Gaussian estimation under attack uncertainty

We consider the estimation of a standard Gaussian random variable under an observation attack where an adversary may add a zero mean Gaussian noise with variance in a bounded, closed interval to an otherwise noiseless observation. A straightforward approach would entail either ignoring the attack and simply using an optimal estimator under normal operation or taking the worst-case attack into account and using a minimax estimator that minimizes the cost under the worst-case attack. In contrast, we seek to characterize the optimal tradeoff between the MSE under normal operation and the MSE under the worst-case attack. Equivalently, we seek a minimax estimator for any fixed prior probability of attack. Our main result shows that a unique minimax estimator exists for every fixed probability of attack and is given by the Bayesian estimator for a least-favorable prior on the set of possible variances. Furthermore, the least-favorable prior is unique and has a finite support. While the minimax estimator is linear when the probability of attack is 0 or 1, our numerical results show that the minimax linear estimator is far from optimal for all other probabilities of attack and a simple nonlinear estimator does much better.