Detailed Explanation: Targeted adversarial examples exploit vulnerabilities in machine learning models by introducing subtle perturbations to input data, such as images, text, or audio, to deceive the model into making a specific, incorrect prediction. Unlike untargeted adversarial examples, which only aim to cause a misclassification, targeted adversarial examples have a precise misclassification goal. For instance, modifying an image of a cat so that a neural network incorrectly classifies it as a dog. These examples highlight weaknesses in model robustness and security, serving as critical tools for testing and improving the resilience of AI systems against malicious attacks.
Historical Overview: The concept of adversarial examples was first discussed in the context of machine learning around 2014. The specific notion of targeted adversarial examples gained prominence shortly thereafter as researchers explored more sophisticated attack strategies and defenses.
Key Contributors: Notable figures in the development of adversarial example research include Ian Goodfellow, who, along with his colleagues, introduced the concept in their seminal 2014 paper. Other significant contributors include Alexey Kurakin, Nicholas Papernot, and Dawn Song, who have extensively studied and expanded upon adversarial attacks and defenses.
--
Regarding terminology, "targeted adversarial examples" is a precise and widely accepted term in the field. However, "targeted adversarial attacks" could also be used interchangeably and may emphasize the malicious intent behind generating such examples.