Abstract
Recent studies show that deep neural networks are extremely vulnerable, especially for adversarial examples of image classification models. However, existing defenses suffer from limited adaptability across attacks, an unfavorable tradeoff between clean accuracy and robustness, and substantial training-time overhead. To tackle these problems, we present a novel component, named the redundant fully connected layer, which can be combined with existing model backbones in a pluggable manner. Specifically, we design a tailor-made loss function for it that leverages cosine similarity to maximize the difference and diversity of multiple fully connected parts. We conduct extensive experiments against 12 representative attacks (white-box and black-box), based on two popular datasets. The empirical evaluations show that our scheme realizes significant outcomes against various attacks with negligible additional training overhead, while hardly bringing collateral damage for clean-instance accuracy.