2nd/10_Wiki/Topics/DevOps_and_Security/Media-Literacy.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-media-literacy	Media Literacy	10_Wiki/Topics	verified	self	Information Literacy Source Evaluation Digital Literacy	none	A	0.85	applied	media-literacy information verification deepfake security		2026-05-10	pending	language framework python c2pa

Media Literacy

매 한 줄

"매 source 의 verify, claim 의 cross-check, framing 의 detect — 매 information 의 evaluate skill". 매 1990s NAMLE 시작, 매 2026 LLM-generated content + deepfake + C2PA provenance + AI watermark 의 era 에 매 default skill.

매 핵심

매 Core skills (5)

• Access: 매 reliable source 의 find.
• Analyze: bias, framing, omission 의 detect.
• Evaluate: credibility, evidence quality.
• Create: ethical content production.
• Act: misinformation 의 counter.

매 SIFT method (Caulfield)

• Stop: 매 click 전 pause.
• Investigate: source 의 background.
• Find: better/original coverage.
• Trace: claim 의 original context.

매 응용

1. Deepfake detection: C2PA provenance + ML classifier.
2. LLM output: hallucination 의 detect.
3. News pipeline: source ranking.

💻 패턴

C2PA manifest verification

# 매 image 의 provenance 의 verify
from c2pa import Reader
reader = Reader.from_file('photo.jpg')
manifest = reader.json()
print(f"Producer: {manifest['active_manifest']['claim_generator']}")
print(f"AI generated: {manifest.get('ai_generated', False)}")
print(f"Signature valid: {reader.validation_status()}")

AI watermark detection (SynthID-like)

# 매 LLM output 매 watermark 의 detect
import torch
def detect_watermark(text: str, key: bytes, threshold=0.6) -> bool:
    tokens = tokenize(text)
    # green-list ratio (Kirchenbauer 2023)
    green = sum(1 for t in tokens if hash_token(t, key) % 2 == 0)
    z = (green - 0.5*len(tokens)) / (0.25*len(tokens))**0.5
    return z > threshold * 5  # 매 strict threshold

Reverse image search (TinEye API)

import httpx
def reverse_search(image_path: str, api_key: str):
    with open(image_path, 'rb') as f:
        r = httpx.post('https://api.tineye.com/rest/search/',
                       files={'image_upload': f},
                       auth=(api_key, ''))
    matches = r.json()['results']['matches']
    return [(m['image_url'], m['domain'], m['crawl_date']) for m in matches[:5]]

Source credibility score

TRUSTED_DOMAINS = {'reuters.com': 0.95, 'apnews.com': 0.93, 'nature.com': 0.97}
SUSPICIOUS = {'.tk', '.click'}

def score_source(url: str) -> float:
    from urllib.parse import urlparse
    domain = urlparse(url).netloc.lower().lstrip('www.')
    if domain in TRUSTED_DOMAINS: return TRUSTED_DOMAINS[domain]
    if any(domain.endswith(s) for s in SUSPICIOUS): return 0.1
    return 0.5  # unknown

Deepfake classifier (FaceForensics++)

import torch
from transformers import AutoModelForImageClassification, AutoImageProcessor

model = AutoModelForImageClassification.from_pretrained(
    'prithivMLmods/Deep-Fake-Detector-v2-Model')
proc = AutoImageProcessor.from_pretrained('prithivMLmods/Deep-Fake-Detector-v2-Model')

def is_deepfake(img) -> tuple[bool, float]:
    inputs = proc(images=img, return_tensors='pt')
    with torch.no_grad():
        logits = model(**inputs).logits
    probs = logits.softmax(-1)[0]
    fake_prob = probs[1].item()
    return fake_prob > 0.5, fake_prob

Cross-reference fact-check

import asyncio, httpx

async def fact_check(claim: str):
    async with httpx.AsyncClient() as c:
        r = await c.get('https://factchecktools.googleapis.com/v1alpha1/claims:search',
                        params={'query': claim, 'key': 'KEY'})
    results = r.json().get('claims', [])
    return [(x['text'], x['claimReview'][0]['textualRating']) for x in results]

Browser ext: provenance badge

// content.ts
async function annotateImages() {
  for (const img of document.querySelectorAll('img')) {
    const r = await fetch(`/api/c2pa-check?url=${encodeURIComponent(img.src)}`);
    const { aiGenerated, verified } = await r.json();
    if (aiGenerated) img.style.outline = '3px solid orange';
    if (!verified) img.title = '매 provenance unverified';
  }
}

매 결정 기준

상황	Approach
News article	SIFT method
Image authenticity	C2PA + reverse search + deepfake classifier
LLM output	watermark detect + cross-reference
Suspicious domain	credibility score < 0.3 → reject

기본값: SIFT + tooling-augmented (C2PA, fact-check API).

🔗 Graph

• 부모: Information Literacy
• 변형: Source Evaluation
• 응용: Deepfake-Detection
• Adjacent: Conversational-Maxims · Procedural-Rhetoric

🤖 LLM 활용

언제: claim cross-reference, framing analysis, summary 의 bias detect. 언제 X: 매 LLM 자체 매 hallucinate — 매 외부 source 와 cross-check 필수.

❌ 안티패턴

• Headline reading: 매 click 만 하고 article body 매 읽지 X.
• Single source: corroboration 매 X.
• Bothsidesism: 매 lopsided evidence 의 false equivalence.
• No provenance check: image 매 viral spread 후 reverse search X.

🧪 검증 / 중복

• Verified (NAMLE Core Principles, C2PA spec 2.0, SIFT method by Mike Caulfield).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — SIFT + C2PA + deepfake tooling