We propose in this paper an on-line algorithm based on Bloom filters to detect port scan attacks in IP traffic. Only relevant information about destination IP addresses and destination ports are stored in two steps in a two-dimensional Bloom filter. This algorithm can be indefinitely performed on a real traffic stream thanks to a new adaptive refreshing scheme that closely follows traffic variations. It is a scalable algorithm able to deal with IP traffic at a very high bit rate thanks to the use of hashing functions over a sliding window. Moreover it does not need any a priori knowledge about traffic characteristics. When tested against real IP traffic, the proposed on-line algorithm performs well in the sense that it detects all the port scan attacks within a very short response time of only 10 seconds without any false positive.