MFFDR: An Advanced Multi-branch Feature Fusion and Dynamic Reconstruction Framework for Enhancing Adversarial Robustness

Deep Neural Networks (DNNs) are highly sus-ceptible to adversarial noise, which can lead to erroneous predictions. In high-stakes scenarios, such as autonomous driving and medical diagnosis, DNNs inaccuracies can be dire. To address this issue, Adversarial Training (AT) has been widely adopted as an effective defense method. However, our analysis reveals two critical flaws in the traditional AT approach that hinder its adversarial robustness: (1) focus only on a subset of robust features during the training process. This narrow focus limits the model's ability to learn and perceive a diverse range of features. (2) tend to overlook potential cues in non-robust features that could be beneficial for the model to make correct predictions. These cues, referred to as “positive activations” for simplicity, contain valuable information that can enhance the model's perception and understanding of the input data. In this way, we propose a novel and plug-and-play framework called Multi-branch Feature Fusion and Dynamic Reconstruction (MFFDR), which leverages multi-branch attention mechanisms to enhance the model's perception of robust features and enrich the diversity of learned features. Moreover, we employ a dynamic weighting strategy to reconstruct non-robust features in order to utilize the positive activations embedded within them. Extensive experiments demonstrate that our method significantly improves the model's adversarial robustness and outperforms previous state-of-the-art methods.