FGSM (Fast Gradient Sign Method) is an adversarial attack technique that adds small, calculated perturbations to input data to fool CNNs. Proposed by Ian Goodfellow in 2014, it generates adversarial examples to mislead the model's predictions.