Skip to content

3.5 Deepfakes — The Evolving Threat

The Arms Race

graph TD
    A["2017: Basic<br>face swaps<br>(minutes to create,<br>easy to detect)"] --> B["2019: DeepFaceLab<br>(hours, moderate<br>detection)"]
    B --> C["2021: Real-time<br>face swap<br>(instant, harder<br>to detect)"]
    C --> D["2023: LivePortrait<br>One-shot animation<br>(seconds from single<br>photo, very hard)"]
    D --> E["2025+: Neural<br>avatars, 3D Gaussians<br>(real-time, near<br>indistinguishable)"]

Why Deepfakes Are Uniquely Dangerous for Banking

  1. Low barrier to entry: Free tools, consumer GPUs, tutorial videos
  2. Real-time capability: Can respond to active liveness challenges live
  3. Scalability: One model can generate unlimited attack attempts
  4. Improving faster than detection: Generative AI advances outpace defensive AI
  5. Cross-modal: Can combine face + voice for multi-modal bypass

Detection Signals

Signal What To Look For Reliability
Temporal flickering Face boundary flickers between frames Good for lower-quality deepfakes; fading for state-of-the-art
Blending boundary Visible seam where swapped face meets original Good — most swaps still show subtle boundaries
Frequency artifacts GAN upsampling creates checkerboard patterns in FFT Good but increasingly addressed by newer models
Eye reflection inconsistency Reflections in left vs right eye should match; deepfakes often don't Moderate — improving in newer methods
Physiological absence (rPPG) No detectable heart rate signal in synthetic face Excellent — extremely hard to fake
Lip sync quality Subtle timing/shape mismatches between audio and mouth Moderate — tools like Wav2Lip are very good
Hair/ear/neck artifacts Deepfakes struggle with fine hair detail, ear consistency, neck blending Good supplementary signal
Background consistency Face processing may leave background untouched, creating lighting/color mismatch Moderate

Anti-Deepfake Architecture

graph TD
    A["Input Video<br>(N frames)"] --> B["Frame-level<br>Deepfake Detector<br>(per-frame score)"]
    A --> C["Temporal<br>Consistency<br>Analyzer"]
    A --> D["rPPG Extractor<br>(physiological<br>signal)"]
    A --> E["Forensic<br>Frequency<br>Analyzer"]

    B --> F["Ensemble<br>Fusion"]
    C --> F
    D --> F
    E --> F

    F --> G{"Deepfake<br>Probability"}
    G -->|"> 0.7"| H["❌ Deepfake<br>Detected"]
    G -->|"0.3 - 0.7"| I["⚠️ Uncertain<br>Escalate"]
    G -->|"< 0.3"| J["✅ Likely<br>Genuine"]
  • FaceForensics++ (Rössler et al., 2019) — benchmark dataset
  • Thinking in Frequency (Qian et al., 2020) — frequency analysis approach
  • Multi-Attentional Deepfake Detection (Zhao et al., 2021) — attention-based
  • DeepfakeBench (Yan et al., 2023) — comprehensive benchmark
  • Implicit Identity Leakage (Dong et al., 2023) — identity-based detection

Next: Social Engineering & Process Attacks →