Paper 2019/037
Improving Attacks on Speck32/64 using Deep Learning
Aron Gohr
Abstract
This paper presents a very practical key recovery attack on Speck32/64 reduced to 11 rounds based on a novel type of differential distinguisher using machine learning. These distinguishers exceed distinguishers based on the entire differential distribution table of Speck32/64 in accuracy, specificity and sensitivity. We show that they obtain significant gain from features of the output distribution that are invisible to the differential distribution table. The key recovery attack has been completely verified empirically and has an average runtime of approximately three minutes on a desktop computer with a fast graphics card or about 30 minutes on the same machine when not using the graphics card. This corresponds to roughly 41 bits of remaining security for 11-round Speck32/64, which is a substantial improvement over previous literature. The average data complexity of our attack is slightly lower than the best previous attack on the same number of rounds. While our attack is based on a known input difference taken from the literature, we also show that neural networks can be used to rapidly (within a matter of minutes on our machine) find good input differences without using prior human cryptanalysis.
Metadata
- Available format(s)
- Category
- Secret-key cryptography
- Publication info
- Preprint. MINOR revision.
- Keywords
- Speck and Deep Learning and Cryptanalysis
- Contact author(s)
- aron gohr @ gmail com
- History
- 2019-08-15: last of 2 revisions
- 2019-01-17: received
- See all versions
- Short URL
- https://ia.cr/2019/037
- License
-
CC BY