Deepfake Fraud Set to Surge 495% in 2026

Key Takeaways

Four findings from analysed fraud attempts

What a Deepfake Actually Is, and Why It Scales Like Software

A deepfake is a synthetic media asset, video, image, or audio, in which artificial intelligence is used to fabricate or manipulate a person's likeness or voice. What was once a computationally expensive, specialist capability has become accessible through open source models and low cost cloud infrastructure.

The barrier to entry has collapsed. Anyone with a consumer device and an internet connection can now generate a convincing fake in minutes. That shift from expensive to cheap, from rare to widespread, is precisely why deepfake fraud is scaling the way software scales: marginal cost approaches zero as volume approaches infinity.

The shift is from craft to assembly line. A few years ago, a believable fake face needed a skilled editor and hours of work. Today an attacker feeds one image or a text prompt into a generative model and gets back a fully animated face that blinks, turns its head, lip-syncs, and reacts in real time, ready to drop into a selfie check or a live video call.

That industrialization is why the growth in this report climbs the way it does. When fraud becomes software, it scales like software.

The Deepfake Fraud Stack Behind the Surge

Four AI-driven categories make up the threat. Synthetic Identity account for the largest share, but all four are climbing, and Document Deepfakes is accelerating the hardest.

The mix is no longer dominated by any single technique. Synthetic Identity lead, live video deepfakes and Face Swaps's follow, and document deepfakes, though the smallest share today, is rising fastest:

All four AI attack types: Projected Annualized Growth for 2026

Document Deepfake - The Fastest Growing Type

One category is driving the spike. Document Deepfakes is the fastest-growing of the four, projected to rise nearly 3892% year over year in 2026 (annualized estimate), by far the sharpest mover in the dataset.

A note on method: 2026 figures cover January to May only. When we project a full year, we annualize the five-month run rate and label the result 2026E, an estimate, not a confirmed figure.

Synthetic Identity: The Largest Type, Still Climbing

Full Synthetic Identities, fabricated personas with no real world counterpart, represent the largest single AI fraud category and continue to grow. First tracked as a distinct category in 2025, they are projected to rise approximately 173 in 2026. The synthetic identity problem is scaling, not slowing.

Face Swap: A Newer Category Gaining Ground

Face-swap attacks map a victim's face onto an attacker's head in real time. Tracked as a distinct category from 2025, they're projected to grow about +121 in 2026, a smaller but steady climb that widens the range of live-video threats teams have to block.

Deepfake Video: Climbing Every Year, Now an Established Threat

Deepfake video, manipulated live streams during verification, has the longest track record of the four types, and it has grown every year since 2023. By 2025 it had become the second-largest AI fraud type. It remains a core, persistent threat that layered liveness is built to catch.

2026 omitted: only Jan–May data is available and is not yet annualized for this category.

How the Attacks Actually Reach the Camera

Attackers reach the camera in three broad ways, and most real campaigns combine them. Understanding the overlap is the difference between blocking one technique and blocking the chain.

Deepfake Presentation attacks

Hold something up to a real camera, a silicone mask, a printed photo, or a live video deepfake playing on a screen.

Deepfake Injection attacks

Skip the camera entirely, piping an AI-generated feed straight into the verification app through virtual-camera or emulator software.

Deepfake-Enabled Synthetic Identity Creation

Stitch a generated face to fabricated documents to manufacture a brand-new person who has never existed.

How Attack Vectors Combine

Most campaigns overlap — a layered defense is required.

Why Human Review Can’t Catch Up to High-Quality Deepfakes

Manual review is no longer a reliable backstop against high-quality deepfakes, because humans are less likely to spot a live video deepfake than a deepfake image.
When the human eye is overconfident, and the fakes are this good, a reviewer won’t be enough.

How Modern Detection Works: Three Layers, Not One Selfie Check

Effective deepfake detection is layered, no single signal is enough, so defenses stack independent checks an attacker has to beat all at once. Shufti uses a three-layer approach.

Shufti’s models carry independent validation: iBeta Level 3 passive-liveness conformance with 0% APCER and 0% BPCER across iOS and Android, and a 98.49% true-accept rate with zero false template creation at the DHS Remote Identity Validation Rally.

Trust, Regulation, and the Financial Stakes

The regulatory and threat trajectories point the same way: detecting synthetic inputs is becoming a baseline expectation, not a differentiator. Three external markers stand out.

• Standalone selfie checks are losing trust. Gartner predicts that by 2026, 30% of enterprises will no longer consider identity verification reliable in isolation because of AI-generated deepfakes, and noted injection attacks rose 200% in 2023.
• The financial stakes are rising. Deloitte's Center for Financial Services estimates generative AI could enable fraud losses of up to $40 billion in the US by 2027, up from $12.3 billion in 2023.

How Shufti Catches These Attacks

Shufti has been developed to detect those four deepfake attack types on remote onboarding systems. Its deepfake and biometric stack is 100% proprietary, with no third-party models in the chain, and Shufti is a certified iBeta Level 3 liveness provider. Each attack type meets a control designed for it, so a synthetic input has to beat several independent checks at once.

Shufti was built to stop exactly the four attack types in this report. Its deepfake and biometric stack is 100% proprietary, with no third-party models in the chain, and Shufti is a certified iBeta Level 3 liveness provider. Each attack type meets a control designed for it, so a synthetic input has to beat several independent checks at once.

360° Deepfake Detection: Shufti's Seven Gates of Defense

Shufti runs deepfake detection as a seven-part forensic model. Rather than trusting one cue, it puts every face and video through seven independent authenticity checks, each probing a different trace a fake leaves behind. Media is only cleared when all seven align, so no single trick can slip past, and the checks are built to hold up even after compression, screenshots, and re-uploads.