Spark ignition engines are often desired to be operated close to its knock borderline limit, when MBT (maximum brake torque) cannot be achieved, to optimize engine efficiency. Traditionally, engine knock is closed-loop controlled using a dual-rate PID scheme with the help of feedforward control based on off-line calibrated borderline knock limit along with associated corrections. Note that the feedforward control is often not accurate and very conservative. This paper proposed a data-driven knock baseline control architecture consisting of two-key components: (a) offline training knock borderline Surrogate models under the worst and lightest knock conditions and (b) online updating the composite Surrogate model for adaptation. A Bayesian-based multi-objectives optimization algorithm is used to obtain optimal knock control parameters with significantly improved calibration efficiency, where the offline trained models are obtained using spark and intake valve timings as control parameters, and indicated specific fuel consumption (ISFC) and knock intensity (KI) as two competing performance measures to be minimized. Based on our early results, two Kriging surrogate models with two corresponding Pareto fronts (most and least advanced timing) can be obtained, and the purpose of this paper is to generate a composite real-time Kriging model for compensating engine aging and operational environmental changes such as fuel type, temperature, humidity, etc. In order to reduce the number of control parameters used in online updating, principal component analysis is conducted to find dominated control parameter, and in this case, the spark timing is the most sensitive factor of Pareto front. During the online updating process, a likelihood ratio controller with short- and long-term buffers was proposed to update the surrogate model in real-time and to adapt to fast and slow environmental and engine variations so that the optimal borderline knock limit can be found based on the intersection of mean and variance of Pareto front from the real-time updated composite Kriging model; and accordingly, the optimal control parameters can be located in the design space using surrogate model. Both simulation and test results indicate that the proposed online updating scheme is able to update the machine-learned stochastic surrogate models and adjust the feedforward borderline knock control parameters adapted to engine aging and operational environment.