A system and method for detecting cyber-attacks using quantile regression analysis are disclosed. The method includes: identifying at least one hit quantile out of a plurality of quantiles, wherein the at least one identified hit quantile falls within quantile edges of a sample of traffic directed at a protected entity, wherein each of the plurality of quantiles is characterized by a probability distribution of at least one feature of a data stream, each of the plurality of quantiles having a respective probability estimate; updating the probability estimates of the plurality of quantiles when the at least one hit quantile has been identified; and when the probability estimate of the at least one hit quantile is above a threshold, taking an action to mitigate existence of a cyber-attack.