Authors: Yatong Bai, Mo Zhou, Vishal M. Patel, Somayeh Sojoudi

Published on: February 03, 2024

Impact Score: 8.3

Arxiv code: Arxiv:2402.02263

Summary

• What is new: Introducing ‘MixedNUTS’, a training-free method that leverages the ‘benign confidence property’ of models to enhance both accuracy and robustness in an ensemble setting.
• Why this is important: Adversarial robustness typically reduces accuracy, making robust classification models less effective for real-life applications.
• What the research proposes: MixedNUTS processes output logits of both robust and non-robust classifiers through nonlinear transformations optimized with an efficient algorithm, improving accuracy and robustness without additional training.
• Results: Achieved significant improvements in clean accuracy (e.g., +7.86 points on CIFAR-100) with minimal loss in robust accuracy (-0.87 points).

Technical Details

Technological frameworks used: nan

Models used: Robust classifier, standard non-robust classifier

Data used: CIFAR-10, CIFAR-100, and ImageNet datasets

Potential Impact

Cybersecurity firms, AI defense solutions, tech companies leveraging AI for image processing or classification tasks

Want to implement this idea in a business?

We have generated a startup concept here: Confidensify.