Fooling the classifier: Ligand antagonism and adversarial examples

Thomas J. Rademaker, Emmanuel Bengio, Paul FrançoisarXiv 2018 · DOI · OpenReview

Abstract: Machine learning algorithms can be fooled by small well-designed adversarial perturbations. This is reminiscent of cellular decision-making where ligands (called antagonists) prevent correct signalling, like in early immune recognition. We draw a formal analogy between neural networks used in machine learning and models of cellular decision-making (adaptive proofreading). We apply attacks from machine learning to simple decision-making models, and show explicitly the correspondence to antagonism by weakly bound ligands. Such antagonism is absent in more nonlinear models, which inspired us to implement a biomimetic defence in neural networks filtering out adversarial perturbations. We then apply a gradient-descent approach from machine learning to different cellular decision-making models, and we reveal the existence of two regimes characterized by the presence or absence of a critical point for the gradient. This critical point causes the strongest antagonists to lie close to the decision boundary. This is validated in the loss landscapes of robust neural networks and cellular decision-making models, and observed experimentally for immune cells. For both regimes, we explain how associated defence mechanisms shape the geometry of the loss landscape, and why different adversarial attacks are effective in different regimes. Our work connects evolved cellular decision-making to machine learning, and motivates the design of a general theory of adversarial perturbations, both for in vivo and in silico systems.

Introduction: The Parallel Between AI and Biology

Both artificial neural networks and biological cells function as complex classifiers. An image recognition model must distinguish between a "panda" and a "gibbon," while a T-cell in the immune system must classify a biological molecule (a ligand) as "self" (harmless) or "non-self" (a threat). In machine learning, classifiers are known to have blind spots: adding tiny, often invisible perturbations to an image—an adversarial attack—can completely fool the network.

Figure 1: Comparison of adversarial attacks in machine learning (MNIST digits and traffic signs) with biological ligand binding times and receptor responses.

Surprisingly, a similar phenomenon occurs in biology. The immune system can be fooled by "antagonists"—numerous weakly binding ligands that suppress the cell's ability to detect an actual threat. Viruses like HIV and certain cancers exploit this to escape immune detection. This research formally bridges these two fields, demonstrating that the mathematical underpinnings of adversarial attacks are identical to biological ligand antagonism.

Methodology: Bridging Neural Networks and Adaptive Proofreading

The authors established a formal mathematical analogy between a simple neural network and biophysical models of cellular decision-making, specifically adaptive proofreading models. To explore this link, they applied the Fast Gradient Sign Method (FGSM)—a standard machine learning technique for generating adversarial noise—to biological models to observe how an "attack" manifests in a biochemical context.

Conversely, they looked to biology for defense strategies. Inspired by kinetic proofreading—a biological mechanism that filters out weak, short-lived chemical signals—they implemented a "biomimetic defense" for neural networks. This involved applying a non-linear Hill function to suppress weak pixel values (mimicking receptor sensitivity) and local mean-filtering (mimicking receptor clustering and signal averaging over time).

Key Results: Biomimetic Defenses and Robustness

The study revealed that an FGSM attack in machine learning is mathematically equivalent to ligand antagonism in biology. Both exploit the classifier's linear aggregation of weak features. However, by implementing the biological proofreading defense, the neural network successfully filtered out FGSM perturbations, maintaining accuracy where a naive network failed.

Figure 3: Performance of naive vs. robust classifiers on interpolated digits and ligand mixtures, showing how biomimetic defenses force adversarial examples toward meaningful ambiguity.

In the biological model, increasing the proofreading parameter ($m$) caused a sudden phase transition in robustness. For $m \ge 2$, the system became highly resistant to small noise, requiring 3 to 4 orders of magnitude more effort for an "attacker" to reach the decision boundary. Crucially, when these robust systems are finally fooled, the adversarial examples are no longer invisible noise; they become genuinely ambiguous. For instance, an attacked "6" transitioning toward a "9" ends up looking like a hybrid "5" to a human observer, just as a maximally antagonizing ligand acts as a borderline weak agonist in biological systems.

Discussion and Future Directions

This work demonstrates that evolved cellular decision-making can provide a blueprint for more robust artificial intelligence. While this study focused on binary classification and simple models, the implications for complex systems are significant. Future research may involve mapping high-dimensional data into lower-dimensional latent spaces to apply these biological analogies to more complex images.

Furthermore, these findings have vital implications for cancer immunotherapy. Tumors often produce mutated ligands (neoantigens) that act as "ambiguous" adversarial attacks against the immune system. Understanding these biological adversarial strategies could help predict the success of personalized therapies, such as checkpoint inhibitors, which effectively "remove the blindfold" from the immune classifier.

Summary updated

Abstract

Machine learning algorithms can be fooled by small well-designed adversarial perturbations. This is reminiscent of cellular decision-making where ligands (called antagonists) prevent correct signalling, like in early immune recognition. We draw a formal analogy between neural networks used in machine learning and models of cellular decision-making (adaptive proofreading). We apply attacks from machine learning to simple decision-making models, and show explicitly the correspondence to antagonism by weakly bound ligands. Such antagonism is absent in more nonlinear models, which inspired us to implement a biomimetic defence in neural networks filtering out adversarial perturbations. We then apply a gradient-descent approach from machine learning to different cellular decision-making models, and we reveal the existence of two regimes characterized by the presence or absence of a critical point for the gradient. This critical point causes the strongest antagonists to lie close to the decision boundary. This is validated in the loss landscapes of robust neural networks and cellular decision-making models, and observed experimentally for immune cells. For both regimes, we explain how associated defence mechanisms shape the geometry of the loss landscape, and why different adversarial attacks are effective in different regimes. Our work connects evolved cellular decision-making to machine learning, and motivates the design of a general theory of adversarial perturbations, both for in vivo and in silico systems.
BibTeX
@misc{rademaker2018fooling,
  author = {Thomas J. Rademaker and Emmanuel Bengio and Paul François},
  title = {{Fooling the classifier: Ligand antagonism and adversarial examples}},
  year = {2018},
  eprint = {1807.04270},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1807.04270},
}
Figure