// phd candidate · lirmm × thales · final year

A little bit of DL-SCA

Deep-learning side-channel analysis: neural networks that pull secret keys out of a chip's power consumption — and the metrics that sometimes lie about how well they do it.

read --writingcat papers.bibstatus: open to security research / engineering roles from 2027
// live attacks, computed in your browser
demo://cpa · aes-128 · round 1 SubBytes · HW leakage● done
  1. 2bk00
  2. 7ek01
  3. 15k02
  4. 16k03
  5. 28k04
  6. aek05
  7. d2k06
  8. a6k07
  9. abk08
  10. f7k09
  11. 15k10
  12. 88k11
  13. 09k12
  14. cfk13
  15. 4fk14
  16. 3ck15

key recovered · 16/16 bytes · 173 simulated traces, correlated at build time. With JavaScript on, the attack runs live in your browser.

demo://template · aes-128 · 1st-order boolean masking○ idle

Every S-box output is split with a fresh random mask, so first-order CPA sees nothing. A template attack first learns the device on a clone, then recovers the key from the masked traces, racing a second-order CPA on the same data.

01 profile · clone device, key and masks known0 traces
02 attack · target device, key unknown, same masked traces for both

template · posterior per guess

  1. ··
  2. ··
  3. ··
  4. ··
  5. ··
  6. ··
  7. ··
  8. ··
  9. ··
  10. ··
  11. ··
  12. ··
  13. ··
  14. ··
  15. ··
  16. ··

cpa · 2nd order · |ρ| of the centred product

  1. ··
  2. ··
  3. ··
  4. ··
  5. ··
  6. ··
  7. ··
  8. ··
  9. ··
  10. ··
  11. ··
  12. ··
  13. ··
  14. ··
  15. ··
  16. ··
race · key bytes found vs traces (log)

ready

ls ./writing

ls -a →

animated · interactive

Is the gradient to blame for the plateau? TL;DR: no.

A learner that never computes a gradient still stalls on masked data, for just as long. The CHES 2026 poster, interactive.

cascade-2026 · animated · interactive

Is it really broken?

When the key isn't uniformly distributed, the metrics we use to call an attack successful can declare victory too early.

tches-2025 · animated · interactive

Scoop, explained: optimizing against higher-order masking

Why standard optimizers stall on masked implementations, and what changes when you account for it.

whoami

Arthemis team, LIRMM · Advanced Security Team, Thales. Supervised by Loïc Masure, Karine Heydemann, Vincent Migairou and Philippe Maurine.

Teaching: Foundations of Deep Learning, Random Signals. ./teaching →

0x01

Side-channel analysis as a learning problem

The plateau effect (Masure et al., 2023), its causes and how to overcome it; links between information theory and deep-learning side-channel attacks.

0x02

Optimization for deep learning

Second-order optimization and Bregman-based descent methods, and where they find appropriate use cases in deep learning.

cat papers.bib

talks & posters →