Spaces:
Sleeping
Sleeping
File size: 32,194 Bytes
be5f1db 380fdf3 be5f1db 233a692 1367cc2 50cb337 ee37188 50cb337 ee37188 233a692 ee37188 9bfa211 380fdf3 9bfa211 380fdf3 3c1f41b 4466506 9bfa211 3c1f41b 9bfa211 ee37188 1367cc2 ee37188 380fdf3 ee37188 380fdf3 ee37188 380fdf3 18433b8 1367cc2 ee37188 1367cc2 ee37188 380fdf3 ee37188 380fdf3 ee37188 d2ce864 40aa67d 4466506 ee37188 1367cc2 380fdf3 1367cc2 ee37188 be5f1db 4466506 be5f1db 380fdf3 be5f1db ee37188 fd22ff8 fa509a2 fd22ff8 fa509a2 fd22ff8 1367cc2 ee37188 1367cc2 fa509a2 1367cc2 fa509a2 1367cc2 ee37188 4466506 ee37188 380fdf3 ee37188 1367cc2 d9f59d1 ee37188 1367cc2 ee37188 1367cc2 d9f59d1 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 d9f59d1 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 d9f59d1 ee37188 1367cc2 d9f59d1 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 d9f59d1 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 d9f59d1 ee37188 1367cc2 d9f59d1 ee37188 d9f59d1 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 380fdf3 1367cc2 380fdf3 1367cc2 ee37188 1367cc2 ee37188 380fdf3 ee37188 1367cc2 380fdf3 1367cc2 4466506 1367cc2 4466506 1367cc2 4466506 1367cc2 4466506 1367cc2 ee37188 1367cc2 ee37188 1367cc2 380fdf3 1367cc2 ee37188 1367cc2 380fdf3 1367cc2 380fdf3 1367cc2 ee37188 1367cc2 d9f59d1 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 380fdf3 1367cc2 380fdf3 1367cc2 ee37188 3c1f41b ee37188 1367cc2 380fdf3 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 380fdf3 ee37188 380fdf3 ee37188 380fdf3 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 380fdf3 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 380fdf3 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 18433b8 ee37188 14142b1 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 1367cc2 ee37188 233a692 ee37188 1367cc2 ee37188 14142b1 ee37188 233a692 ee37188 1367cc2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 | ---
title: TEXT-AUTH — Evidence-Based Text Forensics System
emoji: 🔍
colorFrom: blue
colorTo: purple
sdk: docker
sdk_version: "4.36.0"
app_file: text_auth_app.py
pinned: false
license: mit
---
<div align="center">
# 🛡️ TEXT-AUTH: Evidence-First Text Forensics & Authenticity Assessment












</div>
---
## 📋 Table of Contents
- [Abstract](#abstract)
- [Overview](#overview)
- [Key Differentiators](#key-differentiators)
- [System Architecture](#system-architecture)
- [Workflow / Data Flow](#workflow--data-flow)
- [Forensic Signals & Mathematical Foundation](#forensic-signals--mathematical-foundation)
- [Ensemble Methodology](#ensemble-methodology)
- [Domain-Aware Analysis](#domain-aware-analysis)
- [Performance Characteristics](#performance-characteristics)
- [Evaluation & Benchmarks](#-evaluation--benchmarks)
- [Project Structure](#project-structure)
- [API Endpoints](#api-endpoints)
- [Installation & Setup](#installation--setup)
- [Model Management & First-Run Behavior](#model-management--first-run-behavior)
- [Frontend Features](#frontend-features)
- [Business Model & Market Analysis](#business-model--market-analysis)
- [Research Impact & Future Scope](#research-impact--future-scope)
- [Infrastructure & Deployment](#infrastructure--deployment)
- [Security & Risk Mitigation](#security--risk-mitigation)
- [Continuous Improvement Pipeline](#continuous-improvement-pipeline)
- [License & Acknowledgments](#license--acknowledgments)
---
## 📝 Abstract
**TEXT-AUTH** is a research-oriented, production-minded **text forensics system** that evaluates written content using multiple independent linguistic, statistical, and semantic signals.
Rather than claiming authorship or identifying a generation source, the platform performs **evidence-based probabilistic assessment** of textual consistency patterns. It reports confidence-calibrated signals, uncertainty estimates, and human-interpretable explanations to support downstream decision-making.
TEXT-AUTH is designed as a **decision-support and forensic analysis tool**, not a binary classifier or attribution oracle.
- *For Architectural details, see [Architecture](docs/ARCHITECTURE.md).*
- *For detailed technical documentation, see [Technical Docs](docs/BLOGPOST.md).*
- *For research methodology, see [Whitepaper](docs/WHITE_PAPER.md).*
- *For API documentation, see [API Documentation](docs/API_DOCUMENTATION.md).*
---
## 🚀 Overview
**Problem.** Modern text—whether human-written, assisted, edited, or fully generated—often exhibits patterns that are difficult to evaluate using binary classifiers.
**Solution.** A domain-aware analysis system combining six orthogonal evidence signals (Perplexity, Entropy, Structural, Semantic, Linguistic, Multi-perturbation stability) analysis into a confidence‑calibrated ensemble. Outputs are explainable with sentence‑level highlighting, and downloadable reports (JSON/PDF).
**Live Deployment Link:** [AI Text Authenticator Platform](https://huggingface.co/spaces/satyaki-mitra/AI_Text_Authenticator)
**MVP Scope.** End‑to‑end FastAPI backend, lightweight HTML UI, modular metrics, Hugging Face model auto‑download, and a prototype ensemble forensic analyzer. Model weights are not committed to the repo; they are fetched at first run.
---
## 🎯 Key Differentiators
| Feature | Description | Impact |
|---|---:|---|
| **Domain‑Aware Detection** | Calibrated thresholds and metric weights for 16 content types (Academic, Technical, Creative, Social Media, etc.) | Improved signal calibration and reduced false positives compared to generic binary systems |
| **6-Signal Evidence Ensemble** | Orthogonal statistical, syntactic, and semantic indicators | Robust assessments with reduced false positives |
| **Explainability** | Sentence‑level scoring, highlights, and human‑readable reasoning | Trust & auditability |
| **Auto Model Fetch** | First‑run download from Hugging Face, local cache, offline fallback | Lightweight repo & reproducible runs |
| **Extensible Design** | Plug‑in metrics, model registry, and retraining pipeline hooks | Easy research iteration |
### 📊 Supported Domains & Threshold Configuration
The platform supports domain-aware forensic analysis tailored to the following 16 domains, each with specific synthetic-text consistency thresholds and metric weights defined in `config/threshold_config.py`. These configurations are used by the ensemble forensic analyzer to adapt its evidence aggregation process.
**Domains:**
* `general` (Default fallback)
* `academic`
* `creative`
* `ai_ml`
* `software_dev`
* `technical_doc`
* `engineering`
* `science`
* `business`
* `legal`
* `medical`
* `journalism`
* `marketing`
* `social_media`
* `blog_personal`
* `tutorial`
**Threshold Configuration Details (`config/threshold_config.py`):**
Each domain is configured with specific thresholds for the six detection metrics and an ensemble threshold. The weights determine the relative importance of each metric's output during the ensemble aggregation phase.
* **High-Consistency Threshold:** If a metric's synthetic-consistency score exceeds this value, it contributes stronger evidence toward a synthetic-consistency assessment for that metric.
* **Low-Consistency Threshold:** If a metric's Authentic probability falls below this value, it contributes evidence toward higher human-authored consistency for that metric.
* **Weight:** The relative weight assigned to the metric's result during ensemble combination (normalized internally to sum to 1.0 for active metrics).
### Confidence-Calibrated Aggregation (High Level)
1. Start with domain-specific base weights (defined in `config/threshold_config.py`).
2. Adjust these weights dynamically based on each metric's individual confidence score using a scaling function.
3. Normalize the adjusted weights.
4. Compute the final weighted aggregate probability.
---
## 🏗️ System Architecture
```mermaid
%%{init: {
"theme": "dark",
"themeVariables": {
"fontSize": "10px",
"fontFamily": "Segoe UI, Helvetica, Arial, sans-serif"
}
}}%%
flowchart TD
classDef frontend fill:#4CAF50,stroke:#2E7D32,color:white;
classDef api fill:#2196F3,stroke:#0D47A1,color:white;
classDef orchestrator fill:#FF9800,stroke:#E65100,color:white;
classDef metrics fill:#9C27B0,stroke:#4A148C,color:white;
classDef core fill:#607D8B,stroke:#263238,color:white;
classDef storage fill:#795548,stroke:#3E2723,color:white;
A[Web UI<br/>📄 Upload & Input]:::frontend
B[Dashboard<br/>📊 Live Results]:::frontend
C[FastAPI<br/>]:::api
D[Domain Classifier]:::orchestrator
E[Preprocessor]:::orchestrator
F[Metric Coordinator]:::orchestrator
P1[Perplexity]:::metrics
P2[Entropy]:::metrics
P3[Structural]:::metrics
P4[Linguistic]:::metrics
P5[Semantic]:::metrics
P6[Stability]:::metrics
G[Evidence Aggregator<br/>⚖️ Ensemble + Calibration]:::core
H[Reporter<br/>📝 Highlights • PDF/JSON]:::core
I["Models<br/>🤗 HF Cache"]:::storage
J[(Storage<br/>💾 Logs • Reports)]:::storage
A --> C
B --> C
C --> D
D --> E
E --> F
F --> P1
F --> P2
F --> P3
F --> P4
F --> P5
F --> P6
P1 --> G
P2 --> G
P3 --> G
P4 --> G
P5 --> G
P6 --> G
G --> H
H --> C
I --> F
C --> J
```
**Notes:** The orchestrator schedules parallel metric computation, handles timeouts, and coordinates with the model manager for model loading and caching.
---
## 🔁 Workflow / Data Flow
```mermaid
%%{init: {
"theme": "dark",
"themeVariables": {
"fontSize": "10px",
"fontFamily": "Segoe UI, Helvetica, Arial, sans-serif"
}
}}%%
sequenceDiagram
participant U as 👤 User<br/>(Web / API)
participant API as 🚪 FastAPI<br/>
participant O as 🧠 Orchestrator<br/>Domain + Preprocess
participant M as 📊 Metrics Pool<br/>6 Detectors (ParallelGroup)
participant E as ⚖️ Ensemble<br/>Domain Based<br/>Confidence Calibration
participant R as 📝 Reporter<br/>PDF/JSON Export
U->>API: 📤 Submit text or file
API->>O: ✅ Validate & enqueue job
O->>M: ⚡ Run metrics in parallel
M-->>O: 📈 Return evidence scores
O->>E: 🔗 Aggregate & calibrate
E-->>O: 🎯 Verdict + uncertainty
O->>R: 🖨️ Generate report
R-->>API: 📦 JSON/PDF ready
API-->>U: 🔗 Return analysis + download
```
---
## 🧮 Forensic Evidence Signals & Mathematical Foundation
This section provides the exact metric definitions implemented in `metrics/` and rationale for their selection. The ensemble combines these orthogonal signals to increase robustness against edited, paraphrased, or algorithmically regularized text.
### Metric summary (weights are configurable per domain)
- Perplexity — 25%
- Entropy — 20%
- Structural — 15%
- Semantic — 15%
- Linguistic — 15%
- Multi-perturbation Stability — 10%
### 1) Perplexity (25% weight)
**Definition**
```math
Perplexity = \exp\left(-\frac{1}{N}\sum_{i=1}^N \log P(w_i\mid context)\right)
```
**Implementation sketch**
```python
def calculate_perplexity(text, model, k=512):
tokens = tokenize(text)
log_probs = []
for i in range(len(tokens)):
context = tokens[max(0, i-k):i]
prob = model.get_probability(tokens[i], context)
log_probs.append(math.log(prob))
return math.exp(-sum(log_probs)/len(tokens))
```
**Domain calibration example**
```python
if domain == Domain.ACADEMIC:
perplexity_threshold *= 1.2
elif domain == Domain.SOCIAL_MEDIA:
perplexity_threshold *= 0.8
```
### 2) Entropy (20% weight)
**Shannon entropy (token level)**
```math
H(X) = -Σ p(x_i) * log₂ p(x_i)
```
**Implementation sketch**
```python
from collections import Counter
def calculate_text_entropy(text):
tokens = text.split()
token_freq = Counter(tokens)
total = len(tokens)
entropy = -sum((f/total) * math.log2(f/total) for f in token_freq.values())
return entropy
```
### 3) Structural Metric (15% weight)
**Burstiness**
```math
Burstiness = \frac{\sigma - \mu}{\sigma + \mu}
```
where:
- μ = mean sentence length
- σ = standard deviation of sentence length
**Uniformity**
```math
Uniformity = 1 - \frac{\sigma}{\mu}
```
where:
- μ = mean sentence length
- σ = standard deviation of sentence length
**Sketch**
```python
def calculate_burstiness(text):
sentences = split_sentences(text)
lengths = [len(s.split()) for s in sentences]
mean_len = np.mean(lengths)
std_len = np.std(lengths)
burstiness = (std_len - mean_len) / (std_len + mean_len)
uniformity = 1 - (std_len/mean_len if mean_len > 0 else 0)
return {'burstiness': burstiness, 'uniformity': uniformity}
```
### 4) Semantic Analysis (15% weight)
**Coherence (sentence embedding cosine similarity)**
```math
Coherence = \frac{1}{n} \sum_{i=1}^{n-1} \cos(e_i, e_{i+1})
```
**Sketch**
```python
def calculate_semantic_coherence(text, embed_model):
sentences = split_sentences(text)
embeddings = [embed_model.encode(s) for s in sentences]
sims = [cosine_similarity(embeddings[i], embeddings[i+1]) for i in range(len(embeddings)-1)]
return {'mean_coherence': np.mean(sims), 'coherence_variance': np.var(sims)}
```
### 5) Linguistic Metric (15% weight)
**POS diversity, parse tree depth, syntactic complexity**
```python
def calculate_linguistic_features(text, nlp_model):
doc = nlp_model(text)
pos_tags = [token.pos_ for token in doc]
pos_diversity = len(set(pos_tags))/len(pos_tags)
depths = [max(get_tree_depth(token) for token in sent) for sent in doc.sents]
return {'pos_diversity': pos_diversity, 'mean_tree_depth': np.mean(depths)}
```
### 6) MultiPerturbationStability (10% weight)
**Stability under perturbation** (curvature principle)
```math
Stability = \frac{1}{n} \sum_{j} \left| \log P(x) - \log P(x_{perturbed_j}) \right|
```
```python
def multi_perturbation_stability_score(text, model, num_perturbations=20):
original = model.get_log_probability(text)
diffs = []
for _ in range(num_perturbations):
perturbed = generate_perturbation(text)
diffs.append(abs(original - model.get_log_probability(perturbed)))
return np.mean(diffs)
```
---
## 🏛️ Ensemble Methodology
### Confidence‑Calibrated Aggregation (high level)
- Start with domain base weights (e.g., `DOMAIN_WEIGHTS` in `config/threshold_config.py`)
- Adjust weights per metric with a sigmoid confidence scaling function
- Normalize and compute weighted aggregate
- Quantify uncertainty using variance, confidence means, and decision distance from 0.5
```python
def ensemble_aggregation(metric_results, domain):
base = get_domain_weights(domain)
adj = {m: base[m] * sigmoid_confidence(r.confidence) for m, r in metric_results.items()}
total = sum(adj.values())
final_weights = {k: v/total for k, v in adj.items()}
return weighted_aggregate(metric_results, final_weights)
```
### Uncertainty Quantification
```python
def calculate_uncertainty(metric_results, ensemble_result):
var_uncert = np.var([r.synthetic_probability for r in metric_results.values()])
conf_uncert = 1 - np.mean([r.confidence for r in metric_results.values()])
decision_uncert = 1 - 2*abs(ensemble_result.synthetic_probability - 0.5)
return var_uncert*0.4 + conf_uncert*0.3 + decision_uncert*0.3
```
---
## 🧭 Domain‑Aware Detection
Domain weights and thresholds are configurable. Example weights (in `config/threshold_config.py`):
```python
DOMAIN_WEIGHTS = {'academic' : {'perplexity':0.22,'entropy':0.18,'structural':0.15,'linguistic':0.20,'semantic':0.15,'multi_perturbation_stability':0.10},
'technical' : {'perplexity':0.20,'entropy':0.18,'structural':0.12,'linguistic':0.18,'semantic':0.22,'multi_perturbation_stability':0.10},
'creative' : {'perplexity':0.25,'entropy':0.25,'structural':0.20,'linguistic':0.12,'semantic':0.10,'multi_perturbation_stability':0.08},
'social_media' : {'perplexity':0.30,'entropy':0.22,'structural':0.15,'linguistic':0.10,'semantic':0.13,'multi_perturbation_stability':0.10},
}
```
### Domain Calibration Strategy (brief)
- **Academic**: increase linguistic weight, raise perplexity multiplier
- **Technical**: prioritize semantic coherence, maximize Synthetic threshold to reduce false positives
- **Creative**: boost entropy & structural weights for burstiness detection
- **Social Media**: prioritize perplexity and relax linguistic demands
---
## 📊 Evaluation & Benchmarks
### Comprehensive System Validation
TEXT-AUTH has been rigorously evaluated on **2,750 text samples** across 16 domains using a multi-subset benchmark designed to test:
- **Baseline performance** on clean human vs. AI text
- **Cross-model generalization** (different AI models)
- **Adversarial robustness** (paraphrased content)
**Evaluation Dataset: TEXT-AUTH-Eval**
- **1,444 samples** - CLEAN subset (baseline)
- **682 samples** - CROSS_MODEL subset (generalization test)
- **500 samples** - PARAPHRASED subset (robustness test)
- **124 samples** - Classified as UNCERTAIN (appropriate abstention)
### Overall Performance Metrics
| Metric | Value | Status |
|--------|-------|--------|
| **Overall Accuracy** | 78.3% | Production-Ready |
| **F1 Score** | 85.7% | Exceeds Target (>75%) |
| **Precision (AI Detection)** | 84.3% | High Reliability |
| **Recall (AI Detection)** | 87.2% | Strong Coverage |
| **AUROC** | 0.777 | Good Discrimination |
| **AUPRC** | 0.888 | Excellent Precision-Recall |
| **ECE (Calibration)** | 0.080 | Well-Calibrated |
**4-Class System Behavior:**
- **Coverage**: 95.5% (decisive predictions)
- **Abstention Rate**: 4.5% (appropriate uncertainty handling)
- **Hybrid Detection**: 0.5% (mixed-authorship identification)
**Verdict Distribution:**
```
Synthetically-Generated: 73.3% (2,017 samples)
Authentically-Written: 21.7% (596 samples)
Hybrid: 0.5% (13 samples)
Uncertain: 4.5% (124 samples)
```
### Performance by Evaluation Subset
| Subset | Samples | F1 Score | Coverage | Abstention | Hybrid Rate |
|--------|---------|----------|----------|------------|-------------|
| **CLEAN** | 1,444 | 78.6% | 92.4% | 7.6% | 0.6% |
| **CROSS_MODEL** | 682 | **95.3%** ⭐ | 99.1% | 0.9% | 0.1% |
| **PARAPHRASED** | 500 | 86.1% | 100.0% | 0.0% | 0.8% |
**Key Insights:**
- **Exceptional cross-model generalization** (95.3% F1) - system detects AI patterns regardless of specific model
- **Strong adversarial robustness** (86.1% F1) - maintains performance on paraphrased content
- **Adaptive abstention** - higher uncertainty on CLEAN set (7.6%) reflects appropriate caution on ambiguous cases
### Domain-Specific Performance
#### Top Performing Domains (F1 > 90%)
| Domain | F1 Score | Coverage | Abstention | Notes |
|--------|----------|----------|------------|-------|
| **General** | **93.4%** | 91.8% | 8.2% | Encyclopedic content |
| **Creative** | **92.9%** | 83.5% | 16.5% | Literary narratives |
| **Medical** | **90.3%** | 100.0% | 0.0% | Clinical terminology |
| **Journalism** | **90.3%** | 93.1% | 6.9% | News reporting |
#### Strong Performing Domains (F1 85-90%)
| Domain | F1 Score | Coverage | Abstention | Notes |
|--------|----------|----------|------------|-------|
| **AI/ML** | 88.8% | 99.2% | 0.8% | Technical AI content |
| **Academic** | 87.5% | 100.0% | 0.0% | Research papers |
| **Tutorial** | 87.5% | 94.2% | 5.8% | How-to guides |
| **Business** | 86.2% | 94.9% | 5.1% | Business writing |
| **Science** | 86.2% | 95.4% | 4.6% | Scientific content |
| **Technical Doc** | 85.6% | 94.6% | 5.4% | Documentation |
#### Moderate Performing Domains (F1 80-85%)
| Domain | F1 Score | Coverage | Abstention | Hybrid % | Notes |
|--------|----------|----------|------------|----------|-------|
| **Blog/Personal** | 83.8% | 96.7% | 3.3% | 0.0% | Personal blogs |
| **Marketing** | 84.0% | 96.0% | 4.0% | 0.0% | Marketing copy |
| **Engineering** | 82.0% | 100.0% | 0.0% | 1.7% | Technical specs |
| **Software Dev** | 81.9% | 94.9% | 5.1% | 3.9% | Code documentation |
#### Challenging Domains (F1 < 80%)
| Domain | F1 Score | Coverage | Abstention | Hybrid % | Notes |
|--------|----------|----------|------------|----------|-------|
| **Legal** | 77.1% | 94.9% | 5.1% | 1.6% | ⚠️ Formal legal language |
| **Social Media** | 73.3% | 98.9% | 1.1% | 0.8% | ⚠️ Short, informal text |
**Domain Analysis:**
- Achieves >80% F1 on 14 of 16 domains
- Strong performance on structured content (medical, journalism, general)
- Legal domain challenging due to highly formulaic language patterns
- Social media challenging due to brevity and informal style
### Performance by Text Length
| Length Range | Samples | F1 Score | Precision | Recall | Accuracy | Abstention | Avg Time (s) |
|--------------|---------|----------|-----------|--------|----------|------------|--------------|
| **Very Short (0-100)** | 18 | 0.000 | 0.000 | 0.000 | 0.278 | 0.0% | 4.6 |
| **Short (100-200)** | 249 | 0.211 | 0.118 | 0.947 | 0.458 | 0.0% | 8.3 |
| **Medium (200-400)** | 1,682 | **0.885** | 0.901 | 0.869 | 0.813 | 0.6% | 18.2 |
| **Medium-Long (400-600)** | 630 | **0.900** ⭐ | 0.929 | 0.872 | 0.833 | 7.1% | 23.6 |
| **Long (600-1000)** | 15 | 0.000 | 0.000 | 0.000 | 1.000 | 74.6% | 37.1 |
| **Very Long (1000+)** | 19 | 0.000 | 0.000 | 0.000 | 1.000 | 64.8% | 108.4 |
**Length-Performance Insights:**
- **Optimal range: 200-600 words** (F1: 0.885-0.900)
- **Strong performance on medium-length texts** - majority of real-world content
- **Limited signals on very short texts** (<100 words) - insufficient statistical context
- **High abstention on very long texts** (>600 words) - system appropriately defers to human judgment
- **Processing time scales sub-linearly** - efficient even on longer texts
**Statistical Analysis:**
- Pearson correlation coefficient: r = 0.833
- No statistically significant linear correlation (p = 0.374)
- Performance peaks at 400-600 words, then plateaus with increased abstention
### Evaluation Visualizations
#### Confusion Matrix (Decisive Predictions Only)

**Binary Classification Performance:**
- **True Negatives (Human → Human)**: 344 samples
- **True Positives (AI/Hybrid → AI)**: 1,711 samples
- **False Positives (Human → AI)**: 252 samples (14.7% FP rate)
- **False Negatives (AI → Human)**: 319 samples (15.7% FN rate)
#### Length-Based Analysis

**Key Patterns:**
1. **Performance peaks at medium-length texts** (400-600 words)
2. **Abstention rate increases dramatically** for texts >600 words
3. **Processing time grows sub-linearly** with text length
4. **Sample distribution heavily concentrated** in 200-600 word range (84% of dataset)
---
### 🔍 Key Findings & Insights
#### Strengths
1. **High Precision-Recall Balance** (84.3% precision, 87.2% recall)
- Minimizes both false positives and false negatives
- Well-suited for high-stakes decision support
2. **Exceptional Cross-Model Generalization** (95.3% F1)
- Detects AI patterns regardless of specific generation model
- Robust to model diversity in real-world scenarios
3. **Appropriate Uncertainty Handling** (4.5% abstention)
- System abstains on genuinely ambiguous cases
- Prevents overconfident incorrect predictions
4. **Well-Calibrated Confidence** (ECE = 0.080)
- Reported confidence scores match actual accuracy
- Reliable uncertainty estimates for decision-making
5. **Domain Adaptability** (14/16 domains >80% F1)
- Performs well across diverse writing styles
- Domain-specific thresholds reduce false positives
#### Challenges & Limitations
1. **Very Short Texts** (<100 words)
- Insufficient statistical signals for reliable analysis
- Recommendation: Require minimum 100-word samples
2. **Highly Formulaic Content**
- Legal documents and templates show lower F1 (77.1%)
- Human-written formulaic text can resemble AI patterns
- Mitigation: Review Hybrid verdicts in formal domains
3. **Very Long Texts** (>1000 words)
- High abstention rates (65-75%)
- System appropriately defers to human review
- Recommendation: Analyze long documents in sections
---
### 📚 Evaluation Methodology
**Dataset Construction:**
- Human-written texts sourced from Wikipedia, arXiv, C4, PubMed, Project Gutenberg
- AI-generated texts created using Ollama (mistral:7b baseline)
- Cross-model texts generated with llama3:8b (generalization test)
- Paraphrased texts created by rephrasing AI outputs (robustness test)
**Metrics Computation:**
- **Binary metrics** (Precision, Recall, F1) computed on decisive predictions only
- **Hybrid verdicts** counted as successful AI detection
- **Uncertain verdicts** excluded from accuracy calculation (appropriate abstention)
- **Coverage** = percentage of samples with decisive predictions
**Evaluation Infrastructure:**
- 2,750 total samples evaluated
- 16 domains × 3 subsets (clean, cross-model, paraphrased)
- Processing time: ~1.5 hours on standard hardware
- Results exported to JSON, CSV, and visualization formats
For detailed evaluation methodology and reproduction instructions, see [EVALUATION.md](docs/EVALUATION.md).
---
## 📁 Project Structure (as in repository)
```text
text_auth/
├── config/
│ ├── model_config.py
│ ├── settings.py
| ├── enums.py
| ├── constants.py
| ├── schemas.py
│ └── threshold_config.py
├── data/
│ ├── reports/
| ├── validation_data/
│ └── uploads/
├── services/
│ ├── reasoning_generator.py
│ ├── ensemble_classifier.py
│ ├── highlighter.py
│ └── orchestrator.py
├── metrics/
│ ├── base_metric.py
│ ├── multi_perturbation_stability.py
│ ├── entropy.py
│ ├── linguistic.py
│ ├── perplexity.py
│ ├── semantic_analysis.py
│ └── structural.py
├── models/
│ ├── model_manager.py
│ └── model_registry.py
├── processors/
│ ├── document_extractor.py
│ ├── domain_classifier.py
│ ├── language_detector.py
│ └── text_processor.py
├── reporter/
│ └── report_generator.py
├── ui/
│ └── static/index.html
├── utils/
│ └── logger.py
├── validation/
├── requirements.txt
├── README.md
├── Dockerfile
├── .gitignore
├── test_integration.py
├── .env.example
├── requirements.txt
└── text_auth_app.py
```
---
## 🌐 API Endpoints
### `/api/analyze` — Text Analysis (POST)
Analyze raw text. Returns ensemble assessment, per‑metric signals, highlights, and explainability reasoning.
**Request (JSON)**
```json
{
"text":"...",
"domain":"academic|technical_doc|creative|social_media",
"enable_highlighting": true,
"use_sentence_level": true,
```
**Response (JSON)** — abbreviated
```json
{
"status": "success",
"analysis_id": "analysis_170...",
"assessment": {
"final_verdict": "Synthetic / Authentic / Hybrid",
"overall_confidence": 0.89,
"uncertainty_score": 0.23
},
"metric_signals": {
"perplexity": { "score": 0.92, "confidence": 0.89 }
},
"highlighted_html": "<div>...</div>",
"reasoning": {
"summary": "...",
"key_indicators": ["...", "..."]
}
}
```
> **Note:** The final verdict represents a probabilistic consistency assessment, not an authorship or generation claim.
### `/api/analyze/file` — File Analysis (POST, multipart/form-data)
Supports PDF, DOCX, TXT, DOC, MD. File size limit default: 10MB. Returns same structure as text analyze endpoint.
### `/api/report/generate` — Report Generation (POST)
Generate downloadable JSON or PDF reports for a given analysis id.
### Utility endpoints
- `GET /health` — health status, models loaded, uptime
- `GET /api/domains` — supported domains and thresholds
- `GET /api/models` — detectable model list
---
## ⚙️ Installation & Setup
### Prerequisites
- Python 3.8+
- 4GB RAM (8GB recommended)
- Disk: 2GB (models & deps)
- OS: Linux/macOS/Windows (WSL supported)
### Quickstart
```bash
git clone https://github.com/satyaki-mitra/text_authentication.git
cd text_authentication
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
# Copy .env.example -> .env and set HF_TOKEN if using private models
python text_auth_app.py
# or: ./run.sh
```
**Dev tips**
- Use `DEBUG=True` in `config/settings.py` for verbose logs
- For containerized runs, see `Dockerfile` template (example included in repo suggestions)
---
## 🧠 Model Management & First‑Run Behavior
- The application **automatically downloads** required model weights from Hugging Face on the first run and caches them to the local HF cache (or a custom path specified in `config/model_config.py`).
- Model IDs and revisions are maintained in `models/model_registry.py` and referenced by `models/model_manager.py`.
- **Best practices implemented**:
- Pin model revisions (e.g., `repo_id@v1.2.0`)
- Resumeable downloads using `huggingface_hub.snapshot_download`
- Optional `OFFLINE_MODE` to load local model paths
- Optional integrity checks (SHA256) after download
- Support for private HF repos using `HF_TOKEN` env var
**Example snippet**
```python
from huggingface_hub import snapshot_download
snapshot_download(repo_id="satyaki-mitra/statistical-text-reference-v1", local_dir="./models/text-detector-v1")
```
---
## 🎨 Frontend Features (UI)
- Dual‑panel responsive web UI (left: input / upload; right: live analysis)
- Sentence‑level color highlights with tooltips and per‑metric breakdown
- Progressive analysis updates (metric-level streaming)
- Theme: light/dark toggle (UI respects user preference)
- Export: JSON and PDF report download
- Interactive elements: click to expand sentence reasoning, copy text snippets, download raw metrics
---
## 💼 Business Model & Market Analysis
**TAM**: $20B (education, hiring, publishing) — see detailed breakdown in original repo.
**Use cases**: universities (plagiarism & integrity), hiring platforms (resume authenticity), publishers (content verification), social platforms (spam & SEO abuse).
**Competitive landscape** (summary)
- Binary authorship-claim systems (e.g., GPTZero-style tools) — our advantages: domain adaptation, explainability, evidence transparency, lower false positives and competitive pricing. TEXT-AUTH explicitly avoids authorship claims in favor of evidence-based forensic assessment.
**Monetization ideas**
- SaaS subscription (seat / monthly analyze limits)
- Enterprise licensing with on‑prem deployment & priority support
- API billing (per‑analysis tiered pricing)
- Onboarding & consulting for institutions
---
## 🔮 Research Impact & Future Scope
**Research directions**
- Adversarial robustness (paraphrase & synonym attacks)
- Cross‑model generalization & zero‑shot detection
- Explainability: counterfactual examples & feature importance visualization
**Planned features (Q1‑Q2 2026)**
- Multi‑language support (Spanish, French, German, Chinese)
- Real‑time streaming API (WebSocket)
- Institution‑specific calibration & admin dashboards
*Detailed research methodology and academic foundation available in our [Whitepaper](docs/WHITE_PAPER.md). Technical implementation details in [Technical Documentation](docs/BLOGPOST.md).*
---
## 🏗️ Deployment notes**
- Containerize app with Docker, orchestrate with Kubernetes or ECS for scale
- Autoscaling groups for API servers & worker nodes
- Use spot GPU instances for retraining & large metric compute jobs
- Integrate observability: Prometheus + Grafana, Sentry for errors, Datadog if available
---
## 🔐 Security & Risk Mitigation
**Primary risks & mitigations**
- Model performance drift — monitoring + retraining + rollback
- Adversarial attacks — adversarial training & input sanitization
- Data privacy — avoid storing raw uploads unless user consents; redact PII in reports
- Secrets management — use env vars, vaults, and avoid committing tokens
- Rate limits & auth — JWT/OAuth2, API key rotation, request throttling
**File handling best practices (examples)**
```python
ALLOWED_EXT = {'.txt','.pdf','.docx','.doc','.md'}
def allowed_file(filename):
return any(filename.lower().endswith(ext) for ext in ALLOWED_EXT)
```
---
## 📄 License & Acknowledgments
This project is licensed under the **MIT License** — see [LICENSE](LICENSE) in the repo.
Acknowledgments:
- DetectGPT (Mitchell et al., 2023) — inspiration for perturbation-based detection
- Hugging Face Transformers & Hub
- Open-source NLP community and early beta testers
---
<div align="center">
**Built for Evidence-based text forensics, transparency, and real-world readiness.**
*Version 1.0.0 — Last Updated: October, 2025*
</div>
|