TL;DR
This paper uses a simplified "adversarial spheres" toy dataset to rigorously test how different Bayesian inference methods handle adversarial examples. The authors demonstrate that while accurate Bayesian methods like MCMC can successfully flag adversarial points as uncertain, common approximations—such as the Laplace approximation or bootstrap ensembling—often fail to provide reliable uncertainty estimates in high-dimensional, noiseless settings.
What Problem Does This Paper Address?
Deep neural networks are notoriously overconfident, often assigning near-100% probability to "adversarial examples"—input points that have been subtly manipulated to cause misclassification. Bayesian Neural Networks (BNNs) are theoretically a great solution because they capture epistemic uncertainty (uncertainty about which model is correct). In theory, a BNN should be uncertain in regions of the input space where it hasn't seen much training data.
However, analyzing BNNs is difficult because they are complex and high-dimensional. To cut through the noise, the authors use the adversarial spheres setup: a dataset of two concentric hyperspheres where the "true" decision boundary is known to be a sphere. This creates a noiseless environment where the only challenge is capturing the model's uncertainty correctly in high dimensions ($D=500$).
Methods
The authors apply Bayesian Logistic Regression to squared features, which allows a linear model to learn the spherical decision boundaries. They compare several inference strategies:
- MAP (Maximum A Posteriori): A standard non-Bayesian baseline with weight regularization.
- Bootstrap Ensembling: Training multiple models on different subsets of data, a common "poor man's Bayesian" approach.
- Laplace Approximation: Fitting a Gaussian distribution to the posterior by looking at the curvature at the mode.
- MCMC (Markov Chain Monte Carlo): Specifically slice sampling, considered the "gold standard" for accurate posterior estimation.
- Variational Inference (SVI): A faster approximation that fits a simpler distribution (like a Gaussian) to the posterior.
- Hierarchical Models: A more advanced approach that uses priors to exploit the known symmetry of the problem (i.e., the fact that the decision boundary is a sphere).
Datasets
The study uses the Adversarial Spheres dataset (Gilmer et al., 2018). It consists of two concentric hyperspheres in 500-dimensional space. Points are distributed uniformly on the surface of each sphere. The setup is designed so that even if a model achieves 100% accuracy on a massive validation set, an optimizer can still easily find adversarial points on the sphere surface that the model confidently misclassifies.
Key Insights and Ideas
- The Failure of Laplace: The authors found that the Laplace approximation performs poorly here. Because the data is linearly separable and the prior is broad, the posterior distribution has a "flat top" rather than a sharp peak. Fitting a Gaussian to this flat region results in an unrealistically wide approximation that makes the model uncertain about everything, including valid data points.
- Bootstrap is Insufficient: Bootstrap ensembles remained highly confident on adversarial examples. The adversarial optimization procedure simply learned to exploit the shared blind spots of the entire ensemble.
- Symmetry as a Defense: The most robust results came from using a hierarchical prior that encouraged the model to exploit the symmetry of the problem. If the model "knows" the boundary should be spherical, it becomes much harder to fool.
- Variational Trade-offs: Standard Variational Inference (SVI) was surprisingly good at matching MCMC's predictions, but it tended to rule out the very large weight values needed to represent the "true" steepest decision boundaries.
Key Results
The quantitative results, summarized in the table below, highlight the gap between standard methods and sophisticated Bayesian inference.
- MAP was nearly 100% confident in its wrong predictions (Adversarial Error of 0.999).
- MCMC significantly reduced this error to 0.558, meaning it flagged many adversarial points as uncertain.
- MCMC with a non-zero mean prior (exploiting symmetry) achieved the best performance, dropping adversarial error to 0.341 while maintaining perfect confidence on validation data.
- Laplace failed completely, with an average confidence of only 0.501, essentially guessing randomly on all points.
Limitations and Future Work
The primary limitation is that this study focuses on a "toy" linear model (logistic regression on squared features) rather than deep, multi-layer neural networks. While this allows for precise analysis, the authors note that the challenges of choosing the right priors and handling non-Gaussian posteriors will only intensify in deeper models.
Future work involves exploring families of priors that can capture more complex symmetries and invariances in real-world datasets, which the authors believe is the key to making Bayesian methods truly robust against adversarial attacks in deep learning.
Summary updated