Genetic improvement (GI) in Deep Neural Networks (DNNs) has traditionally enhanced neural architecture and trained DNN parameters. Recently, GI has supported large language models by optimizing DNN operator scheduling on accelerator clusters. However, with the rise of adversarial AI, particularly model extraction attacks, there is an unexplored potential for GI in fortifying Machine Learning as a Service (MLaaS) models. We suggest a novel application of GI — not to improve model performance, but to diversify operator parallelism for the purpose of a moving target defense against model extraction attacks. We discuss an application of GI to create a DNN model defense strategy that uses probabilistic isolation, offering unique benefits not present in current DNN defense systems.
Tue 16 AprDisplayed time zone: Lisbon change
11:00 - 12:30 | Presentation Session 1GI@ICSE at Vianna da Motta Chair(s): Gabin An Korea Advanced Institute of Science and Technology, Justyna Petke University College London | ||
11:00 30mTalk | Deep Mutations have Little Impact GI@ICSE William Langdon University College London | ||
11:30 30mTalk | Grammar evolution and symbolic regression for astrometric centering of Hubble Space Telescope images GI@ICSE R. Sarmiento Universidad Internacional de la Rioja (UNIR), Spain, M. de la Cruz Universidad Internacional de la Rioja (UNIR), Spain, A. Ortega Universidad Internacional de la Rioja (UNIR), Spain, R. Baena-Galle Universidad Internacional de la Rioja (UNIR), Spain, T.M. Girard Southern Connecticut State University, USA, D.I. Casetti-Dinescu Southern Connecticut State University, USA, A. Cervantes Universidad Internacional de la Rioja (UNIR), Spain | ||
12:00 15mTalk | Genetic Improvement for DNN Security GI@ICSE Hunter Baxter Vanderbilt University, Yu Huang Vanderbilt University, Kevin Leach Vanderbilt University | ||
12:15 15mOther | Discussion GI@ICSE | ||