Dashboard Structural Analysis anonym.plus SD2 IRREVERSIBILITY Case Study
← Previous Next →
anonym.plus SD2 IRREVERSIBILITY
Case Study 16 of 30

De-Identification of Facial Features in Magnetic Resonance Images: Software Development Using Deep Learning Technology

Jeong, Yeon Uk, Yoo, Soyoung, Kim, Young-Hak et al. · Journal of Medical Internet Research (2020)

Research Source

De-Identification of Facial Features in Magnetic Resonance Images: Software Development Using Deep Learning Technology
Jeong, Yeon Uk, Yoo, Soyoung, Kim, Young-Hak et al. · Journal of Medical Internet Research · 2020 · Source: doaj

BackgroundHigh-resolution medical images that include facial regions can be used to recognize the subject’s face when reconstructing 3-dimensional (3D)-rendered images from 2-dimensional (2D) sequential images, which might constitute a risk of infringement of personal information when sharing data.

Executive Summary

This research paper examines a critical privacy challenge related to IRREVERSIBILITY — once pii propagates, it cannot be un-propagated.

anonym.plus addresses this through 100% local processing with AES-256-GCM encrypted vault — PII processed and stored locally, never touching any external server.

Root Cause: SD2 — IRREVERSIBILITY

Once PII propagates, it cannot be un-propagated. The arrow of data only points one direction. PII exposure is a one-way function with no inverse.

Irreducible truth: Information entropy only increases. You cannot recall a broadcast signal. You cannot un-train a neural network. You cannot selectively erase a backup tape. Every deletion mechanism is an approximation — and the original exposure persists.

The Solution: How anonym.plus Addresses This

Detection Capabilities

anonym.plus identifies 200+ entity types including names, emails, phone numbers, medical records, training data with PII. The local Presidio 2.2.357 + spaCy 3.8.11 architecture uses Presidio 2.2.357 deterministic recognizers with 121 built-in presets for structured identifiers and spaCy 3.8.11 with 23 language models, all running locally via FastAPI sidecar for contextual references.

Anonymization Methods

Replace is recommended for this pain point: substituting PII in training data with realistic synthetic alternatives preserves statistical properties while preventing memorization. Redact provides an alternative — removing PII entirely from training data eliminates memorization risk at the cost of reduced training diversity. For scenarios requiring reversibility, Encrypt (AES-256-GCM) enables authorized recovery of original values.

Architecture & Deployment

Documents and datasets are batch-anonymized before ML training. The 200+ entity types with 121 presets cover common training data PII patterns. Processed data never leaves the machine.

Compliance Mapping

This pain point intersects with GDPR Article 25 data protection by design, Article 5(1)(c) minimization.

anonym.plus’s GDPR (data never leaves device), HIPAA (local processing) compliance coverage, combined with 100% local — data never leaves device hosting, provides documented technical measures organizations can reference in their compliance documentation and regulatory submissions.

Product Specifications

SpecificationValue
App Versionv8.10.5
Entity Types200+ built-in, up to 50 custom
Detection EnginePresidio 2.2.357 + spaCy 3.8.11 (23 models)
Languages48 UI, 23 NLP models
Document FormatsPDF, DOCX, XLSX, TXT, CSV, JSON, XML + Image OCR
Anonymization MethodsReplace, Redact, Mask, Hash (SHA-256/512/MD5), Encrypt (AES-256-GCM)
ArchitectureTauri 2.x (Rust + React) + FastAPI sidecar (~370 MB)
PlatformsWin/Mac/Linux
LicensingEd25519 signed, machine-fingerprinted, max 5 machines
Processing100% local — data never leaves device
ComplianceGDPR, HIPAA (data residency guaranteed by local processing)
← Previous Next →