ZEN Accepted to NDSS 26

My paper, Achieving Zen Combining Mathematical and Programmatic Deep Learning Model Representations for Attribution and Reuse has been accepted to the the 33rd Network and Distributed System Security (NDSS) Symposium.

Here is a synopsis of it, more to come in a future blog post with the link to the presentation and paper:

Existing techniques for recovering and analyzing AI models from black-box systems are insufficient for practical use as they only extract the model’s weights, layers, and structure while ignoring the underlying code that actually defines how the model uses those recovered components. This paper introduces ZEN, a novel framework that bridges this gap by recovering a model’s “unified representation”, essentially a unique fingerprint combining both its mathematical structure (weights, layers, graph structure, etc.) and its programmatic implementation (code) from a memory image. Using this fingerprint, ZEN automatically attributes an unknown black-box model to a known open-source base model and generates code patches to replicate the black-box model’s custom functionality. In evaluations across 21 models, ZEN successfully attributed every custom model to its correct base with 100% accuracy, enabling full model reuse for white-box analysis.

Thank you CyFI Lab!!