Heuristic Approaches
summary
A deep‑learning ensemble called NNBits trains networks to predict missing bits in long avalanche sequences and flags bits whose prediction exceeds chance, detecting non‑randomness beyond NIST STS — e.g., spotting the SPECK32 input difference up to SPECK32 7 rounds (bits 716 and 732) and publishing code and notebooks. A separate line of work reframes SLP to SLPA/SLPD, adds division optimization and extending base heuristics, prunes AES MixColumns from 103 to 102 XORs in hours and improves ~3,300 of ~4,000 test matrices down to as low as 85 XORs, reducing area, power and latency.
AI-written summary · about
chapters (21)
- 0:00Introduction to talk
- 1:06Cryptanalysis and randomness testing
- 6:27Explainability in AI for cryptanalysis
- 9:03NNBits system explained
- 11:55Dataset and experiments
- 13:21S2 experiments and limitations
- 14:05Bit prediction results on SPECK
- 17:01Using NNBits and source code
- 18:30Conclusion and future work
- 19:12Q&A on key recovery
- 20:45Q&A on random number tests
- 21:58Lightweight cryptography background
- 24:09Linear layer optimization focus
- 26:47SLP problem and extensions
- 28:34Forward and backward search methods
- 31:15Division optimization technique
- 32:19Backward search properties
- 35:22General optimization framework
- 37:42Application to AES MixColumns
- 38:28Results on 4000 matrices
- 39:56Conclusion and future steps
related talks
40mImplementation
20mBreaking the Blindfold: Deep Learning-based Blind Side-channel Analysis
40mCryptanalysis I
19mReality Check on Side-Channels: Lessons learnt from breaking AES on ARM Cortex-A72...
23mSmall-Box Cryptography
44mLPN Based IO Part 1
34mHAWKEYE – Recovering Symmetric Cryptography From Hardware Circuits
32mBlack box side-channel leakage verification using statistical approach