Neural Defend
Start Here
Deepfake ThreatsDeepfake Threats4 min read

What Are the Main Types of Deepfake Attacks?

The main types of deepfake attacks are executive and colleague impersonation on video or voice calls, voice-clone scams against families and bank customers, identity verification bypass with face swaps and forged IDs, remote hiring fraud, sexual extortion with fabricated images, and disinformation or market manipulation. Each targets a decision point where a face or voice releases money, access or trust. Neural Defend detects the manipulated media at those points.

Organization authorNeural Defend Editorial
Updated October 24, 2025
Sumit Singh, Co-founder and CDO of Neural Defend.
Sumit SinghCo-founder & CDO
Conceptual 3D render illustrating What Are the Main Types of Deepfake Attacks?

What Are the Main Types of Deepfake Attacks?

A deepfake attack uses AI-generated or AI-altered media to make a target believe they are seeing or hearing someone they trust, or something that happened. The media is only the delivery mechanism. What defines the attack is the decision it is trying to trigger, so the most useful way to classify deepfake attacks is by the workflow they target.

The table maps the six attack types that public agencies have documented most often to the modality used, the workflow under attack and the control that breaks it. Most real incidents combine more than one, for example a phishing email followed by a deepfaked video call.

Attack typeModalityTarget decisionControl that breaks it
Executive or colleague impersonationLive video, cloned voice, emailPayment release, data transferCallback on a directory number, dual approval, payment hold
Voice-clone scamsCloned voice on phone or voice noteFamily emergency payment, bank account accessCode word, callback, voice deepfake detection
Identity verification bypassFace-swapped selfie, injected video, forged IDAccount opening, account recoveryDeepfake detection on captures, injection and liveness checks, document forensics
Remote hiring fraudLive face swap, voice spoofing, stolen identityJob offer and system accessIdentity proofing, live challenges, background checks
Sexual extortion and harassmentFabricated explicit images or videoPayment under threat, reputational harmTakedown, reporting, forensic support for victims
Disinformation and market manipulationFabricated video, images or audio of officials or executivesVotes, trades, public safety responsesVerified channels, provenance, rapid forensic review
Deepfake attack types by target workflow, with the controls that break them.
A tray of image, audio and video samples connected by dotted lines to pause cards and inspection tools such as a magnifier and compass.
Light enterprise still life for deepfake attack types: what-related verification objects on a white surface, no people.

How Do Impersonation and Voice-Clone Attacks Work?

Impersonation attacks borrow the authority of someone the target already trusts. In January 2024 a finance employee at Arup in Hong Kong doubted an email from the UK office asking for a secret transaction, then joined a video call where the chief financial officer and other colleagues appeared. Hong Kong police said all of them were deepfake recreations. The employee made 15 transfers totalling HK$200 million, and Arup confirmed to CNN in May 2024 that fake voices and images were used.

Voice-clone attacks need even less. The US Federal Trade Commission warned on 20 March 2023 that a scammer needs only a short audio clip of a family member's voice, which may come from content posted online, to stage a convincing emergency call. The FBI's Internet Crime Complaint Center added in December 2024 that criminals clone account holders' voices to gain access to bank accounts.

How Do Deepfakes Attack Identity Verification and Hiring?

Identity attacks aim at onboarding and account recovery, where a face and a document open an account. FinCEN's alert FIN-2024-Alert004, issued on 13 November 2024, reported an increase in suspicious activity reports from financial institutions describing suspected deepfake media, mainly identity documents created with generative AI to circumvent identity verification and authentication. Attackers pair those documents with face-swapped selfies or fully synthetic faces, and inject the video through virtual cameras or emulators so it never passes through a real camera.

Hiring attacks use the same tools on recruiters. The FBI's IC3 warned on 28 June 2022 of applicants using deepfakes and stolen personal information to apply for remote IT, programming and database roles, some with access to customer and financial data. Interviewers noticed that lip movements did not fully match the audio and that coughs and sneezes were out of step with the video.

Layered translucent panels of mesh, noise and pixel grids linked by a chain of round tokens to separate texture, waveform and noise samples.

How Are Deepfakes Used for Extortion and Disinformation?

Extortion attacks fabricate compromising material. The FBI's December 2024 announcement lists fake explicit images of victims used to demand payment in sextortion schemes, alongside images used in market manipulation and fabricated disaster or conflict imagery used to solicit fake charity donations. Disinformation attacks fabricate statements by officials or executives to move opinion, votes or prices.

Fabricated evidence is a related threat inside organizations. In April 2024 Baltimore County police announced that a recording attributed to a Pikesville High School principal, which had circulated widely, was AI-generated, and charged the school's athletic director. Workplace and school disputes increasingly involve recordings whose authenticity is contested.

Where Do Deepfake Attacks Enter an Organization's Workflows?

Attacks enter through channels where media is accepted as proof: selfie and document capture in onboarding apps, video meetings, phone lines and voicemail, help desk and contact centre calls, uploaded claims photos, and social media posts about the company. Each channel has a natural place for a check, usually just before the decision the attacker wants.

Detection is most effective when it runs automatically at those points and passes its result to the system that makes the decision. A selfie can be scored before the account is approved, a claims photo before the payout, and a recorded call before a password reset is completed. Live video meetings are harder, which is why procedural controls carry more weight there.

A tray holding a checklist card, photo prints, a waveform strip, film negatives, pause and search cards and a magnifying glass.

Which Controls Stop Deepfake Attacks?

Effective defence layers technical and procedural controls so that one failure does not decide the outcome. Technical layers include deepfake detection on images, video and audio; liveness and injection-attack checks at capture; document forensics; and provenance checks where content carries C2PA credentials. Procedural layers include callbacks to numbers on file, dual approval above a payment threshold, holds on new payees, code words for families and executives, and a clear route for staff to pause a request without penalty.

Training matters less than policy. People cannot reliably spot good deepfakes by eye, so instruct them to verify requests rather than judge media, and give them the authority to do so. Measure each control by the attacks it would have stopped and the genuine customers or colleagues it slows down.

How Does Neural Defend Cover These Attack Types?

NeuroVerify covers identity attacks at onboarding, account recovery and claims. One API call checks a face image, video or audio and returns a unified risk score with evidence from six signals: Visual Safety Screening, Live Presence Verification, Face Manipulation Detection, AI Generative Model Analysis, AI Pixel-Level Manipulation and AI Content Compliance. Eligible image requests return in under 200 ms p95. SDKs cover Android, iOS and Flutter on mobile and Python, TypeScript, Go and Java on the backend, with a REST API and an MCP server for AI agents, and integration takes two lines of code.

NeuroDetect covers investigation of extortion material, disputed recordings and suspicious media in a case-based workspace with heatmaps, frame and audio-segment views and an exportable forensic report. NeuroCall (Beta) covers voice and call deepfakes. All run on ARCAS 1B and deploy as a managed cloud API, private cloud or VPC, secure container, on premises or air-gapped. NeuroVerify does not analyse ID documents, so document checks need a separate tool.

Evidence boundary

Limitations

  • Live video meetings remain hard to protect with detection alone; procedural controls such as callbacks carry most of the weight there.
  • NeuroVerify does not analyse ID documents, and NeuroCall is in Beta.
  • Public case data is incomplete: most deepfake fraud is never reported, so the attack types here reflect what agencies and companies have disclosed.
  • Detection produces a risk signal; it does not identify the attacker or decide what action to take.

Frequently asked questions

What modalities do deepfake attacks target?

Deepfake attacks use video, audio, still images and documents, and the most damaging ones combine several. Video is used for live impersonation on calls and for selfie videos in onboarding. Audio is used for voice-clone phone scams and voice notes. Images appear as fake profile photos, face-swapped selfies, fabricated explicit material and edited claims photos. Documents appear as generative AI identity cards, which FinCEN reported in November 2024 as the most common deepfake use in suspicious activity reports.

Are control layers only technical?

No. Technical layers such as deepfake detection, liveness and injection checks, and document forensics catch manipulated media at capture. Procedural layers, including callbacks on numbers already on file, dual approval for payments, holds on new payees and family code words, stop attacks even when the media is perfect. The Arup loss in 2024 is the standard example of a case a callback would have stopped. Strong programmes use both, and give staff authority to pause any request.

Can deepfake detection be integrated into existing systems?

Yes. Detection is normally delivered as an API that your onboarding, claims, contact centre or case management system calls at the point of decision, with the result fed into the existing risk engine. NeuroVerify, for example, offers SDKs for Android, iOS and Flutter, backend SDKs for Python, TypeScript, Go and Java, a REST API and an MCP server, and can be deployed in your cloud, on premises or air-gapped.

How quickly can deepfake detection respond?

Fast enough for live onboarding on still images, and more slowly for long video. NeuroVerify returns results in under 200 ms at p95 on eligible image requests under defined test conditions. In the public VendorBench-100 v2 test setup, mean still-image latency among commercial APIs ranged from 175 ms for Neural Defend to 21.9 seconds for Reality Defender, figures the paper ties to its own setup. Long video and forensic case analysis take longer and usually run asynchronously.

Why are deepfakes considered more dangerous than ransomware by some security teams?

Because they bypass the technical defences built for ransomware. A deepfake attack needs no malware, no network intrusion and no stolen password; it persuades an authorized person to act, so endpoint and network tools see nothing unusual. The losses are also immediate and often irrecoverable once a payment clears. Ransomware remains a severe threat with its own costs, and the two call for different controls: verification policy and media detection for deepfakes, and backups and patching for ransomware.

Source record

Sources and references

  1. CNNArup revealed as victim of $25 million deepfake scam involving Hong Kong employeePublished May 16, 2024. Verified October 19, 2026.
  2. Financial Crimes Enforcement NetworkFinCEN Issues Alert on Fraud Schemes Involving Deepfake Media Targeting Financial Institutions (FIN-2024-Alert004)Published November 13, 2024. Verified October 19, 2026.
  3. FBI Internet Crime Complaint CenterCriminals Use Generative Artificial Intelligence to Facilitate Financial Fraud (I-120324-PSA)Published December 3, 2024. Verified October 19, 2026.
  4. FBI Internet Crime Complaint CenterDeepfakes and Stolen PII Utilized to Apply for Remote Work Positions (I-062822-PSA)Published June 28, 2022. Verified October 19, 2026.
  5. Federal Trade CommissionScammers use AI to enhance their family emergency schemesPublished March 20, 2023. Verified October 19, 2026.
  6. Baltimore County GovernmentAthletic Director Charged in Pikesville High School AI CasePublished April 25, 2024. Verified August 7, 2026.
  7. National Institute of Standards and TechnologyReducing Risks Posed by Synthetic Content (NIST AI 100-4)Verified October 19, 2026.

Connected authority

Continue through related concepts.

Deepfake ThreatsHow Do You Stop CEO Deepfake Scams and Verify Payment Requests?4 min read
Deepfake ThreatsVoice Cloning Attacks: How They Work and How to Respond4 min read
Deepfake ThreatsWhat Is a Digital Injection Attack in Identity Verification and How Do You Detect It?6 min read
Deepfake ThreatsHow Do Deepfake Impersonation Attacks Work? The Attack Lifecycle4 min read
Deepfake ThreatsHow to Stop Deepfake Fraud: Controls That Work4 min read
Deepfake ThreatsWhat to Do After Deepfake Extortion or Sextortion5 min read
Insights libraryBrowse the complete authority collection

Protect the decision point

Add NeuroVerify where faces and voices release money or access

One API call returns a unified risk score from six signals, under 200 ms p95 on eligible image requests, with SDKs for mobile and backend.

Explore NeuroVerify