Detection of Poisoned Data Using Genetic Algorithm : Case Study on Badnet Attack in Image Classification Models

Abstract

The increasing reliance on deep learning models, particularly in image classification tasks, has raised critical concerns regarding their security and robustness. Among the most seri ous threats are data poisoning attacks, especially backdoor attacks such as BadNets, where malicious patterns are embedded into training data to manipulate model behavior. These at tacks are difficult to detect due to their stealthy nature and minimal impact on overall model performance. This thesis proposes a detection framework based on Genetic Algorithms (GAs) to iden tify poisoned data within image datasets. The approach uses evolutionary optimization tech niques to explore the data space and distinguish between clean and compromised samples, supported by a customized fitness function to improve detection accuracy. Experimental results demonstrate that the proposed method achieves promising perfor mance in detecting poisoned data and isolating malicious samples, highlighting the effective ness of evolutionary algorithms in enhancing the security of deep learning systems.