Random Spiking and Systematic Evaluation of Defenses Against Adversarial Examples

Huangyi Ge, Sze Yiu Chau, Bruno Ribeiro, Ninghui Li

Introduction

Adversarial examples—small, often imperceptible perturbations to input data—pose a significant threat to the reliability of deep learning models. Despite numerous proposed defenses, many are quickly circumvented by more sophisticated attacks. This paper addresses the fragility of current models by introducing Random Spiking (RS), a novel defense mechanism, and a rigorous Translucent-Box evaluation framework designed to simulate realistic, resourceful attackers.

Random Spiking: Injecting Resilience

Random Spiking (RS) is a training-time defense that generalizes the concept of Dropout. While Dropout sets activations to zero, RS replaces them with random noise sampled from the activation range of the specific layer.

By training with this stochasticity, the network learns to maintain stable internal representations despite perturbations. During inference, a Monte Carlo approach is employed: the input is processed multiple times through the noisy network, and the results are averaged. This process effectively approximates a model ensemble, enhancing robustness. The authors found that injecting noise early in the network (e.g., after the first convolution) is particularly effective, as subsequent layers can learn to filter or ignore the noise.

The Translucent-Box Evaluation Framework

To move beyond the "cat-and-mouse" cycle of weak evaluations, the authors propose the Translucent-Box framework. This middle ground assumes the attacker knows the model architecture and training data but lacks the specific trained parameters.

To simulate a potent adversary, the framework uses a Multi-surrogate with Validation strategy:

  1. Generation: An attacker trains 16 surrogate models and generates adversarial examples that successfully fool at least 8 of them.
  2. Validation: These examples are then tested against 7 additional "validation" surrogates. Only those that fool at least 5 validation models are used against the target defense.

This ensures that the defense is tested against highly transferable and robust adversarial examples, rather than "naive" attacks that exploit single-model artifacts.

Experimental Results and Model Stability

The researchers evaluated RS against standard benchmarks: MNIST, Fashion-MNIST, and CIFAR-10. A key finding is that robust models should also exhibit stability against non-adversarial noise, such as Gaussian blur or JPEG compression.

Model stability under Gaussian noise across MNIST, Fashion-MNIST, and CIFAR-10.

As shown in Figure 1, models utilizing Random Spiking combined with Adversarial Training (RS-1-Adv) maintain high prediction stability even as the amount of Gaussian noise increases. In contrast, other defenses often show a more rapid decline in performance.

Prediction stability on Fashion-MNIST under varying JPEG compression quality.

The advantage of RS is even more pronounced under JPEG compression. On Fashion-MNIST (Figure 2), RS-based models maintain approximately 85% stability even at low quality settings. Other defenses, such as MagNet, experience significant drops in stability, falling to 60% or lower.

Key Insights

  • Randomness as a Shield: By increasing the stochasticity of the training process, RS makes the "blind spots" of a model unique to its specific training run, significantly hindering the transferability of adversarial attacks.
  • Superior Defense: The RS-1-Adv configuration consistently outperformed existing methods like defensive distillation, requiring significantly higher $L_2$ distortion for successful label flipping on CIFAR-10.
  • Breaking Naive Attacks: The study demonstrated that multi-surrogate attacks are far more effective than single-surrogate ones, highlighting the need for more rigorous evaluation standards in the field.

Conclusion and Future Directions

Random Spiking offers a promising path toward more robust neural networks by leveraging internal randomness and ensemble-like inference. While the computational cost of Monte Carlo averaging and the limitations of $L_p$ distance metrics remain challenges, the principles of RS and the Translucent-Box framework provide a more realistic and rigorous foundation for developing and testing adversarial defenses. Future work may extend these concepts to other domains, such as audio or natural language processing.

Summary updated

Abstract

Image classifiers often suffer from adversarial examples, which are generated by strategically adding a small amount of noise to input images to trick classifiers into misclassification. Over the years, many defense mechanisms have been proposed, and different researchers have made seemingly contradictory claims on their effectiveness. We present an analysis of possible adversarial models, and propose an evaluation framework for comparing different defense mechanisms. As part of the framework, we introduce a more powerful and realistic adversary strategy. Furthermore, we propose a new defense mechanism called Random Spiking (RS), which generalizes dropout and introduces random noises in the training process in a controlled manner. Evaluations under our proposed framework suggest RS delivers better protection against adversarial examples than many existing schemes.
BibTeX
@misc{ge2018random,
  author = {Huangyi Ge and Sze Yiu Chau and Bruno Ribeiro and Ninghui Li},
  title = {{Random Spiking and Systematic Evaluation of Defenses Against Adversarial Examples}},
  year = {2018},
  eprint = {1812.01804},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1812.01804},
}
Figure