacks, disinformation, and identity theft
Key takeaway: Verify through separate communication channels—never authorize actions based on video or voice alone

What Are Deepfakes?

Deepfakes are AI-generated synthetic media that present someone's identity in an inauthentic context. The technology relies on Generative Adversarial Networks (GANs), where two AI systems compete: a generator creates fake content while a discriminator attempts to detect forgeries. Through thousands or millions of iterations, this process produces increasingly convincing synthetic media.

The term "deepfake" originally described face-swapped videos, but now encompasses voice cloning, full-body synthesis, and even text-to-video generation. Modern deepfake tools can clone your voice from just three seconds of audio, swap faces in real-time video calls, and generate entire conversations that never happened.

How Deepfakes Are Used in Attacks

Criminals deploy deepfakes in several ways:

Voice cloning scams are one of the most prevalent deepfake threats. Scammers harvest brief audio clips from social media or voicemails, then use AI to clone voices. They call family members claiming to be in an emergency, requesting immediate wire transfers. These attacks have escalated alongside other social engineering techniques that exploit trust relationships.

Business email compromise attacks increasingly leverage deepfake technology. Attackers create synthetic video calls impersonating executives to authorize fraudulent transactions, similar to the sophisticated BEC campaigns targeting the energy sector that combined multiple attack vectors. In one high-profile case, scammers used deepfake video to impersonate a CFO during a video conference, convincing an employee to transfer $25 million.

Identity Fraud: Deepfakes bypass facial recognition systems for account takeovers, loan applications, and remote identity verification. Attackers combine stolen identity documents with AI-generated video to defeat "liveness checks" used by financial institutions.

Disinformation campaigns represent a growing threat, with nation-state actors and threat groups deploying deepfakes to manipulate public opinion, disseminate false narratives about political or military matters, and generate widespread confusion.

Visual Clues That Reveal Deepfakes

Modern deepfakes fail at the edges of human behavior. Watch for these telltale signs:

Eyes and Blinking

Real humans blink spontaneously every 2-10 seconds. AI-generated faces often stare without blinking for unnaturally long periods or display mechanical blinks lacking the subtle muscle movements around the eyes that accompany genuine blinks. When they do blink, the motion looks robotic rather than organic.

Head Movements and Profile Views

Most deepfake models train primarily on front-facing data. When a synthetic face rotates to a full profile, the rendering breaks down—the ear might blur, the jawline detaches from the neck, or glasses melt into skin. Look for inconsistencies when the subject turns their head.

Skin Texture and Aging

Pay attention to whether skin appears too smooth or too wrinkly. Check if the agedness of the skin matches the hair and eyes, as deepfakes may be incongruent across these dimensions. Real skin has visible pores, blemishes, and texture that AI often over-smooths or renders inconsistently.

Hair Physics

Watch how hair moves. Real hair flows as individual strands responding to gravity and movement. Deepfake hair often moves as a solid mass or displays unnatural physics, especially around the face and shoulders.

Jewelry and Accessories

Deepfakes struggle with objects that partially occlude the face. Jewelry may morph, disappear, or reappear inconsistently during movement. Glasses can display strange reflections or appear to merge with facial features.

Lighting and Shadows

Look for lighting inconsistencies—shadows that don't match the apparent light source, reflections in eyes that don't correspond to the environment, or unnatural color differences between edited and unedited portions of the image.

Audio Patterns That Expose Synthetic Voices

Human speech includes natural breathing patterns. AI audio often inserts breath sounds at syntactically wrong moments or loops identical breath sounds throughout. Listen for these red flags:

Unnatural Breathing: Real speakers breathe at grammatical breaks—between sentences or during pauses for thought. Synthetic voices place breaths randomly or use repetitive breathing sounds.

Phoneme-Viseme Mismatches: In video, watch whether lip movements match the sounds being produced. Deepfakes sometimes display timing issues between mouth shapes (visemes) and speech sounds (phonemes).

Voice Timbre Inconsistencies: AI-generated voices may shift in pitch or timbre in ways human vocal cords can't replicate, especially during emotional speech or transitions between words.

Speech Pattern Anomalies: Synthetic voices can sound rehearsed or mechanical, lacking the natural stumbles, corrections, and filler words ("um," "uh") that characterize genuine human speech.

Detection Tools and Their Limitations

Several software tools attempt to identify deepfakes, but they're fighting a losing battle:

Available Detection Software

Microsoft's Video Authenticator analyzes still photos or videos to provide a confidence score, detecting the subtle grayscale elements and blending boundaries undetectable to the human eye. Intel's FakeCatcher examines blood flow patterns in facial pixels, identifying the subtle color changes caused by oxygen levels in blood.

Other tools include Sensity AI, Deepware Scanner, and Deepfake-o-Meter. These systems analyze facial features, movement patterns, audio-visual synchronization, and behavioral indicators to assess whether media has been manipulated.

Why Detection Software Fails

No detection tool is foolproof. Accuracy drops when lighting conditions, facial expressions, or video quality differ from the training data. Each new generative model is specifically engineered to defeat existing detection algorithms—it's an arms race where attackers iterate faster than defenders.

Tools claiming "90% Real" confidence scores provide no authenticity guarantee. The NIST guide on synthetic media emphasizes that digital media claimed to be deepfakes should be authenticated to certify whether it's synthetic, rather than relying on automated detection alone.

Content Credentials (C2PA)

The Content Authenticity Initiative developed C2PA, a standard that cryptographically signs digital content at the moment of capture, creating a tamper-evident chain of custody. Camera manufacturers are beginning to embed C2PA signatures in photos and videos.

The problem: most social media platforms strip metadata to reduce file size, effectively deleting the C2PA manifest. Until platforms preserve content credentials, this authentication method remains limited.

Practical Defense Strategies

Human verification procedures outperform technological solutions. Here's how to protect yourself:

Personal Protection

Establish Safe Words: Create a random "safe word" with family members for voice verification calls. If someone calls claiming to be your child in distress, ask for the safe word. Scammers can't respond because they don't know it—simply hang up if they can't provide it.

Restrict Social Media Visibility: Limit personal videos to friends-only visibility. Scammers harvest brief audio clips from public social media posts for voice cloning attacks.

Enable Device Security: Use biometric protections like Apple's "Stolen Device Protection" to prevent attackers from using deepfakes to unlock your devices.

Organizational Protection

Multi-Channel Verification: Never authorize financial transactions based solely on video or voice communication. Require verification through a separate, trusted channel—call the person back at a known number, send a text message, or use a pre-established authentication method.

Request Physical Actions: On video calls, ask the person to perform specific physical actions like holding up fingers or moving their hand in front of their face. Current deepfakes struggle to render hand occlusions convincingly.

Implement Transaction Controls: Establish approval workflows requiring multiple verifications for high-value or unusual transactions. The business email compromise attacks that targeted energy companies relied on single points of verification that deepfakes could defeat.

Train Employees: Educate staff about deepfake threats, especially those with authority to authorize payments or access sensitive data. Include deepfake awareness in security training programs.

The Detection Arms Race

The gap between generation and detection is widening. Here's why:

Generative models are trained on diverse, high-quality data and designed specifically to produce content that defeats detection algorithms. Detection models, by contrast, must train on examples of both real and synthetic media—but they're always playing catch-up, training on yesterday's deepfakes rather than tomorrow's.

New generation techniques like diffusion models leave different "fingerprints" than GANs. Detection tools trained to spot GAN artifacts fail against diffusion-generated media. By the time detection catches up, the next generation of tools has already deployed.

No single detection method reliably catches all deepfakes. Research published by NIST emphasizes that combining multiple detection methods—visual analysis, audio analysis, behavioral patterns, and metadata verification—is necessary to achieve high detection rates.

What This Means for You

The era of passive trust is over. Here's your action plan:

Never trust media alone: Verify through separate channels before taking action
Watch for the tells: Blinking patterns, head movement artifacts, audio inconsistencies
Establish verification procedures: Safe words for personal calls, multi-channel confirmation for business
Limit exposure: Restrict public sharing of video and audio that could be weaponized
Stay informed: Detection techniques evolve—stay updated through security resources like our security guides

Detection technology won't save you. The deepfake generators will always be ahead because they control the generation process, while detection systems must reverse-engineer artifacts from finished media. Your best defense is skepticism combined with verification procedures that deepfakes can't defeat.

Frequently Asked Questions

Can I trust deepfake detection apps? No. Detection apps provide confidence scores, not guarantees. They're useful for flagging suspicious media, but should never be your only verification method. New generative models are specifically designed to fool existing detection tools.

How long does it take to create a convincing deepfake? With modern tools, voice cloning takes as little as three seconds of audio. Face-swapped videos can be generated in minutes using consumer-grade hardware. Real-time deepfakes that work during live video calls now exist.

Should I avoid posting videos online? Not necessarily, but restrict personal videos to friends-only visibility and avoid posting audio clips that clearly capture your voice. The more public audio and video of you that exists, the easier it becomes for attackers to create convincing deepfakes.