Discovering associations between elements occurring in a stream is applicable in numerous applications, including predictive caching and fraud detection. These applications require a new model of association between pairs of elements in streams. We develop an algorithm, Streaming-Rules, to report association rules with tight guarantees on errors, using limited processing per element, and minimal space. The modular design of Streaming-Rules allows for integration with current stream management systems, since it employs existing techniques for finding frequent elements. The presentation emphasizes the applicability of the algorithm to fraud detection in advertising networks. Such fraud instances have not been successfully detected by current techniques. Our experiments on synthetic data demonstrate scalability and efficiency. On real data, potential fraud was discovered.