A Comprehensive Analysis of Robust Machine Learning Against Adversarial Attacks

This research paper presents a comprehensive analysis of robust machine learning techniques against adversarial attacks, focusing on the vulnerability of machine learning models to small, intentionally designed input perturbations. The study reviews six relevant research works covering PGD Adversarial Training, TRADES, Randomized Smoothing, AutoAttack, Obfuscated Gradients, and the use of generated data for robust training. A reproducible experimental study is conducted using the UCI Optical Recognition of Handwritten Digits dataset, containing 1,797 handwritten digit images. Four approaches are evaluated: Standard Training, PGD Adversarial Training, TRADES, and Gaussian Noise Training with Randomized Smoothing. The models are evaluated using clean accuracy, FGSM accuracy, PGD-20 accuracy, transfer attacks, certified accuracy, and training time across three random seeds. Experimental results demonstrate that the standard model achieves 97.6% clean accuracy but drops to 0.0% under a PGD attack with a perturbation budget of ε = 0.2. In comparison, PGD Adversarial Training and TRADES retain 51.9% and 51.2% accuracy, respectively, under the same attack budget. The study also demonstrates the importance of certified robustness, strong adversarial evaluation, and the trade-off between model accuracy, robustness, and computational cost. This work provides a transparent, low-compute, and reproducible comparative study of adversarial robustness techniques. The results are limited to a small dataset and a simple neural network and should not be interpreted as a large-scale performance benchmark. Author: Arya Anil JaradDepartment: Artificial Intelligence and Data Science EngineeringInstitution: JSPM's Narhe Technical Campus, Pune, Maharashtra, IndiaGuide: Prof. Kalyani Kulkarni Document status: Independent Preprint, Version 1.0, November 2026. Not peer reviewed.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-15
DOI
https://doi.org/10.5281/zenodo.24239857
Primary Topic
Adversarial Robustness in Machine Learning
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

A Comprehensive Analysis of Robust Machine Learning Against Adversarial Attacks

Arya Anil Jarad
Zenodo (CERN European Organization for Nuclear Research)
Adversarial Robustness in Machine Learning
preprint

A Comprehensive Analysis of Robust Machine Learning Against Adversarial Attacks

Arya Anil Jarad
preprint en

Abstract

This research paper presents a comprehensive analysis of robust machine learning techniques against adversarial attacks, focusing on the vulnerability of machine learning models to small, intentionally designed input perturbations. The study reviews six relevant research works covering PGD Adversarial Training, TRADES, Randomized Smoothing, AutoAttack, Obfuscated Gradients, and the use of generated data for robust training. A reproducible experimental study is conducted using the UCI Optical Recognition of Handwritten Digits dataset, containing 1,797 handwritten digit images. Four approaches are evaluated: Standard Training, PGD Adversarial Training, TRADES, and Gaussian Noise Training with Randomized Smoothing. The models are evaluated using clean accuracy, FGSM accuracy, PGD-20 accuracy, transfer attacks, certified accuracy, and training time across three random seeds. Experimental results demonstrate that the standard model achieves 97.6% clean accuracy but drops to 0.0% under a PGD attack with a perturbation budget of ε = 0.2. In comparison, PGD Adversarial Training and TRADES retain 51.9% and 51.2% accuracy, respectively, under the same attack budget. The study also demonstrates the importance of certified robustness, strong adversarial evaluation, and the trade-off between model accuracy, robustness, and computational cost. This work provides a transparent, low-compute, and reproducible comparative study of adversarial robustness techniques. The results are limited to a small dataset and a simple neural network and should not be interpreted as a large-scale performance benchmark. Author: Arya Anil JaradDepartment: Artificial Intelligence and Data Science EngineeringInstitution: JSPM's Narhe Technical Campus, Pune, Maharashtra, IndiaGuide: Prof. Kalyani Kulkarni Document status: Independent Preprint, Version 1.0, November 2026. Not peer reviewed.

Zenodo (CERN European Organization for Nuclear Research)
JSPM University (IN)
Adversarial Robustness in Machine Learning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon