THE INFLUENCE-BASED DEFENSE AGAINST TO THE DATA POISONING ATTACKS IN ONLINE LEARNING

Data poisoning attacks pose a significant threat to online learning systems by introducing malicious data into the training process, thereby compromising model integrity and performance. Traditional defense mechanisms often fall short in identifying and mitigating the subtle influences of poisoned data, especially in dynamic, real-time learning environments. This paper introduces an influence-based defense strategy designed to detect and neutralize data poisoning attacks in online learning systems. By leveraging influence functions from robust statistics, our approach assesses the impact of individual data points on the model's predictions, enabling the identification of anomalous instances indicative of poisoning. We integrate this influence analysis with data sanitization techniques to filter out potentially harmful data before it adversely affects the learning process. Our method operates in real-time, making it suitable for online learning scenarios where data arrives sequentially. Empirical evaluations across various datasets and attack strategies demonstrate the effectiveness of our approach in maintaining model accuracy and robustness against data poisoning attempts. The proposed defense mechanism offers a proactive solution to safeguard online learning models from adversarial data manipulations, ensuring their reliability and trustworthiness in practical applications.