Backdoors in MLMs

Researchers have developed various tricks to hide their own sample backdoors in machine learning models. 

But the approach has been largely trial and error, lacking formal mathematical analysis of how well those backdoors are hidden.

Researchers are now starting to analyze the security of machine learning models in a more rigorous way. 

In a paper presented at last year’s Foundations of Computer Science conference, a team of computer scientists demonstrated how to plant undetectable backdoors whose invisibility is as certain as the security of state-of-the-art encryption methods.

The mathematical rigor of the new work comes with trade-offs, like a focus on relatively simple models. 

But the results establish a new theoretical link between cryptographic security and machine learning vulnerabilities, suggesting new directions for future research at the intersection of the two fields.