Abstract
We propose two training techniques for improving the robustness of Neural Networks to adversarial attacks—manipulations of the inputs that are maliciously crafted to fool networks into incorrect predictions. Both methods are independent of the chosen attack and leverage random projections of the original inputs, with the purpose of exploiting both dimensionality reduction and some characteristic geometrical properties of adversarial perturbations. The first technique is called RP-Ensemble and consists of an ensemble of networks trained on multiple projected versions of the original inputs. The second one, named RP-Regularizer, adds instead a regularization term to the training objective.
Introduction
Deep learning models are notoriously brittle. A tiny, often invisible-to-humans tweak to an image—an "adversarial perturbation"—can trick a state-of-the-art network into misclassifying a stop sign as a speed limit sign. The current gold standard for defense is "adversarial training," where the model is shown millions of these fake examples during training. However, this often becomes a cat-and-mouse game: models become resilient to specific known attacks but remain vulnerable to new ones. Furthermore, as input dimensionality increases, there are more "hiding spots" for these malicious tweaks. This paper addresses the need for a general, efficient defense that does not require prior knowledge of the attacker's strategy.
Proposed Methods
The authors propose two distinct strategies based on Random Projections (RP)—a technique that squashes high-dimensional data into a smaller space while preserving the distances between points.
RP-Ensemble
Instead of training one giant model on the original data, the authors project the data into several different lower-dimensional "subspaces." They train a separate, smaller classifier for each subspace. To make a final prediction, they combine the votes from all these sub-classifiers with a pre-trained model. Because each sub-classifier only sees a specific "slice" of the data, it is statistically difficult for a single adversarial tweak to fool all of them simultaneously.
RP-Regularizer
This fine-tuning approach adds a mathematical penalty (a regularizer) to the training objective. This penalty forces the model to maintain stable gradients across different random projections. Essentially, it teaches the model that small changes in the input should not lead to wild swings in the output, effectively "smoothing" the decision boundaries where attackers usually hide.
Key Insights and Geometry
- The Manifold Hypothesis: High-dimensional data (like a 1024-pixel image) typically lives on a much simpler, low-dimensional "manifold." Adversarial attacks often try to push data off this manifold.
- Codimension and Projection: Most attacks exploit directions that are "normal" (perpendicular) to the data manifold. By randomly selecting directions via projections, there is a high statistical probability of "catching" and neutralizing these adversarial features.
- Attack Independence: Because these methods rely on the inherent geometry of the data rather than simulating specific attacks, they provide a more generalized defense.
Experimental Results
The researchers evaluated their methods on the MNIST (handwritten digits) and CIFAR-10 (color images) benchmarks against a variety of attacks, including FGSM, PGD, DeepFool, and C&W.
- Broad Robustness: On MNIST, the RP-Ensemble significantly outperformed the baseline. Against the PGD attack, where baseline accuracy dropped to near 0%, the RP-Ensemble maintained over 50% accuracy.
- Competitive Performance: The RP-Regularizer achieved results on par with state-of-the-art (SOTA) models on MNIST, particularly against sophisticated attacks like DeepFool and Carlini & Wagner (C&W).
- Efficiency: Training the RP-Ensemble can be faster than training a standard model because the input data for each sub-classifier is much smaller, and the workload can be parallelized across multiple processors.
Limitations and Future Work
While results on MNIST were stellar, performance on the more complex CIFAR-10 dataset was less striking. The RP-Regularizer, in particular, struggled to maintain the same level of boost on color images compared to simple digits. Future work will focus on optimizing the RP-Regularizer's efficiency and exploring deeper connections between random projections and the geometric "shape" of adversarial regions in high-dimensional space.
Summary updated