We are working on anomaly detection technologies to detect things like DDOS attacks; we're about to publish some results and will let you know when that is ready. In the mean time, you might want to look at attacks against ML itself, see [0], [1], and [2]. The bottom line here is that even simple linear models are susceptible to adversarial attacks, so things like the autoencoders we used to do binary classification (e.g., anomaly detection) are susceptible; see slides 16+ of [3].
[0] https://arxiv.org/pdf/1312.6199.pdf
[1] https://arxiv.org/pdf/1412.6572.pdf
[2] https://arxiv.org/pdf/1602.02697.pdf
[3] http://www.1-4-5.net/~dmm/ml/talks/2016/cor_ml4networking.pdf