Background EEG signals can be represented as the sum of a conventional AR process and an innovation process. It is know that conventional estimation techniques, such as least square estimates (LSE) or Gauasian maximum likelihood estimates (MLE-G) are optimal when the innovation process satisfies the Gaussian or presumed distribution. When the data are contaminated by outliers, however, these assumptions are not met and the power spectrum estimated by conventional estimation techniques may be fatally biased. EEG signal may be affected by artifacts, which are outliers in the statistical term. So the robust filtering estimation technique is used against those artifacts and it performs well for the contaminated EEG signal.