Integrating Evidence Accumulation with Multi-dimensional mmWave Features for Robust Object Recognition

Deep learning (DL) models are widely used in mmWave-based object recognition, but the sensitivity of traditional DL models to adversarial attacks limits their application in safety-critical scenarios. Existing defense methods based on generative adversarial networks (GANs) target specific attacks and lack generalization capabilities. Inspired by the decision-making principles of the human brain, we introduce a novel defense framework into the mmWave-based object recognition task, which integrates test-phase dropout and an evidence accumulation mechanism to transform the traditional single-shot decision-making process into a dynamic evidence collection process, thereby effectively defending against different types of adversarial attacks. In addition, we use multi-dimensional mmWave features at different scales to make multi-stage decisions, which improves the recognition ability of the model in complex signal environments. Experimental results show that our method enhances the robustness of the mmWave recognition model under adversarial interference, making it more suitable for practical applications.