Multi-Modal Plagiarism Detection Framework

Multi-Modal Similarity Analysis Across Code, PDFs & Visuals

An automated cross-artifact similarity analysis framework uniting Sentence-BERT semantic vectors, CodeBERT AST embeddings, and CLIP visual patch heatmaps into a shared metric space. Detect unauthorized copying, semantic paraphrasing, and structural code refactoring without manual supervision.

Select or hover any modality node to see how inputs project into the joint manifold and compute similarity in real time.

Launch Ingestion
Sentence-BERTCodeBERT ASTCLIP ViT-B/32Patch HeatmapJoint Manifold
Source CodeCodeBERT AST
Text / PDFSentence-BERT
Heat Map4×4 Patch Grid
Visual GraphicCLIP ViT-B/32
Cosine NormSimilarity cos(θ)
CodeBERT AST82.5% Clone
def shortest_path(g):Ref
def find_min_route(a):Clone
AST Tree Structural Invariance Matched
Source Code · 82.5% Clone Flagged
The Core Scientific Problem

Why Traditional Plagiarism Detectors Fail

Standard industry tools were engineered in the 2000s for static word-overlap. Modern university assignments involve source code, architecture diagrams, and AI-assisted paraphrasing that evade legacy algorithms.

Traditional Detectors

Lexical N-Gram & Exact String Matching

🔴 Vulnerable to Variable Renaming

Renaming variables (e.g., curr_node to frontier_item) destroys character n-grams, dropping match confidence from 100% to under 15%.

🔴 Blind to Diagrams, Figures & Charts

Completely ignores image files, flowcharts, vector graphs, and screenshots pasted into submitted academic papers and documents.

🔴 Bypassed by Synonym Substitution

Replacing every third word with a synonym prevents string hashing collision algorithms from detecting the underlying conceptual copy.

🔴 Zero Cross-Modal Correlation

Cannot identify that an algorithmic implementation in Python was directly plagiarized from a paper’s pseudocode or architectural diagram.

OriginaX AI Multi-Modal Engine

AST Invariance & 768-D Joint Embedding Space

🔹 Abstract Syntax Tree (AST) Invariance

Parses Python, C++, Java, and JS into abstract syntax trees. Renaming variables, rearranging helper functions, or swapping loops produces identical AST hashes (98.4% match).

🔹 Vision Transformer (ViT) Patch Matcher

Slices diagrams into 16x16 feature patches to detect rotated, re-colored, cropped, or slightly edited architectural diagrams and flowcharts.

🔹 Deep Contextual NLP Representation

Projects entire paragraphs into high-dimensional semantic vectors. Retains high similarity scores even when text is rewritten using AI paraphrasing tools.

🔹 Unified Cross-Modal Correlation

Projects code implementations and textual descriptions into a shared coordinate space, detecting cross-format derivation automatically.

Feature-by-Feature Engineering Comparison:

Detection Modality Scope
Single-Modal (Text)5 Heterogeneous Modalities
Legacy Detectors:

Plain text strings and ASCII characters only. Completely blind to images, diagrams, or source code ASTs.

OriginaX AI:

Unified Multi-Modal Space: Ingests raw Text, Source Code (ASTs), Vector Diagrams, Flowcharts, and Mathematical proofs.

Paraphrase & Synonym Substitution
Brittle Lexical MatchDeep Semantic Invariance
Source Code Plagiarism
0% AST Structural AwarenessAST Graph Isomorphism
Visual Figures & Architecture Diagrams
Blind to Visual ArtifactsViT Patch Heatmap Analysis
Cross-Modal Congruence
UnsupportedJoint-Space Projection
Explainability & Audit Trail
Basic String DiffStructural Explainability
The 5-Stage Scientific Pipeline

How OriginaX Unifies Multi-Modal Vectors

Through unified multi-modal alignment, artifacts of any type are tokenized, normalized, and projected onto a shared hypersphere manifold.

01

Multi-Modal Ingestion

Accepts raw PDFs, Python/C++ source code, image diagrams (PNG/JPG), and mathematical equations.

Direct Upload & Paste
02

AST & ViT Preprocessing

Generates Abstract Syntax Trees for code; extracts layout OCR from PDFs; slices diagrams into 16x16 patches.

Structural Normalization
03

Joint-Space Projection

Embeddings mapped to 768-dimensional coordinates where semantic proximity reflects conceptual overlap.

768-D Shared Manifold
04

Cosine & AST Scoring

Weighted formula: 40% Text, 35% AST Code, 25% Visual Graphic. Strict bipartite Hungarian assignment.

Calibrated Scoring Formula
05

Unified Report & Diff

Produces interactive visual diffs, syntax node alignment, and audit-ready PDF reports for evaluators and institutions.

Audit-Ready Evidence
Empirical Academic Evaluation

Validated Performance Benchmarks

Tested against benchmark datasets containing over 1,200 cross-modality student submission pairs.

98.4%
Code Obfuscation Accuracy
Resistant to renaming & control transposition
96.2%
Academic Paraphrase Recall
Evaluated on multi-sentence research summaries and abstracts
94.7%
Diagram & Flowchart Match
ViT patch overlap on altered graphics
180ms
Average Inference Latency
Sub-second matrix generation
Transparent Academic & Institutional Pricing

Flexible Subscription Plans

Start with our 10-check free sandbox, upgrade to full multi-modal Pro, or equip your department with collaborative team seats and enterprise custom integrations.

Starter1 Seats

Free Plan

10 checks free

$0/forever
What's Included:
  • 10 Free Checks Included
  • Natural Text & PDF Document Extraction
  • Basic Source Code AST Analysis
  • Local Zero-Knowledge Privacy Engine
  • Instant Overlap Detection
Most Popular1 Seats

Pro Plan

500 scans/mo

$20/month
What's Included:
  • $20 / Month Active Access
  • 500 Multi-Modal Scans Monthly
  • Full Code AST Identifier Anonymizer
  • Deep Semantic Vector Projections
  • Synchronized Split/Unified Diff Inspector
  • Priority AI Inference Gateway
Collaboration5 Seats

Team Plan

2,500 scans/mo · 5 Seats

$60/month
What's Included:
  • 5 Team Seats Included
  • Cohort N x N Collusion Heatmap Matrix
  • Cross-Assignment Batch Comparison
  • Shared Team Folders & Assignment Pools
  • Team Activity Logs & Exportable Audit Trail
  • Expedited Priority Queue
InstitutionalUnlimited

Enterprise Plan

Unlimited scans & custom seats

Custom
What's Included:
  • Custom Seat & Scan Volume
  • Direct Contact: contact@originax.online
  • LMS Integration (Canvas, Moodle, Blackboard)
  • Dedicated On-Premise Vectorizer Appliance
  • Zero-Knowledge FERPA/HIPAA Compliance SLA
  • Dedicated Account Executive & 24/7 SLA
Need a custom institutional deployment for your university?

We provide custom on-premise vectorizers, Canvas/Moodle LTI integration, and departmental volume billing.

Email: contact@originax.online

Ready to Analyze Submissions with OriginaX?

Sign in with your institutional credentials or launch the researcher workspace to run single-pair AST comparisons and N×N classroom cross-examinations.

Open User Dashboard