Year: 2024
Type: conference-paper
Abstract: Artificial intelligence systems are increasingly used in security-critical applications with limited computing resources, which makes them vulnerable to such disturbances as adversarial attack noise, out-of-distribution data, and fault injections. To absorb disturbances and adapt the AI system during its life cycle, it is necessary to expand the Machine learning operations structure with stages related to the implementation of resilience mechanisms. In this case, increasing resilience in one form or another is associated with the introduction of redundancy in the form of additional resources to absorb disturbances and quickly recover performance. It is proposed to provide Affordable Resilience for resource-constrained AI systems by implementing a resilience optimization stage and adding add-ons with a small number of parameters that will allow for uncertainty calibration and rapid adaptation to labeled and unlabeled data. This approach is also intended to separate the work related to the development and deployment of the basic AI model that solves the applied problem and the work related to ensuring resilience in the Machine learning operations structure. The experiments were conducted on CIFAR-10 and CIFAR-100 datasets using the MobileViT network, which is a modern network architecture for visual image analysis in conditions of limited computing resources. We have experimentally confirmed the increase in the resilience of an AI system at different stages of its life cycle by implementing the stages of resilience optimization, uncertainty calibration, and Test-Time Adaptation. Post-hoc resilience optimization improved robustness to fault injection by 5% and robustness to adversarial attack by 7%. Moreover, tuning with 10% of the test data allowed for an additional 6% increase in robustness to fault injection and 7.1% increase in robustness to adversarial attack on the new data. Also, the use of post-hoc resilience optimization increased the integral indicator of resilience to task changes by 10.5%. Post-hoc uncertainty calibration makes it possible to further increase the robustness of fault injection models by an average of 4.4% and the robustness to adversarial attacks by an average of 1.3%. Test-Time Adaptation increases robustness to Fault Injection by 6.9% and robustness to Adversarial Attack by 4.72%.
Institution: Sumy State University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W7166394157