Similarity-based fraud detection techniques are provided in adaptive authentication systems. A method is provided for determining if an event is fraudulent by obtaining a plurality of tagged events and one or more untagged events, wherein the tagged events indicate a likelihood of whether the corresponding event was fraudulent; constructing a graph, wherein each node in the graph represents an event and has a value representing a likelihood of whether the corresponding event was fraudulent and wherein similar transactions are connected via weighted links; diffusing through weights in the graph to assign values to nodes such that neighbors of nodes having non-zero values receive similar values as the neighbors; and classifying whether at least one of the one or more untagged events is fraudulent based on the assigned values.