GPU Undervolting Boosts CNN Adversarial Robustness and Energy Efficiency

Researchers have developed a method to enhance the adversarial robustness of Convolutional Neural Networks (CNNs) by undervolting their GPUs during training. This technique introduces stochastic perturbations that act as implicit regularization, improving model resilience against adversarial attacks while simultaneously reducing energy consumption. Experiments with LeNet, VGG-6, and MobileNetV3 models on MNIST and CIFAR-10 datasets demonstrated that undervolting consistently led to higher adversarial accuracy compared to standard training at nominal voltage, even when adversarial training was already employed. This hardware-level approach offers a practical way to achieve both robustness and energy efficiency without requiring algorithmic modifications. AI

IMPACT Offers a hardware-level technique to improve AI model security and reduce operational costs.

RANK_REASON Academic paper detailing a novel method for improving model robustness. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• CIFAR-10
• graphics processing unit
• MNIST database
• MobileNetV3
• VGG-6

• paper
• safety
• infra

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