Cyber threats of machine learning

This paper offers a detailed and comprehensive analysis of the various cyber threats that target machine learning models and applications. It begins by characterizing basic classifiers and exploring the objectives of intentional attacks on typical classifiers, providing a foundational understanding of the threat landscape. The paper then thoroughly examines the vulnerabilities that machine learning systems face, alongside the methods for detecting, countering, and responding to these cyber threats. Special attention is given to specific types of threats, including attacks on machine learning models, adversarial attacks, poisoning attacks, and backdoor attacks. The paper also addresses critical issues such as attacks on data protection mechanisms, replay attacks, denial of service attacks, learning model theft, malware, and breaches in data privacy. Each of these threats is analyzed in detail, with a focus on their potential impact and the strategies that can be employed to mitigate them. In its conclusion, the paper provides recommendations on regulatory measures and best practices to safeguard machine learning models and applications against these evolving cyber threats. These recommendations emphasize the necessity for a robust regulatory framework to ensure the security, reliability, and integrity of machine learning systems in an increasingly digital and interconnected world.

---

Opracowanie rekordu ze środków MNiSW, umowa nr POPUL/SP/0154/2024/02 w ramach programu "Społeczna odpowiedzialność nauki II" - moduł: Popularyzacja nauki (2025).