Universal adversarial perturbations

Abstract

Given a state-of-the-art deep neural network classifier, we show the existence of a universal (image-agnostic) and very small perturbation vector that causes natural images to be misclassified with high probability. We propose a systematic algorithm for computing universal perturbations, and show that state-of-the-art deep neural networks are highly vulnerable to such perturbations, albeit being quasi-imperceptible to the human eye. We further empirically analyze these universal perturbations and show, in particular, that they generalize very well across neural networks. The surprising existence of universal perturbations reveals important geometric correlations among the high-dimensional decision boundary of classifiers. It further outlines potential security breaches with the existence of single directions in the input space that adversaries can possibly exploit to break a classifier on most natural images.

Accepted at IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017

References in corpus (5)

• Very Deep Convolutional Networks for Large-Scale Image Recognition
• Explaining and Harnessing Adversarial Examples
• Caffe: Convolutional Architecture for Fast Feature Embedding
• Going Deeper with Convolutions
• Return of the Devil in the Details: Delving Deep into Convolutional Nets

Cited by in corpus (15)

• Adversarial Examples, Uncertainty, and Transfer Testing Robustness in Gaussian Process Hybrid Deep Networks
• Adversarial Examples that Fool Detectors
• Fast Feature Fool: A data independent approach to universal adversarial perturbations
• UPSET and ANGRI : Breaking High Performance Image Classifiers
• Blocking Transferability of Adversarial Examples in Black-Box Learning Systems
• Exploring the Space of Black-box Attacks on Deep Neural Networks
• Standard detectors aren't (currently) fooled by physical adversarial stop signs
• Confidence estimation in Deep Neural networks via density modelling
• Art of singular vectors and universal adversarial perturbations
• LatentPoison - Adversarial Attacks On The Latent Space
• Verification of Binarized Neural Networks via Inter-Neuron Factoring
• HeNet: A Deep Learning Approach on Intel${}^{\circledR }$ Processor Trace for Effective Exploit Detection
• Butterfly Effect: Bidirectional Control of Classification Performance by Small Additive Perturbation
• Adequacy of the Gradient-Descent Method for Classifier Evasion Attacks
• Where Classification Fails, Interpretation Rises