InferNet exploits GPU profiles for DNN architecture inference

Researchers have developed InferNet, a novel method for inferring the architecture of deep neural networks (DNNs) by analyzing aggregate GPU profiles. This technique bypasses the need for complex, fine-grained data analysis, instead utilizing coarse-grained system-level information like GPU kernel calls and memory events. InferNet can accurately predict both the general architecture family and specific variants, achieving 100% model extraction accuracy in evaluations across different AI frameworks, DNN types, and hardware platforms. AI

IMPACT This research could lead to new methods for detecting and mitigating DNN model extraction attacks, enhancing AI security.

RANK_REASON The item is a research paper detailing a new method for inferring DNN architectures. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Bert
• deep neural network
• GoogLeNet
• graphics processing unit
• Inceptionv3
• InferNet
• Jonah O'Brien Weiss
• NVIDIA Quadro RTX 8000
• Nvidia T4
• PyTorch
• Tensorflow

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →