Detection of Adversarial Attacks and Characterization of Adversarial\n Subspace

2019/10/26 by Mohammad Esmaeilpour, Esmaeilpour, Mohammad, Patrick Cardinal +3 · 1 citation
Computer Science · #Anomaly Detection Techniques and Applications #Adversarial Robustness in Machine Learning #Digital Media Forensic Detection

paper · pdf · doi:10.48550/arxiv.1910.12084

Abstract

Adversarial attacks have always been a serious threat for any data-driven\nmodel. In this paper, we explore subspaces of adversarial examples in unitary\nvector domain, and we propose a novel detector for defending our models trained\nfor environmental sound classification. We measure chordal distance between\nlegitimate and malicious representation of sounds in unitary space of\ngeneralized Schur decomposition and show that their manifolds lie far from each\nother. Our front-end detector is a regularized logistic regression which\ndiscriminates eigenvalues of legitimate and adversarial spectrograms. The\nexperimental results on three benchmarking datasets of environmental sounds\nrepresented by spectrograms reveal high detection rate of the proposed detector\nfor eight types of adversarial attacks and outperforms other detection\napproaches.\n

Cited by

• Diffusion Models are Robust Pretrainers

Related

• A Robust Approach for Securing Audio Classification Against Adversarial\n Attacks
• Subspace Defense: Discarding Adversarial Perturbations by Learning a Subspace for Clean Signals
• AdvEst: Adversarial Perturbation Estimation to Classify and Detect Adversarial Attacks against Speaker Identification
• The Adversarial Attack and Detection under the Fisher Information Metric
• Adversarial Attacks in Sound Event Classification