David Oygenblik (Georgia Institute of Technology), Dinko Dermendzhiev (Georgia Institute of Technology), Filippos Sofias (Georgia Institute of Technology), Mingxuan Yao (Georgia Institute of Technology), Haichuan Xu (Georgia Institute of Technology), Runze Zhang (Georgia Institute of Technology), Jeman Park (Kyung Hee University), Amit Kumar Sikder (Iowa State University), Brendan Saltaformaggio (Georgia Institute of Technology)

Prior work has developed techniques capable of extracting deep learning (DL) models in universal formats from system memory or program binaries for security analysis. Unfortunately, such techniques ignore the recovery of the DL model’s programmatic representation required for model reuse and any white-box analysis techniques. Addressing this, we propose a novel recovery methodology, and prototype ZEN, that automatically recovers the DL model programmatic representation complementing the recovery of the mathematical representation by prior work. ZEN identifies novel code in an unknown DL system relative to a base model and generates patches uch that the recovered DL model can be reused. We evaluated ZEN on 21 SOTA DL models, including models across the language and vision domains, such as Llama 3 and YoloV10. ZEN successfully attributed custom models to their base models with 100% accuracy, enabling model reuse.

View More Papers

Defending Job Platforms from Non-Genuine Applications Using Layered Detection...

Rama Rohit Reddy Gangula (Indeed), Vijay Vardhan Alluri (Indeed), Saif Jawaid (Indeed), Dhwaj Raj (Indeed), Udit Jindal (Indeed)

Read More

Strategic Games and Zero Shot Attacks on Heavy-Hitter Network...

Francesco Da Dalt (ETH Zürich), Adrian Perrig (ETH Zurich)

Read More

MVP-ORAM: a Wait-free Concurrent ORAM for Confidential BFT Storage

Robin Vassantlal (LASIGE, Faculdade de Ciências, Universidade de Lisboa, Portugal), Hasan Heydari (LASIGE, Faculdade de Ciências, Universidade de Lisboa, Portugal), Bernardo Ferreira (LASIGE, Faculdade de Ciências, Universidade de Lisboa, Portugal), Alysson Bessani (LASIGE, Faculdade de Ciências, Universidade de Lisboa, Portugal)

Read More