Adversarial Detection without Model Information
share
Most prior state-of-the-art adversarial detection works assume that the underlying vulnerable model is accessible, i,e., the model can be trained or its outputs are visible. However, this is not a practical assumption due to factors like model encryption, model information leakage and so on. In this work, we propose a model independent adversarial detection method using a simple energy function to distinguish between adversarial and natural inputs. We train a standalone detector independent of the underlying model, with sequential layer-wise training to increase the energy separation corresponding to natural and adversarial inputs. With this, we perform energy distribution-based adversarial detection. Our method achieves state-of-the-art detection performance (ROC-AUC > 0.9) across a wide range of gradient, score and decision-based adversarial attacks on CIFAR10, CIFAR100 and TinyImagenet datasets. Compared to prior approaches, our method requires 10-100x less number of operations and parameters for adversarial detection. Further, we show that our detection method is transferable across different datasets and adversarial attacks. For reproducibility, we provide code in the supplementary material.
READ FULL TEXT
