2nd/10_Wiki/Topics/AI_and_ML/Deepfake-Technology.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-deepfake	Deepfake Technology	10_Wiki/Topics	verified	self	deepfake face swap voice cloning synthetic media FaceForensics C2PA ElevenLabs	none	B	0.85	applied	deepfake generative-ai face-swap voice-cloning synthetic-media c2pa detection ethics		2026-05-10	pending	language framework Python Diffusers / Roop / Stable Diffusion / ElevenLabs / Whisper

Deepfake Technology

매 한 줄

"매 truth 의 boundary 의 erode". 매 GAN / Diffusion / Autoencoder 의 face + voice 의 synthesize. 매 commercial application + 매 election / fraud / 비동의 abuse 의 dual-use. 매 detection arms race + 매 C2PA provenance 의 standard.

매 핵심 technique

Face swap

• Roop / DeepFaceLab: 매 open-source.
• Autoencoder-based: 매 encode → 매 decode 의 다른 face.
• InstantID / PhotoMaker: 매 single image.
• Diffusion-based: 매 InstantID + ControlNet.

Face reenactment

• 매 source 의 expression → 매 target.
• First Order Motion Model.
• DPE (Disentangled Portrait Editing).

Voice cloning

• ElevenLabs: 매 commercial.
• OpenVoice (MyShell): 매 open.
• Tortoise TTS.
• 3 sec sample 의 sufficient (modern).

Lip sync

• Wav2Lip: 매 audio + face.
• SadTalker.

Full body / pose

• AnimateAnyone.
• MagicAnimate.

Video generation (modern, 2024+)

• Sora (OpenAI).
• Veo (Google).
• Runway Gen-3.

매 detection

• FaceForensics++: 매 dataset benchmark.
• CLIP-based: 매 zero-shot.
• Frequency domain.
• Inconsistency (lighting, eye blink rate).
• Liveness check (camera, depth).

매 disclosure / provenance

• C2PA (Adobe + others): 매 cryptographic chain.
• SynthID (Google): 매 watermark.
• Statistical watermark (LLM).

매 legal landscape

• EU AI Act (2024): 매 disclosure required.
• TAKE IT DOWN Act (US 2025): 매 NCII 의 takedown.
• California: 매 election deepfake 의 ban.
• Korea: 매 형법 244-2 의 sexual deepfake 의 처벌.
• Civil: 매 right of publicity, defamation.

매 dual-use

Positive	Negative
Film (de-aging)	Election interference
Education (historical figure)	NCII (non-consensual intimate imagery)
Accessibility (sign language)	Identity theft
Game / VR	Fraud (CEO voice scam)
Localization (lip sync)	Deepfake harassment

매 mitigation strategy

1. Training data filter: 매 NCII / illegal 의 prevent.
2. Watermarking (Glaze, Nightshade, SynthID).
3. Disclosure mandate.
4. Detection at platform.
5. Liveness for high-stakes auth.
6. Provenance (C2PA chain).
7. Legal recourse.

💻 패턴 (응용 — defense + ethical use)

Liveness check (anti-deepfake auth)

def liveness_check(video_stream):
    """매 camera challenge: 매 head movement + blink + utterance."""
    # 매 random challenge
    challenge = random.choice(['blink twice', 'turn head left', 'say YES'])
    show_to_user(challenge)
    
    response = capture_response(video_stream, duration=3)
    
    return {
        'blink_detected': detect_eye_blink(response),
        'head_movement': detect_head_motion(response),
        'utterance_match': verify_speech(response, expected=challenge),
        'depth_check': detect_depth_inconsistency(response),  # 매 2D photo 의 detect
    }

Deepfake detection (CLIP-based)

import open_clip
import torch

model, _, preprocess = open_clip.create_model_and_transforms('ViT-L-14', pretrained='openai')

def detect_deepfake(image_path, threshold=0.6):
    image = preprocess(Image.open(image_path)).unsqueeze(0)
    
    candidates = ['a real photo of a person', 'an AI-generated synthetic face']
    text_emb = model.encode_text(open_clip.tokenize(candidates))
    img_emb = model.encode_image(image)
    
    sim = (100 * img_emb @ text_emb.T).softmax(-1)
    return {
        'real_score': sim[0, 0].item(),
        'synthetic_score': sim[0, 1].item(),
        'is_deepfake': sim[0, 1].item() > threshold,
    }

Frequency-domain detection

import numpy as np
from scipy.fft import fft2, fftshift

def fft_anomaly_score(image):
    """매 GAN 의 typical 의 frequency artifact."""
    gray = np.mean(image, axis=-1)
    spectrum = np.abs(fftshift(fft2(gray)))
    
    # 매 high-frequency 의 GAN typical
    high_freq_energy = spectrum[image.shape[0]//4:].mean()
    return high_freq_energy

C2PA verification

from c2pa import C2pa

def verify_c2pa(image_path):
    c2pa = C2pa()
    try:
        manifest = c2pa.read_manifest(image_path)
        return {
            'has_provenance': True,
            'chain': manifest.actions,
            'signature_valid': c2pa.verify_signature(manifest),
            'creator': manifest.author,
            'tools_used': manifest.softwareAgents,
        }
    except Exception:
        return {'has_provenance': False}

SynthID-style watermark detection

def detect_watermark(image, watermark_key):
    """매 invisible statistical watermark."""
    expected_pattern = generate_pattern(watermark_key)
    actual_pattern = extract_low_freq_signal(image)
    correlation = np.corrcoef(expected_pattern.flatten(), actual_pattern.flatten())[0, 1]
    return correlation > 0.7  # 매 threshold

Glaze / Nightshade (artist protection)

def glaze_protect(artist_image, target_style='abstract', epsilon=0.05):
    """매 ML 의 train 의 disrupt — 매 imperceptible perturbation."""
    perturbed = artist_image.clone().requires_grad_()
    
    for _ in range(100):
        # 매 push to wrong style space
        loss = -style_distance(perturbed, target_style)
        loss.backward()
        perturbed.data -= 0.001 * perturbed.grad.sign()
        perturbed.data = torch.clamp(perturbed, artist_image - epsilon, artist_image + epsilon)
    
    return perturbed.detach()

Ethical use validation (commercial)

def commercial_deepfake_check(generation_request):
    """매 commercial use 의 consent + license check."""
    issues = []
    
    if not generation_request.has_consent_signed:
        issues.append('Missing consent from likeness owner')
    
    if generation_request.purpose == 'fake_attribution':
        issues.append('Cannot fabricate attribution / quotation')
    
    if generation_request.target_minor:
        issues.append('Minor — special protection required')
    
    if generation_request.election_period and not generation_request.disclosure:
        issues.append('Election period — disclosure required')
    
    return {'allowed': len(issues) == 0, 'issues': issues}

NCII detection (incoming user upload)

def detect_ncii_attempt(image, source_user):
    """매 nudity + 매 face match 의 다른 person → 매 likely NCII."""
    if not contains_nudity(image): return None
    
    detected_faces = face_recognize(image)
    user_face = source_user.profile_face
    
    for face in detected_faces:
        if not similar(face, user_face):
            return {
                'risk': 'high',
                'reason': 'nudity + non-self face',
                'action': 'block + report',
            }
    return None

매 결정 기준

응용	Approach
Film / VFX	Consent + C2PA + disclosure
Education	Historical figure + clear context
Accessibility	Sign language synthesis
Auth / KYC	Liveness check + 3D depth
Content moderation	Detection + reporting
Artist protection	Glaze / Nightshade
Commercial likeness	Contract + consent
Election	Detection + takedown

기본값: 매 disclosure + 매 consent + 매 detection + 매 watermark.

🔗 Graph

• 부모: Generative-AI · Computer Vision · AI-Ethics
• 변형: Face-Swap · Voice-Cloning
• 응용: ElevenLabs
• Mitigation: C2PA
• Adjacent: Authenticity · Arts · Algorithmic Fairness · Anthropomorphism · AI Safety

🤖 LLM 활용

언제: 매 deepfake risk assessment. 매 detection system. 매 disclosure policy. 매 ethical use review. 언제 X: 매 manipulative use (election, NCII, fraud).

❌ 안티패턴

• No consent: 매 personality right violation.
• Election deepfake without disclosure: 매 illegal (some jurisdiction).
• No watermark: 매 trust 의 long-term destroy.
• Detection only (no provenance): 매 false negative.
• Commercial without contract: 매 lawsuit.
• NCII: 매 criminal.

🧪 검증 / 중복

• Verified (FaceForensics++, C2PA spec, EU AI Act 2024, TAKE IT DOWN Act 2025).
• 신뢰도 B.
• Related: Authenticity · Arts · Anthropomorphism · AI Safety · Brand Consistency Maintenance · Commercial AI Art Production.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — technique + legal + 매 liveness / detection / C2PA / Glaze / NCII code