This paper proposes a novel non-intrusive system failure prediction technique using available information from developers and minimal information from raw logs (rather than mining entire logs) but keeping the data entirely private with the data owners. A neural network based multi-class classifier is developed for failure prediction, using artificially generated anonymous data set, applying a combination of techniques, viz., genetic algorithm (steps), pattern repetition, etc., to train and test the network. The proposed mechanism completely decouples the data set used for training process from the actual data which is kept private. Moreover, multi-criteria decision making (MCDM) schemes are used to prioritize failures meeting business requirements. Results show high accuracy in failure prediction under different parameter configurations. On a broader context, any classification problem, beyond failure prediction, can be performed using the proposed mechanism with artificially generated data set without looking into the actual data as long as the input features can be translated to binary values (e.g. output from private binary classifiers) and can provide classification-as-a-service.