Random Projections for Improved Adversarial Robustness

Ginevra Carbone, Guido Sanguinetti, Luca Bortolussi

TL;DR

Researchers have developed two new training techniques—RP-Ensemble and RP-Regularizer—that use random mathematical projections to shield neural networks from adversarial attacks. Unlike traditional defenses that study specific "enemy" tactics, these methods are attack-agnostic, leveraging the geometry of data to make models naturally more resilient and computationally efficient.

What Problem Does This Paper Address?

Deep learning models are notoriously brittle. A tiny, 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 you show the model millions of these fake examples during training. However, this is like a cat-and-mouse game: the model often becomes great at defending against the specific attacks it saw but remains vulnerable to new ones. Furthermore, as the input space gets larger (higher dimensionality), there are more "hiding spots" for these malicious tweaks. This paper addresses the need for a general, efficient defense that doesn't require knowing the attacker's strategy in advance.

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.

  1. 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 "slice" of the data, it's harder for a single adversarial tweak to fool all of them at once.
  2. RP-Regularizer: This is a "fine-tuning" approach. They add a new mathematical penalty (a regularizer) to the training objective. This penalty forces the model to have stable gradients across different random projections. Essentially, it teaches the model that small changes in the input shouldn't lead to wild swings in the output, effectively "smoothing" the decision boundaries where attackers usually hide.

Datasets

The researchers evaluated their methods on two classic computer vision benchmarks:

  • MNIST: A dataset of 70,000 small images of handwritten digits (0-9).
  • CIFAR-10: A more complex dataset of 60,000 color images across 10 categories (like airplanes, dogs, and ships).

They tested these against a "rogue's gallery" of famous attacks, including FGSM, PGD, DeepFool, and C&W.

Key Insights and Ideas

  • The Manifold Hypothesis: The authors lean on the idea that high-dimensional data (like a 1024-pixel image) actually lives on a much simpler, low-dimensional "manifold" (like a curved sheet of paper in a 3D room). Adversarial attacks often try to push data off this sheet.
  • Codimension is the Key: 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" 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.
  • Efficiency through Parallelism: Since the sub-classifiers in the RP-Ensemble are independent, they can be trained simultaneously on different processors, making the defense very scalable.

Key Results

  • Broad Robustness: On MNIST, the RP-Ensemble significantly outperformed the baseline across all attacks. For example, against the PGD attack, the baseline accuracy dropped to near 0%, while 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).
  • Speed: Training the RP-Ensemble can actually be faster than training a standard model because the input data for each sub-classifier is much smaller, and the work can be split up (parallelized).

Limitations and Future Work

While the results on MNIST were stellar, the authors admit the results on the more complex CIFAR-10 dataset were "less striking." The RP-Regularizer, in particular, struggled to maintain the same level of boost on color images compared to simple digits.

The authors suggest that the trade-off between computational cost and accuracy needs more fine-tuning. 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 the data space.

Summary updated

Abstract

We propose two training techniques for improving the robustness of Neural Networks to adversarial attacks, i.e. 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.
BibTeX
@inproceedings{carbone2021random,
  author = {Ginevra Carbone and Guido Sanguinetti and Luca Bortolussi},
  title = {{Random Projections for Improved Adversarial Robustness}},
  booktitle = {IJCNN},
  year = {2021},
  url = {https://openreview.net/forum?id=jgjaNp8_LtR},
}

Loading related content...

Figure