Data Poisoning Attacks in Cognitive Computing

Data poisoning attacks pose a significant threat to the integrity of machine learning (ML) models, impacting their ability to generate accurate predictions. This research delves into the realm of adversarial ML 1, specifically focusing on poisoning attacks 4 during the training phase. A taxonomy of adversarial ML techniques is presented, classifying poisoning attacks into different categories. Real-world examples, including incidents involving Google's VirusTotal 8 scanning service, highlight the practical implications of data poisoning. The study explores defense mechanisms such as training data sanitization and robust optimization, acknowledging the nuanced balance required between model robustness and accuracy. The paper concludes with recommendations for safeguarding ML models against evolving adversarial threats, underscoring the ongoing importance of vigilance and education in the cybersecurity landscape.