Basics of poisoning attacks

One of the subtle yet potent threats to ML systems is poisoning attacks. Unlike traditional cyber attacks that target software vulnerabilities, poisoning attacks target the data – the lifeblood of ML systems. In this class of adversarial attacks, the attacker subtly manipulates the training data to compromise the learning process and maliciously influence the model outcomes at inference time. Adversaries perform poisoning attacks for a variety of reasons, including the following:

• Bias induction: Introducing biases in the model, making it perform unfairly or inaccurately for certain inputs
• Backdoor insertion: Inserting backdoors that can be triggered with specific inputs, allowing unauthorized access or behaviors
• Disruption: Degrading the model’s overall performance, undermining trust in its outputs
• Competitive sabotage: Harming the reputation or competitive advantage of the entity using the compromised model
• Ransom and...