Deepfake Security - KI-generierte Täuschungen als Angriffsmittel

Deepfakes are AI-generated media—audio, video, or images—that are so realistic that people cannot distinguish them from genuine recordings. As a means of attack, deepfakes represent a new level of social engineering: whereas an attacker could previously only imitate a CEO, they can now create deceptively realistic replicas of the CEO’s voice (voice cloning) or face (face swap).

Deepfake Attack Vectors

Deepfakes as an attack vector:

1. Voice Cloning – CEO Fraud 2.0:
   Traditional CEO Fraud:
   → Fake email from "CEO" → CFO transfers money
   → Detection: Check email domain, callback
   → Countermeasure: DMARC, verification by phone

   Voice Clone Attack:
   → Attacker collects: CEO interviews, YouTube videos, earnings calls
   → 3 minutes of real voice → ElevenLabs/TortoiseTTS → perfect clone
   → Attacker calls CFO (voice call, not text!)
   → CFO hears familiar CEO voice → transfers funds!

   Real-life case: 2019, UK energy company
   → CFO received a call from the "CEO" (perfect German accent)
   → Authorized a transfer of 220,000 EUR
   → Voice was an AI-generated clone

2. Video deepfakes – video conference fraud:
   Scenario:
   → Attacker creates a deepfake video of the CEO
   → "Video conference" with CFO → CEO’s face on attacker’s camera
   → "We need this transaction IMMEDIATELY"

   Real-life case: Hong Kong 2024
   → Finance employee in a multi-person video conference
   → All other participants: deepfake videos (colleagues and CFO)
   → $25 million transferred!

3. Synthetic Photos - Identity Fraud:
   → LinkedIn profile with AI-generated photo (thispersondoesnotexist.com)
   → Fake employee → Build trust
   → Or: KYC bypass using synthetic ID documents

4. Text Deepfakes (LLM-generated):
   → Hyper-personalized phishing based on the target person
   → Style imitation: Writing style of a colleague learned from emails
   → Also: Comment bots, social media manipulation

Detection of Deepfakes

Technical detection methods:

Visual artifacts (detectable by laypeople):
  → Unnatural blinking patterns (early deepfakes did not blink)
  → Inconsistent lighting: face ≠ background
  → Blurring at the edge of the face (blending artifacts)
  → Teeth/hair: often unnatural
  → Micro-expressions: LLM videos lack sub-expressions

Automatic Deepfake Detection:
  Microsoft Video Authenticator (free):
  → Analyzes at the pixel level for GAN artifacts
  → Confidence score 0–100%

  Hive Moderation API (commercial):
  → Real-time deepfake detection via API
  → Video, audio, images
  → Integration into: KYC systems, video conferencing platforms

  Intel FakeCatcher:
  → Analyzes blood flow patterns in the face (rPPG)
  → Real people: Skin color pulsates subtly
  → Deepfakes: No blood flow → detectable
  → 96% detection rate (as of 2024)

C2PA (Coalition for Content Provenance and Authenticity):
  → Standard for digital content provenance
  → Signatures for authentic media (Nikon, Canon, Adobe)
  → Deepfakes: no valid C2PA signature
  → Supported by: Adobe, BBC, Microsoft, Sony

Audio Deepfake Detection:
  Pindrop Pulse:
  → Voice authentication against voice clones
  → Detects synthetic voices in real time
  → Applications: call centers, banking IVR

  ElevenLabs AI Speech Classifier:
  → Proprietary tool detects voices generated by its own TTS
  → Limitation: detects only its own generations!

Protective Measures Within the Company

Organizational Deepfake Defense:

1. Code Words (can be implemented immediately!):
   → Internal company verification password
   → For every unusual financial request by phone:
     "What is today’s code word?"
   → The CEO knows it; deepfake attackers do not!
   → Change: daily or weekly
   → MANDATORY for executives and the finance team

2. Verification processes:
   □ Unusual financial requests:
     → ALWAYS confirm via a second, independent channel
     → Call back using a known/saved number (NOT the one given in the call!)
     → If in doubt: personal confirmation

   □ Video conference verification:
     → For unknown participants: ask a random question only the real person knows
     → Request a physical gesture (scratch nose, raise hand)
     → Noticeable jerkiness/lag in deepfakes!

3. Technical Controls:
   → Video conferencing platform: Check for end-to-end encryption
     (does not protect against deepfakes, but against MITM)
   → Identity verification for external meetings: Calendar invitation + 2FA
   → Media watermarking for internal videos (C2PA)

4. Awareness Training:
   → Show deepfake examples in security awareness training
   → Employees should ALWAYS be suspicious, even with a "real" voice/video
   → "I recognize the voice" = not a security indicator!
   → Reporting process: report suspicious calls immediately

5. Financial Controls:
   → Dual-control principle for transfers > X EUR (always!)
   → Callback requirement: confirm all transfer requests via email/phone
   → Limits: no instant payments > Y EUR without a 24-hour processing period

Detection checklist for employees:
  □ Unusual request via phone/video?
  □ Urgency and time pressure?
  □ Callback confirmed via a known number?
  □ Is the content of the request consistent with company policy?
  □ Was a code word requested?
  → If in doubt: REFUSE and inform your supervisor/IT!

AI-Generated Threat Landscape

Trends and Developments:

Current AI tools for deepfakes:
  ElevenLabs:     Voice cloning, 3 minutes of audio sufficient
  HeyGen:         Video avatars, lip-sync
  Synthesia:      Professional AI videos
  Midjourney:     Photo-realistic images
  RunwayML:       Real-time video deepfakes
  FaceSwap:       Open-source face swap

Costs for Attackers (March 2026):
  Voice Clone: ~$1/month (ElevenLabs Starter)
  Video Deepfake: ~$30/video (HeyGen)
  → ROI for Attackers: 1 successful BEC = 100,000 EUR → 1 EUR investment!

Future threats:
  → Real-time deepfakes: perfect face swap in video conferences (already happening today!)
  → LLM-powered targeting: automatic personalization of attacks
  → Deepfakes + biometric bypass: circumventing liveness tests
  → Synthetic identities on a large scale

C2PA as an industry standard:
  → Authentic media receive a cryptographic signature
  → Goal: "If no C2PA seal → could be a deepfake"
  → Not yet widespread (2026), but a growing trend
  → Adobe Content Credentials: Photoshop, Stock, Firefly

Regulatory developments:
  → EU AI Act: Mandatory deepfake labeling
  → Germany: Section 184k of the Criminal Code (non-consensual deepfakes) since 2021
  → USA: DEEPFAKES Accountability Act (under discussion)