Data and Model Poisoning refers to which practice?

• A. Manipulating Training/Validation Data To Introduce Vulnerabilities Or Biases
• B. Encrypting Training Data To Secure It
• C. Overfitting A Model To The Data
• D. Sharing Model Weights Publicly

Answer: (A)

Poisoning involves deliberately tampering with the data or the learning process to degrade performance or embed vulnerabilities or biases in a model. By manipulating training data, an attacker can cause the model to learn incorrect associations, or insert backdoors that trigger specific outputs, while altering validation data can skew evaluations to undermine trust or mask the attack. This directly matches the idea of manipulating training or validation data to introduce vulnerabilities or biases, which is why it’s the best fit. Encrypting training data is about protecting data, not compromising it, overfitting is a modeling pitfall rather than an attack, and sharing model weights deals with exposure risks rather than poisoning the learning process.