Towards Safe Deep Learning: Unsupervised Defense Against Generic Adversarial Attacks

Bita Darvish Rouhani, Mohammad Samragh, Tara Javidi, Farinaz Koushanfar

Abstract: Recent advances in adversarial Deep Learning (DL) have opened up a new and largely unexplored surface for malicious attacks jeopardizing the integrity of autonomous DL systems. We introduce a novel automated countermeasure called Parallel Checkpointing Learners (PCL) to thwart the potential adversarial attacks and significantly improve the reliability (safety) of a victim DL model. The proposed PCL methodology is unsupervised, meaning that no adversarial sample is leveraged to build/train parallel checkpointing learners. We formalize the goal of preventing adversarial attacks as an optimization problem to minimize the rarely observed regions in the latent feature space spanned by a DL network. To solve the aforementioned minimization problem, a set of complementary but disjoint checkpointing modules are trained and leveraged to validate the victim model execution in parallel. Each checkpointing learner explicitly characterizes the geometry of the input data and the corresponding high-level data abstractions within a particular DL layer. As such, the adversary is required to simultaneously deceive all the defender modules in order to succeed. We extensively evaluate the performance of the PCL methodology against the state-of-the-art attack scenarios, including Fast-Gradient-Sign (FGS), Jacobian Saliency Map Attack (JSMA), Deepfool, and Carlini&WagnerL2 algorithm. Extensive proof-of-concept evaluations for analyzing various data collections including MNIST, CIFAR10, and ImageNet corroborate the effectiveness of our proposed defense mechanism against adversarial samples.

TL;DR: Devising unsupervised defense mechanisms against adversarial attacks is crucial to ensure the generalizability of the defense.

Keywords: Adversarial Attacks, Unsupervised Defense, Deep Learning

Data: [CIFAR-10](https://paperswithcode.com/dataset/cifar-10), [ImageNet](https://paperswithcode.com/dataset/imagenet), [MNIST](https://paperswithcode.com/dataset/mnist)