Algorithm Developed in Russia to Automatically Detect Vulnerabilities in Autopilot Neural Networks
The algorithm is primarily designed to improve the reliability of computer vision systems.
Researchers at St. Petersburg Electrotechnical University “LETI” have developed an algorithm for detecting vulnerabilities in neural networks that control autonomous vehicles. One of its priority applications is identifying errors in computer vision systems.
“We created a unified database containing all existing tools for finding vulnerabilities in computer vision models and developed an evolutionary optimization algorithm that automatically tests combinations of methods, adjusts their parameters and selects the configuration that most effectively identifies vulnerabilities in a specific model,” said Alla Levina, head of the Laboratory of Fundamental Principles for Building Intelligent Systems and associate professor at the Department of Information Security at St. Petersburg Electrotechnical University “LETI.”
The researchers note that computer vision enables autonomous vehicles to navigate their surroundings and detect obstacles. However, malicious attacks that apply specially designed filters to images can compromise their safety. The new methodology is designed to address such threats more effectively.
