We consider the problem of estimating the mean and covariance of a
distribution from iid samples in $\mathbb{R}^n$, in the presence of an $\eta$
fraction of malicious noise; this is in contrast to much recent work where the
noise itself is assumed to be from a distribution of known typ