Raw signalOct 16, 20261 min read1 source

Code and data released at thu-coai/Backdoor-Data-Extraction.

In one minute

Tsinghua researchers have released the code and processed data for experiments on extracting private fine-tuning examples from open-weight models.

The repository describes a supply-chain scenario in which a model creator trains a backdoor into a base model before release; after a downstream organization fine-tunes it on proprietary data, the creator can use black-box queries to recover examples.

The README reports experiments on four base models ranging from 3B to 32B parameters and two downstream datasets. It says as much as 76.3% of 5,000 queries were perfectly extracted in practical settings, rising to 94.9% under more ideal conditions. The project also reports that its detection-based defense can be bypassed by an improved attack. These are controlled research results; the repository does not establish that this attack has occurred in deployed systems.

The release provides training and evaluation scripts, environment instructions, and processed datasets. It makes the experimental claim more inspectable and reproducible, while leaving open how the method performs against real-world fine-tuning pipelines and defenses.

Evidence / sources

  1. 1github.com/thu-coai/Backdoor-Data-Extraction
All Signals

Next reads

Read next

Open the graph
  1. 01archiveSignalsOpen the signals and the raw-signal archive.
  2. 02sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract