Building a collision for 75-round reduced SHA-1 using GPU clusters

SHA-1 is one of the most widely used cryptographic hash functions. An important property of all cryptographic hash functions is collision resistance, that is, infeasibility of finding two different input messages such that they have the same hash values. Our work improves on differential attacks on SHA-1 and its reduced variants. In this work we describe porting collision search using method of characteristics to a GPU cluster. Method of characteristics employs backtracking search, which leads to low GPU performance due to branch divergence if implemented naively. Using a number of optimizations, we reduce branch divergence and achieve GPU usage efficiency of 50%, which gives 39 × acceleration over a single CPU core. With the help of our application running on a 512-GPU cluster, we were able to find a collision for a version of SHA-1 reduced to 75 rounds, which is currently (February 2012) the world's best result in terms of number of rounds for SHA-1.