Markov chains aren't really relevant here - they're a way to generate random but "seemingly plausible" data - but yes, there is a way to fool the type of convolutional neural networks used in image recognition. Paper here:

http://www.evolvingai.org/fooling