Use the AI Without Handing It Your Secrets

The privacy failures of the AI era rarely look like attacks. They look like a nurse pasting patient notes into a chatbot to get a cleaner summary, or an analyst uploading a payroll file to ask a question about headcount. Nobody broke in. Someone just wanted help, and the fastest way to get it was to hand over the whole document. AnonyMask is built for that moment, on the reasonable assumption that telling people not to use these tools has never once worked.

It is a privacy-preserving tool that automatically detects, masks, and unmasks sensitive data across common file formats including PDF, DOCX, XLSX, XLS, CSV, and TXT. The workflow is the clever part. Before a document goes to an LLM or a RAG system, AnonyMask strips the private data out and leaves a token map behind. The AI analyzes a document that never contained the sensitive values in the first place. When the result comes back, smart unmasking restores the original content from that mapping, so the analyst gets a usable answer without the raw PII ever leaving the machine.

The distinction it draws between explicit and implicit privacy data is what makes it more than a regular expression with ambitions. Explicit data is the easy half, the things that look like what they are, such as an email address or an identity number. Implicit data is the harder problem, information that identifies a person through context rather than format, and AnonyMask handles thirty-three labels using transformer-based models, with multilingual detection across English and Indonesian through models like XLM-RoBERTa. That local language coverage matters more than it might appear, since most privacy tooling is built for English data and quietly fails on everything else.

Control over how masking happens is granular in a way that suggests real use. Redacted masking hides everything behind asterisks. Partial masking keeps fragments so a human can still read the document. Full masking can substitute a category label, a custom value, or randomized replacements, either independently per occurrence or consistently, so that every mention of the same person maps to the same pseudonym and the document’s internal logic survives. Users can add or exclude specific terms, and every operation is logged for traceability and audit.

The architecture reflects its stated principle of privacy by design. Everything runs offline and locally, distributed as a standalone Windows executable with no external dependencies, built with an Angular frontend, a Python backend, and Tauri with Rust wrapping it into a desktop application. Sending your data to a cloud service in order to protect it from a cloud service would rather defeat the purpose. The project is released under Apache 2.0 and was presented at Black Hat Arsenal Europe 2025.

Where This Meets the Curriculum

AnonyMask sits directly on top of Cyber Law, which is where students learn that Indonesia’s UU PDP and Europe’s GDPR are not abstractions but constraints that shape what an organization is allowed to do with data. This tool is what compliance looks like when it becomes engineering rather than policy. It also belongs to the fastest-growing frontier in the field, the security and privacy of AI systems, where the questions are new enough that the answers are still being written. That its authors are BINUS students who took a live regulatory problem, built a working answer, and presented it at Black Hat in Europe is worth stating plainly. This is what the project-based, industry-facing philosophy of the program produces when a student decides to take it seriously.

Explore AnonyMask

Repository: github.com/Caudrey/AnonyMask