Title: C-LEAD: Leveraging Contrastive Learning for Robust Adversarial Defense

Abstract

While deep neural networks (DNNs) have demonstrated exceptional performance in computer vision applications—including object detection, image segmentation, and image classification—they remain highly susceptible to adversarial attacks. These attacks can trick models into making erroneous predictions through minor, often imperceptible, alterations to input images. Ensuring the deployment of secure and resilient deep-learning systems requires addressing this critical vulnerability.

This study introduces a novel defense mechanism that applies contrastive learning to the problem of adversarial robustness, an area that has seen limited prior exploration. The proposed method, C-LEAD, employs a contrastive loss function to strengthen classification models. It achieves this by training the network on a combination of clean images and those subjected to adversarial perturbations. By simultaneously optimizing the model’s parameters and the perturbations, the approach facilitates the learning of robust feature representations that are significantly less prone to adversarial manipulation.

Experimental evaluations demonstrate that this technique yields substantial gains in model resilience across a variety of adversarial attack types. These findings indicate that contrastive loss enables the extraction of more informative and durable features, thereby advancing the state of the art in adversarial robustness for deep learning. The source code for this work is publicly accessible at: https://github.com/suklav/C_Lead .

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC