Adversarial Reprogramming of Neural Cellular Automata
Changing biological system behaviors with machine learning
1Introduction
This article is part of a collection of short articles focusing on differentiable self-organizing systems. These systems aim to discover signaling events that can change system behavior in biological, technological, and social fields.
2Neural Cellular Automata
Neural Cellular Automata (Neural CA) is a model used to train tasks requiring self-organization in an end-to-end differentiable manner. These models are robust to various perturbations and provide quantitative models to understand biological phenomena.
3Adversarial Attacks
In this study, adversarial attacks were trained to reprogram CAs to perform tasks different from what they were trained to do. These attacks are crucial to understanding low-level signals that can change system behavior.
4Cell States and Model Parameters
The system behavior of Neural CA is influenced by individual cell states and model parameters. Cell states are used to diversify cell behaviors and communicate with neighboring cells.
5Perceptive Field
In Neural CA, cells perceive their environment as the eight nearest neighbors and themselves. This perceptive field differs between Growing CA and MNIST CA and significantly affects system behavior.
6Adversarial Cell Injection
The first type of adversarial attacks involves injecting a few adversarial cells into a grid running a pre-trained model. This is an example of injecting CAs with different model parameters to hijack system behavior.
7Global State Perturbation
The second type of adversarial attacks interacts with pre-trained growing CA models by perturbing states within cells. This can be seen as inhibiting or enhancing state values to hijack proper communications among cells.
8Biological Analogies
Such models offer ways to think about adversarial relationships in nature (such as parasitism and evolutionary arms races of genetic and physiological mechanisms) and provide a roadmap for developing regenerative medicine strategies.
9Biomedical Advancements
Understanding how cell groups make collective decisions and how these tissue-level decisions can be subverted by a small group of cells is essential for meaningful progress in regenerative medicine.
10Quantitative Models
Developing quantitative models of these dynamics is necessary to advance regenerative medicine that controls system-level outcomes from the top down.
11Future Steps
Next-generation biomedical research will require computational tools to control infections and other biological processes.
12Conclusion
Adversarial reprogramming of Neural Cellular Automata is a crucial tool for changing biological system behaviors and developing new strategies in regenerative medicine.
Original article: Distill — Sinirsel Hücresel Otomatların Adversaryal Yeniden Programlanması