Giving Neural Networks an Immune System



By Simon Morrow
Figure of Adaptive/Computational Immune Systems

Ren Wang, an assistant professor in the Department of Electrical and Computer Engineering at Illinois Institute of Technology, has received a Ralph E. Powe Junior Faculty Enhancement Award from Oak Ridge Associated Universities (ORAU) for his research using insights from the human adaptive immune system to make artificial intelligence systems more resilient. 

The use of AI in everyday applications is increasing rapidly, such as natural language processing by ChatGPT and facial recognition by smartphones. But the black box nature of these systems has made it difficult to develop techniques for keeping them safe from errors or attacks. 

From humans making themselves invisible to AI using specially designed T-shirts to autonomous cars crashing, AI’s proliferation in real-world applications has led to real consequences when an AI system makes an error or encounters something it wasn’t trained to deal with. 

“This is a very, very dangerous phenomenon,” says Wang. “A very small perturbation can lead to a totally wrong answer. For example, the AI model might predict a speed limit sign when there is actually a stop sign.”

With neural networks already borrowing from the brain, Wang says that looking to the immune system for defense techniques is a natural fit for making AI models more robust. 
Wang is focusing on certain elements of the adaptive immune system, such as examining how B cells are generated, proliferated, and utilized in the body. 

“B cells generate antibodies to defend against certain types of attack, and that attack could be any antigen,” says Wang. “We hope to learn from this whole process and to capture some important patterns that we can use to improve the AI system.”

Wang is aiming to develop general techniques that could be adopted by a broad range of AI systems, but he is also specifically applying his findings to power system applications driven by AI. These applications, such as power system control and stability analysis, require high levels of robustness. 
Incorporating physical constraints in the adaptive immune-inspired learning approach, initially to simple structures and later to complex power grids, will result in robust models that require less data and have a more refined search space.

“I believe that this is a very novel idea,” says Wang, “and I believe this adaptive immune-inspired system will be very powerful.”

The Powe Award is given annually by ORAU to provide seed money for junior faculty members with great potential for scientific research in their field of study. Wang was one of 35 recipients of the award out of 167 applications to be selected in 2023.   

“It is a tremendous honor to have been selected as one of this year’s awardees, and I am immensely grateful,” says Wang.

Wang also recently received a Computer and Information Science and Engineering Research Initiation Initiative award from the National Science Foundation.

Image: The process of the adaptive immune system developing solutions to remove antigens inspires the computational system model for developing solutions to remove adversaries. Information collected from the power grid and physical constraints are used in the practical implementation of the computational system.