Sometimes even the Machine learning algorithms behave so dumb that an image recognition model can be confused by generating an adversarial instance, i.e. by changing few pixels by either taking derivative of model output or exploiting genetic algorithms. Adversarial instances lie in low probability regions which is in contrast with limited instances of high probability regions from which the model was trained. A possible approach to solve this problem is ensemble training - To let multiple models back each other. As we look forward to developing more artificial intelligent systems it would become common to encounter such problems.

You can read more at: http://www.erogol.com/ensembling-against-adversarial-instances/