Enhancing Algorithmic Resilience Against Data Poisoning Using CNN

Abstract

The work aims to improve model resilience and accuracy in machine learning (ML) by addressing data poisoning attacks. Data poisoning attacks are a type of adversarial attack where malicious data is injected into the training data set to manipulate the machine learning model's output, compromising model performance and security. To tackle this, a multi-faceted approach is proposed, including data assessment and cleaning, detecting attacks using outlier and anomaly detection techniques. The authors also train robust models using techniques such as adversarial training, regularization, and data diversification. Additionally, they use ensemble methods that combine the strengths of multiple models, as well as Gaussian processes and Bayesian optimization to improve resilience to attacks. The work aims to contribute to machine learning security by providing an integrated solution for addressing data poisoning attacks and advancing the understanding of adversarial attacks and defenses in the machine learning community.