Deepfake
AI-generated synthetic media, most often produced with GANs, that convincingly fabricates a person’s face or voice and creates new risks for disinformation and identity fraud.
I. Overview
%%{init: { 'theme': 'base', 'themeVariables': { 'edgeLabelBackground': '#fff' }}}%%
flowchart LR
A["Source media\nOriginal video / image"] -- "GAN-based generative model" --> B["Synthesized content\nDeepfake output"]
style A fill:#f9f9f9,stroke:#333,stroke-width:3px
style B fill:#e1f5fe,stroke:#01579b,stroke-width:3px
Definition: Realistic fake video or audio content that uses deep learning — particularly GANs (Generative Adversarial Networks) — to synthesize or edit a specific person’s face or voice.
Features:
( Realistic manipulation ) Achieves a level of realism that is difficult to distinguish from genuine content, beyond what conventional image or video editing can produce.
( Malicious use ) Increasingly used to spread fake news, commit defamation, carry out financial fraud ( voice phishing ), and fuel political disinformation, raising the risk of social disruption.
( Identity theft ) A specific person’s voice or face can be misappropriated without authorization and abused in impersonation crimes.
II. Mechanism & Components
GAN (Generative Adversarial Network)-Based Generative Model
sequenceDiagram
participant G as Generator
participant D as Discriminator
participant T as Training Data
Note over G,D: The two networks learn by competing with each other
G->>D: Generates a fake image (from initial random noise)
D->>T: Learns from real images
D->>G: Feeds back the real/fake judgment
G->>G: Improves the generative model to fool the discriminator (gradient descent)
Note over G,D: Through repeated training, the generator produces increasingly realistic fake images,<br/>while the discriminator gets better at telling real from fake
Deepfake Generation Process
- Data collection: Gather a large volume of source data — the target person’s face, expressions, voice, and so on — for training the AI model.
- Model training: Use a GAN to learn the features of the source data and train a new model that synthesizes faces or voices.
- Content generation: Feed inputs (a target face or voice) into the trained model to generate manipulated deepfake video or audio.
- Post-processing (optional): Edit and refine the generated content so it appears more natural.
III. Adoption Considerations
| Risk | Mitigation |
|---|---|
| Fabricated content spread as disinformation or used in impersonation fraud | AI-based detection of visual/auditory artifacts (unnatural blinking, voice-alteration traces), watermarking at creation time, blockchain-based digital provenance |
| Voice/video impersonation used for social engineering (voice phishing, executive-impersonation fraud) | Out-of-band verification for high-value requests, not caller/sender identity alone |
| Unauthorized use of a person’s likeness or voice, and slow platform takedown | Stronger anti-abuse legislation, mandatory platform detection and takedown obligations, public awareness education |
The organizational fix for deepfake-enabled fraud looks a lot like the fix for email-based CEO fraud a decade ago: a mandatory out-of-band verification step for any high-value request, regardless of how convincing the voice or video sounds. Detection technology will keep losing an arms race against generation technology, so the process control — not detection accuracy — should carry the weight of the defense.
Last updated 18 Aug 2026, 00:00 UTC.