A Novel Two-Stage Training Detection Mechanism Against Adversarial Attacks on Images

Abstract Deep learning (DL) models are prone to a variety of adversarial attacks, which results in monetary losses and reputational damagefor their applications. Detecting adversarial attacks is crucial to take corrective actions for reducing their impact on applications thatare based on the DL approach. Many researchers have developed different detection methods that rely on analytically determinedparameters, often resulting in inadequate outcomes. The present detection algorithm incurs considerable computational overheaddue to complex feature extraction and high-dimensional data processing, which consequently affects the classification performance.To upgrade the performance, a more effective detection mechanism is needed. In this work, a fast and efficient detection methodis approached. Since CNN-based models are widely used for classification of images, the proposed detection technique is testedon common CNN architectures, such as ResNet50, VGG16, InceptionV4, and MobileNetV2. The publicly available CIFAR-100dataset is subjected to various adversarial attacks, such as Basic Iterative Method (BIM), Projected Gradient Descent (PGD), Carliniand Wagner (C&W), SparseFool, and AutoAttack. The performance analysis of different models shows how good the detectionmethod is as the adversarial training-based detection achieves accuracies of 94.16% for ResNet50, 93.81% for VGG16, 91.26% forInceptionV4, and 94.06% for MobileNetV2. The findings show that, in comparison to current techniques, the suggested detectionmethod significantly improves adversarial detection performance.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-16
DOI
https://doi.org/10.21203/rs.3.rs-8338233/v1
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 Novel Two-Stage Training Detection Mechanism Against Adversarial Attacks on Images

Pankaj Kumar Keserwani, Mahesh Chandra Govil, Suman Das
2 links
Research Square
Adversarial Robustness in Machine Learning
preprint

A Novel Two-Stage Training Detection Mechanism Against Adversarial Attacks on Images

Pankaj Kumar Keserwani, Mahesh Chandra Govil, Suman Das
preprint en

Abstract

Abstract Deep learning (DL) models are prone to a variety of adversarial attacks, which results in monetary losses and reputational damagefor their applications. Detecting adversarial attacks is crucial to take corrective actions for reducing their impact on applications thatare based on the DL approach. Many researchers have developed different detection methods that rely on analytically determinedparameters, often resulting in inadequate outcomes. The present detection algorithm incurs considerable computational overheaddue to complex feature extraction and high-dimensional data processing, which consequently affects the classification performance.To upgrade the performance, a more effective detection mechanism is needed. In this work, a fast and efficient detection methodis approached. Since CNN-based models are widely used for classification of images, the proposed detection technique is testedon common CNN architectures, such as ResNet50, VGG16, InceptionV4, and MobileNetV2. The publicly available CIFAR-100dataset is subjected to various adversarial attacks, such as Basic Iterative Method (BIM), Projected Gradient Descent (PGD), Carliniand Wagner (C&W), SparseFool, and AutoAttack. The performance analysis of different models shows how good the detectionmethod is as the adversarial training-based detection achieves accuracies of 94.16% for ResNet50, 93.81% for VGG16, 91.26% forInceptionV4, and 94.06% for MobileNetV2. The findings show that, in comparison to current techniques, the suggested detectionmethod significantly improves adversarial detection performance.

Research Square
National Institute of Technology Sikkim (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