Data Poisoning Attack against Neural Network-Based On-Device Learning Anomaly Detector by Physical Attacks on Sensors

Area of Science:

• Artificial Intelligence
• Machine Learning Security
• Industrial IoT

Background:

• Edge AI systems for industrial anomaly detection are vulnerable to physical attacks.
• On-device learning models can be compromised through data poisoning, degrading performance.
• Risk assessments for physical attacks on Edge AI are limited.

Purpose of the Study:

• To demonstrate a data poisoning attack on an on-device learning Edge AI for anomaly detection.
• To investigate the security risks associated with physical tampering of Edge AI training data.
• To evaluate the impact of data poisoning on the accuracy of machine vibration anomaly detection.

Main Methods:

• An on-device learning anomaly detection system using MEMS accelerometers for factory machine vibration analysis was targeted.
• A data poisoning attack was executed by exposing the MEMS accelerometer to specific acoustic waves, falsifying vibration data.
• The falsified data was used to retrain the anomaly detection model, which included concept drift detection and multiple normal pattern models.

Main Results:

• The data poisoning attack successfully degraded the Edge AI's ability to detect abnormal machine conditions.
• Falsified acceleration data, when used for training, rendered the anomaly detector ineffective.
• The system failed to identify abnormal states after being subjected to the acoustic manipulation attack.

Conclusions:

• Physical attacks, specifically data poisoning via acoustic manipulation, pose a significant threat to the integrity of on-device learning Edge AI systems.
• The accuracy of anomaly detection in industrial settings can be severely compromised by tampering with sensor data.
• Understanding these vulnerabilities is crucial for developing robust security countermeasures for Edge AI in critical infrastructure.