Combatting Adversarial Attacks through Denoising and Dimensionality\n Reduction: A Cascaded Autoencoder Approach

2018/12/07 by Rajeev Sahay, Sahay, Rajeev, Rehana Mahfuz +3
Computer Science · #Adversarial Robustness in Machine Learning #Anomaly Detection Techniques and Applications #Cryptography and Security (cs.CR) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML)

paper · pdf · doi:10.48550/arxiv.1812.03087

openalex publication_date 2018/12/07 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

Machine Learning models are vulnerable to adversarial attacks that rely on\nperturbing the input data. This work proposes a novel strategy using\nAutoencoder Deep Neural Networks to defend a machine learning model against two\ngradient-based attacks: The Fast Gradient Sign attack and Fast Gradient attack.\nFirst we use an autoencoder to denoise the test data, which is trained with\nboth clean and corrupted data. Then, we reduce the dimension of the denoised\ndata using the hidden layer representation of another autoencoder. We perform\nthis experiment for multiple values of the bound of adversarial perturbations,\nand consider different numbers of reduced dimensions. When the test data is\npreprocessed using this cascaded pipeline, the tested deep neural network\nclassifier yields a much higher accuracy, thus mitigating the effect of the\nadversarial perturbation.\n

Related

• A Computationally Efficient Method for Defending Adversarial Deep Learning Attacks
• Mitigating Gradient-based Adversarial Attacks via Denoising and Compression
• A Generative Model based Adversarial Security of Deep Learning and\n Linear Classifier Models
• Divide, Denoise, and Defend against Adversarial Attacks
• DefensiveDR: Defending against Adversarial Patches using Dimensionality Reduction