Centrala begrepp
Bayesian Neural Networks exhibit robustness to gradient-based adversarial attacks due to the averaging effect of the posterior distribution.
Sammanfattning
The article discusses the vulnerability of deep learning models to adversarial attacks and the potential robustness of Bayesian Neural Networks (BNNs). It analyzes the geometry of adversarial attacks in BNNs and demonstrates that BNN posteriors are robust to gradient-based attacks in the over-parameterized limit. The paper provides theoretical proofs and empirical results supporting the robustness of BNNs to both gradient-based and gradient-free adversarial attacks on various datasets. It also explores the convergence of BNNs to Gaussian Processes and the implications for model robustness in safety-critical applications.
Statistik
"Experimental results on the MNIST, Fashion MNIST, and a synthetic dataset with BNNs trained with Hamiltonian Monte Carlo and Variational Inference support this line of arguments."
"BNNs can display high accuracy on clean data and robustness to both gradient-based and gradient-free adversarial attacks."
Citat
"Vulnerability to adversarial attacks is one of the principal hurdles to the adoption of deep learning in safety-critical applications."
"Experimental results on various BNN architectures trained with Hamiltonian Monte Carlo and with Variational Inference empirically show that BNNs are more robust to both gradient-based and gradient-free attacks than their deterministic counterpart."