Defending Machine Learning and Deep Learning Models: Detecting and Preventing Data Poisoning Attacks

Data poisoning attacks present a significant threat to machine learning (ML) and deep learning models, compromising their reliability and effectiveness during inference. This paper draws insights from recent studies to explore various aspects, including the vulnerabilities of ML architectures, especially deep learning models, to these attacks. It introduces novel methodologies to detect and mitigate data poisoning, considering factors such as model robustness, performance degradation, and poisoning ratio thresholds. Special attention is given to backdoor attacks in multimodal learning scenarios, proposing strategies like Clean CLIP to counter spurious associations. By synthesizing findings across different attack types and defenses, the paper underscores the importance of standardized benchmarks and realistic evaluation settings in assessing countermeasures’ efficacy. Ultimately, this work enhances our understanding of data poisoning vulnerabilities in deep learning, offering practical strategies to bolster the resilience of ML systems against evolving adversarial threats.