Breaking the Blindfold: Deep Learning-based Blind Side-channel Analysis
summary
Claims a practical framework, DL-BSCA, that uses deep neural nets plus a new Multi-point Cluster-based labeling to recover keys in blind side-channel settings where plaintext/ciphertext are unavailable. Demonstrates successful key recovery across four real datasets (AES, Ascon, Kyber) on AVR and noisy ARM STM32F4 hardware, including the first blind break of a desynchronization-protected trace set. Uses mixture-of-Gaussians clustering over about 50 points of interest and maximum-likelihood key ranking; experiments show neural models tolerate very noisy labels and actually recover keys where prior methods fail.
AI-written summary · about
chapters (10)
- 0:00Introduction to blind side-channel analysis
- 2:23Challenges in blind side-channel attacks
- 5:59Deep learning approach for blind analysis
- 7:08Theoretical joint distribution calculation
- 9:21Novel labeling with deep neural networks
- 10:46Multipoint clustering technique
- 14:29Training deep neural networks with noisy labels
- 16:38Empirical joint distribution prediction
- 17:42Comparing distributions with maximum likelihood
- 18:26Experimental results and attacks on datasets
related talks
19mReality Check on Side-Channels: Lessons learnt from breaking AES on ARM Cortex-A72...
41mFault Attacks & Side Channels
59mBreaking AES with side channel analysis
32mBlack box side-channel leakage verification using statistical approach
46mIncreasing the Analysis Surface of Large Language Models
38mMLSec Going Deeper
36mIt will slip by, or about Side Channels
30mData Driven Security