Reproduce “On the Effect of Pruning on Adversarial Robustness” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Artur Jordao, Helio Pedrini
Research from the University of Campinas demonstrates that pruning convolutional network structures, such as filters and layers, enhances adversarial robustness by up to 3.68 percentage points for FGSM attacks on ResNet56, while concurrently maintaining or slightly improving generalization on clean images. This method provides a computationally efficient defense using only clean images during pruning and fine-tuning.
Key insights
- Network pruning can effectively serve as an inherent defense mechanism against adversarial attacks, extending its utility beyond mere network compression.
- A weak correlation exists between generalization ability and adversarial robustness in pruned networks, suggesting that optimizing for one does not automatically enhance the other.
- Pruning can result in shallower, yet more robust, networks, which challenges previous assumptions regarding the necessity of very deep architectures for effective adversarial defense.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.