AI Security Concerns In A Nutshell

← All articles
Artificial intelligence · EN

AI Security Concerns In A Nutshell

On 09.03.2023, the German Federal Office for Information Security published the attached guideline. It serves as an introductory resource for developers, shedding light on critical attacks targeting machine learning systems and potential defense strategies. Focusing on applications involving sensitive data, such as autonomous driving and biometric authentication, the guideline underscores the rising threat landscape in tandem with increased machine learning integration:

1. Evasion Attacks

Evasion attacks, seeking to mislead machine learning models during the inference phase. Adversarial examples, manipulated inputs causing misclassification, pose risks in both physical and digital domains. Countermeasures include limiting query numbers and supplying only relevant information during model queries.

2. Information Extraction Attacks

Privacy-centric information extraction attacks, encompassing model stealing, attribute inference, membership inference, and model inversion, aim to reconstruct models or data from training sets. These attacks often require prior knowledge or access to publicly available parts of the training dataset.

3. Poisoning and Backdoor Attacks

Poisoning attacks aim to degrade machine learning model performance by manipulating the training dataset. Flipping labels of inputs to a desired class and injecting them into the training set can be computationally inexpensive but may be detectable through manual inspection.

#AISecurity

This post has a document attached on LinkedIn. View on LinkedIn →

Discuss on LinkedIn