Xabarnavis AI Forensic Platform

Xabarnavis is a multimodal media-authenticity and forensic analysis platform for image, audio, video, and text evidence. It combines model outputs, metadata, signal-level indicators, explainable score fusion, evidence hashing, and professional report generation in one web dashboard.

O'zbekcha: Xabarnavis — rasm, audio, video va matn dalillarini tekshiruvchi ko'p modelli media-forensika platformasi. Tizim AI modellari, metadata, signal belgilari, izohlanadigan baholash, SHA-256 dalil identifikatsiyasi va professional hisobotlarni yagona dashboard'da birlashtiradi.

Navigation

• Overview / Umumiy ko'rinish
• Screenshots / Interfeys
• Mathematical model / Matematik model
• Features / Imkoniyatlar
• Architecture / Arxitektura
• Quick start / Ishga tushirish
• Repository rules

Overview / Umumiy ko'rinish

Xabarnavis associates every analysis with an authenticated user and produces traceable evidence artifacts. The current pipeline supports forensic visualizations, per-model results, fused scores, evidence hashes, QR-linked report pages, and JSON/DOCX exports.

O'zbekcha: Har bir tahlil foydalanuvchi hisobiga bog'lanadi. Natijaga model javoblari, vizual dalillar, umumlashtirilgan baholar, dalil hashi, QR orqali ochiladigan sahifa va JSON/DOCX hisobotlar kiradi.

Important

Xabarnavis scores are decision-support indicators, not proof of authorship and not calibrated legal certainty. Results must be reviewed together with source provenance, metadata, visual evidence, and domain-expert judgment.

Muhim: Xabarnavis baholari mualliflikning mutlaq isboti yoki kalibrlangan huquqiy aniqlik emas. Natija fayl manbasi, metadata, vizual dalillar va mutaxassis xulosasi bilan birga baholanishi kerak.

Screenshots / Interfeys

Audio forensic report / Audio-forensika hisoboti

The audio workspace presents model confidence, waveform evidence, segment-level labels, SHA-256 chain-of-custody data, and a QR-linked report.

O'zbekcha: Audio sahifasi model ishonchini, waveform dalilini, segment belgilarini, SHA-256 dalil zanjirini va QR orqali ochiladigan hisobotni ko'rsatadi.

Image forensic report / Rasm-forensika hisoboti

The image report compares the original evidence, a signal overlay, ELA residuals, and a forensic heatmap. It also records anomaly regions and exports the result as JSON or a legal-style DOCX report.

O'zbekcha: Rasm hisoboti asl dalilni signal overlay, ELA qoldig'i va forensic heatmap bilan taqqoslaydi. Anomaliya hududlari qayd etilib, natija JSON yoki rasmiy uslubdagi DOCX shaklida eksport qilinadi.

Mathematical model / Matematik model

The current MVP uses an explainable weighted fusion of normalized indicators. For any value $x$, the clamp operation keeps the score in the interval $[0,1]$:

$$ \operatorname{clamp}(x)=\min(1,\max(0,x)) $$

AI-generation score

$$ A=\operatorname{clamp}(0.45f+0.25t+0.20g+0.10m) $$

Manipulation score

$$ M=\operatorname{clamp}(0.30e+0.25n+0.15j+0.15l+0.15d+0.05m) $$

Real-camera score

$$ R=\operatorname{clamp}(1-\max(A,M)+0.15c) $$

The selected class is the largest of the three scores:

$$ \hat{y}=\operatorname*{arg,max}_{k\in{R,A,M}} k $$

If $s=\max(R,A,M)$, confidence is assigned using the implemented thresholds:

$$ \operatorname{confidence}(s)= \begin{cases} \text{High}, & s\ge 0.75\\ \text{Medium}, & 0.55\le s<0.75\\ \text{Low}, & s<0.55 \end{cases} $$

Symbol	Indicator	O'zbekcha izoh
$f$	Frequency-domain anomaly	Chastota sohasidagi anomaliya
$t$	Texture uniformity	Teksturaning noodatiy bir xilligi
$g$	Generator-software metadata	AI generator dasturi metadata belgisi
$m$	General metadata anomaly	Umumiy metadata nomuvofiqligi
$e$	Edge inconsistency	Qirralardagi nomuvofiqlik
$n$	Noise inconsistency	Shovqin profilidagi nomuvofiqlik
$j$	JPEG blocking	JPEG blok artefaktlari
$l$	ELA anomaly	Error Level Analysis qoldig'i
$d$	Editor-software metadata	Tahrirlash dasturi metadata belgisi
$c$	Camera provenance	Kamera kelib chiqishi belgisi

These equations document the current explainable MVP fusion in fusion.py. Individual model adapters may also provide their own probabilities and evidence; the values above should be interpreted as heuristic forensic indicators rather than calibrated probabilities.

O'zbekcha: Formulalar joriy MVP fusion mantiqini ifodalaydi. Alohida modellar qo'shimcha ehtimol va dalillar qaytarishi mumkin; yuqoridagi qiymatlar kalibrlangan ehtimol emas, balki tushuntiriladigan forensik indikatorlardir.

Features / Imkoniyatlar

• Image upload, metadata inspection, ELA, residual heatmaps, and anomaly regions
• Audio waveform, anti-spoof/deepfake adapters, and segment evidence
• Video frame/audio extraction and multi-model research adapters
• Text evidence intake and structured case handling
• Local and external model registry with per-model status
• Authenticated user sessions, device history, and private case archives
• SHA-256 evidence identification and chain-of-custody fields
• JSON and DOCX report generation with QR/public report pages
• PostgreSQL production target with SQLite local-development fallback

Architecture / Arxitektura

User
  -> Next.js web application
  -> FastAPI forensic API
  -> Media preprocessing and signal extraction
  -> Local/external AI model adapters
  -> Explainable score fusion
  -> SQLite (local) or PostgreSQL + Redis (production target)
  -> JSON/DOCX report + QR/public report page

apps/       FastAPI API and Next.js web application
ml/         Model registry and model-family metadata
data/       Local datasets; excluded from Git
artifacts/  Model weights, runs, reports, and logs; excluded from Git
storage/    Runtime uploads, reports, profiles, and local database
infra/      Docker, nginx, PostgreSQL, and systemd deployment files
scripts/    Dataset, training, evaluation, and setup utilities
docs/       Architecture, datasets, security, training, and deployment guides

Further reading:

• Architecture
• API
• Datasets
• Model benchmarks
• Training guide
• Deployment

Quick Start / Ishga tushirish

Full website

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
python manage.py runserver

Open:

• Website: http://127.0.0.1:8000
• API health: http://127.0.0.1:8001/health
• Swagger UI: http://127.0.0.1:8001/docs

Stop both development servers:

python manage.py stopserver

Backend only

cd apps\api
python -m uvicorn app.main:app --reload --port 8001

Frontend only

cd apps\web
pnpm install
pnpm dev -- --hostname 127.0.0.1 --port 8000

Repository Rules

Git contains source code, documentation, manifests, and lightweight metadata only. Do not commit generated or private runtime assets:

• apps/web/node_modules/ and apps/web/.next/
• data/datasets/, data/raw/, data/processed/, data/ready/, and data/holdout/
• artifacts/models/, artifacts/runs/, artifacts/reports/, and artifacts/logs/
• storage/uploads/, storage/reports/, storage/profiles/, and local databases
• .env, private keys, credentials, and access tokens

Track datasets and models through lightweight metadata and reproducible scripts:

• docs/datasets.md
• docs/model_benchmarks.md
• ml/registry/models_registry.json
• storage/dataset_inventory.json
• scripts/datasets/

Forensic reporting principle

Avoid unsupported absolute claims such as "100% fake." Prefer probability and confidence language tied to the available evidence:

The image is highly likely to be AI-generated. This conclusion is based on visual, frequency-domain, metadata, and model-based forensic indicators.

O'zbekcha: Mutlaq hukm o'rniga mavjud vizual, chastotaviy, metadata va model dalillariga tayangan ehtimollik hamda ishonch tilidan foydalaning.