Hybrid domain adaptation method defends deep learning models against multi-targeted adversarial examples

A study published in Cluster Computing reports a hybrid adversarial domain adaptation method designed to defend deep learning models against multi-targeted adversarial examples. In evaluations on benchmark datasets, the approach, developed by Lovi Dhamija and Urvashi Bansal, showed accuracy improvements of 2 to 3 percent over state-of-the-art adversarial defenses.

Deep learning models have achieved rapid success across domains such as computer vision and autonomous vehicles, yet they remain highly vulnerable to adversarial attacks. These attacks introduce subtle perturbations that exploit weaknesses in a model and degrade its performance.

The most widely used countermeasure, adversarial training, improves resilience, but the authors note that it often overfits against the specific attacks a model was hardened against. According to the study, such robustness can lack generalization across diverse attacks and across domain shifts.

To address these issues, the researchers proposed a hybrid domain adaptation method intended to defend against both domain-specific and domain-invariant adversarial perturbations. The approach incorporates two components, logit correlation and gradient randomization, into adversarial domain adaptation, with the aim of aligning feature discrepancies between domains.

The team carried out extensive evaluations using benchmark datasets to demonstrate the performance of the model. The experimental assessments were designed to test how well the trained model generalizes against multiple adversarial threats rather than a single attack type.

The reported results cover generalization against four distinct attacks: the Fast Gradient Sign Method (FGSM), the Carlini and Wagner (C&W) attack, Projected Gradient Descent (PGD), and adversarial patches. Testing against this set of attacks forms the basis of the paper’s generalization claims.

Across these evaluations, the method demonstrated accuracy improvements of 2 to 3 percent compared with state-of-the-art adversarial defenses. The authors present these figures as evidence that combining logit correlation and gradient randomization with adversarial domain adaptation can improve robustness across varied adversarial domains.

The findings come from the benchmark dataset experiments described in the study, and the reported gains are limited to the attacks and settings evaluated. The authors state that the results are intended to encourage practitioners to build more robust and trustworthy systems against various adversarial domains.

Subject of Research: A hybrid adversarial domain adaptation defense against multi-targeted adversarial examples in deep learning

Article Title: A hybrid adversarial domain adaption method to defend against multi-targeted adversarial examples

Article References: Dhamija, L., & Bansal, U. (2026). A hybrid adversarial domain adaption method to defend against multi-targeted adversarial examples. Cluster Computing, 29(16), Article 898. https://doi.org/10.1007/s10586-026-06695-5

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06695-5

Keywords: adversarial examples, domain adaptation, deep learning, adversarial training, logit correlation, gradient randomization, FGSM, PGD, adversarial patches, model robustness, Cluster Computing, hybrid