Unsupervised Anomaly Detection for Improving Adversarial Robustness of 3D Object Detection Models

• DOI: 10.3390/electronics14020236
• Publication Date: 2025-01-08T09:54:08Z
• Authors: Mumuxin Cai; Xupeng Wang; Ferdous Sohel; Hang Lei
• Keywords: Robustness

Three-dimensional object detection based on deep neural networks (DNNs) is widely used in safety-related applications, such as autonomous driving. However, existing research has shown that 3D models are vulnerable to adversarial attacks. Hence, the improvement robustness of under attacks investigated this work. A autoencoder-based anomaly method proposed, which a strong ability detect elaborate samples an unsupervised way. The proposed operates given Light Detection and Ranging (LiDAR) scene its Bird’s Eye View (BEV) image reconstructs through autoencoder. To improve performance autoencoder, augmented memory module with typical normal patterns recorded introduced. It designed help model amplify reconstruction errors malicious negligibly affected. Experiments several public datasets show achieves AUC 0.8 improves detection.