Generating imperceptible adversarial examples via low-frequency aware transfer attacks on battery management systems.

Area of Science:

• Electrical Engineering
• Computer Science
• Artificial Intelligence

Background:

• Deep learning models are crucial for state-of-health (SoH) estimation in lithium-ion battery management systems (BMS).
• These models face security risks due to susceptibility to adversarial attacks, potentially causing misdiagnosis and increased maintenance.
• The practical robustness of these models against adversarial perturbations is not well understood.

Purpose of the Study:

• To develop a novel adversarial attack to assess the security risks of deep learning models in BMS.
• To evaluate model robustness in scenarios with limited or no access to model information.
• To understand potential vulnerabilities in real-world BMS applications.

Main Methods:

• Developed a novel adversarial attack targeting deep learning models for SoH estimation.
• Exploited frequency information within signals to capture unique SoH characteristics.
• Generated adversarial examples by manipulating low-frequency signal components while preserving signal relationships.
• Assessed model robustness in black-box scenarios (inaccessible model information).

Main Results:

• Demonstrated the effectiveness of the proposed adversarial attack in uncovering risks.
• Showcased the ability to generate imperceptible adversarial examples.
• Highlighted the vulnerability of SoH estimation models to targeted frequency manipulation.
• Provided insights into the practical security implications for BMS.

Conclusions:

• Deep learning models in BMS are susceptible to novel adversarial attacks that exploit frequency characteristics.
• The developed attack effectively reveals potential security risks in practical, information-limited scenarios.
• Further research is needed to enhance the robustness of BMS against such adversarial threats.