Scientific Integrity Research Platform

ARIA

Automated Research Integrity Analysis Platform

A unified AI-powered platform that integrates three independent forensic analysis systems to detect image manipulation, deepfakes, and paper mill activity in scientific publications — from pixel-level forgery maps to metadata anomaly detection.

Scroll
3 Integrated analysis systems
16 Forensic analysis modules
988 Papers in SILA benchmark
Multi-user Secure role-based platform

A Single Portal for Scientific Forensics

ARIA brings together three independent research-grade forensic pipelines under one multi-user web interface. Each system addresses a distinct aspect of research integrity: image manipulation within papers, deepfake detection in figures, and paper mill identification through metadata patterns.

Researchers upload their scientific articles or figures, select the appropriate analysis system and modules, and receive rich visual results — activation maps, provenance graphs, anomaly scores, and natural-language forensic reports — in a single, downloadable ZIP package.

Image Forensics

Pixel-level detection of copy-move, splicing and panel duplication artefacts in scientific figures

LLM Forensic Reports

Multimodal LLM analysis generating natural-language descriptions of detected manipulation artefacts

Provenance Graphs

Directed graph analysis tracing the origin and history of scientific figures across publications

Anomaly Detection

Machine learning detection of paper mill patterns in metadata, embeddings and citation graphs

Three Specialised Forensic Pipelines

Each system runs in its own containerised environment and can be executed independently. Select the system that matches your analysis goal.

System 1 — SILA

Scientific Image Forensic Analysis

Detects image fabrication, copy-move forgeries, figure duplication and provenance inconsistencies in academic PDFs. Based on the sciint pipeline (Moreira et al., Nature Sci. Rep. 2022).

  • PDF Figure Extraction
  • Panel Segmentation (TensorFlow)
  • Copy-Move Detection (SIFT + CRAFT)
  • Image Ranking (FAISS)
  • Provenance Graph Analysis
  • SIFT / LIFT Feature Analysis
PDF Input Images No GPU required
System 2 — SciForensics

Deepfake & Manipulation Detection

A three-stage pipeline combining deep-learning manipulation detection, multimodal LLM forensic reporting and pixel-level inconsistency mapping for scientific images.

  • Detect & Locate (XceptionFusion)
  • FakeScope — LLaVA 1.5 7B Report
  • FakeScope — mPLUG-Owl2 Report
  • Pixel Inconsistency Mapping (PIM)
Image Input GPU Recommended CUDA 12.4
System 3 — PaperMill

Paper Mill Detection

Detects anomalies in scientific papers indicative of paper mill activity by extracting metadata, generating transformer embeddings and applying unsupervised anomaly detection.

  • PDF Mining (MinerU / PDF-Extract-Kit)
  • Metadata CSV Consolidation
  • BERT/SciBERT/T5 Embeddings + t-SNE/UMAP
  • Train/Test Split (no-leakage strategy)
  • One-Class SVM + Isolation Forest
  • Comparative Graphics (12 types)
PDF Batch Input CPU only

How It Works

1
Request Access

Contact the administrator to request an account on the platform.

2
Upload & Configure

Upload your paper PDFs or images, select the forensic system, and choose analysis modules.

3
Run & Monitor

Start the pipeline with one click. Monitor real-time execution logs as each module completes.

4
Explore & Download

Browse visual results in the interface and download a complete ZIP package of all outputs.

Access Policy & Attribution

ARIA is a research platform operated at IFSP Campus Itapetininga in collaboration with RECOD/Unicamp. Access is restricted to researchers, students and institutional partners. To request an account, please contact the platform administrator.

The platform integrates three independent systems, each with its own academic attribution:

  • SILA — Moreira et al., Nature Scientific Reports 12, 18306 (2022)
  • SciForensics — Based on XceptionFusion, LLaVA 1.5, mPLUG-Owl2, and PIM
  • PaperMill — Internal pipeline using MinerU, BERT/SciBERT/T5, One-Class SVM and Isolation Forest

Hardware notice: The SciForensics pipeline (FakeScope stage) requires a GPU with ≥ 7 GB VRAM. Other systems run on CPU-only hardware.

Interested in Using ARIA?

To request access or discuss research collaborations, please contact the platform administrator directly by email. Include a brief description of your research context and intended use of the platform.

Carlos Santos
Researcher · IFSP Campus Itapetininga & RECOD / Unicamp
carlos.santos@ifsp.edu.br