Non-Profiled Deep Learning-based Side-Channel attacks with - PDF document

Non-Profiled Deep Learning-based Side-Channel attacks with Sensitivity Analysis Benjamin Timon August 28, 2019 Python notebook presentation for CHES 2019 1

Introduction & Motivation Profiled vs Non-Profiled attacks Machine Learning trend 2

This research 3

Differential Deep Learning Analysis (DDLA) Correlation Power Analysis Follow similar strategy for DDLA 4

Demonstration: attack with accuracy and loss Generate simulation traces ■♥ ❬✶❪✿ ❢r♦♠ ❞❡♠♦ ✐♠♣♦rt ❣❡♥❴s✐♠✉❴❞❛t❛ ★ ●❡♥❡r❛t❡ s✐♠✉❧❛t✐♦♥ tr❛❝❡s ❢♦r ❞❡♠♦♥str❛t✐♦♥ ❞❛t❛❴tr❛✐♥✐♥❣ ❂ ❣❡♥❴s✐♠✉❴❞❛t❛✭✮ Network 5

Observe loss and accuracy during training ■♥ ❬✷❪✿ ❢r♦♠ ❞❡♠♦ ✐♠♣♦rt ❞❡♠♦❴❞❞❧❛❴❛❝❝❴❧♦ss ❞❞❧❛ ❂ ❞❡♠♦❴❞❞❧❛❴❛❝❝❴❧♦ss✭❞❛t❛❴tr❛✐♥✐♥❣✮ ❞❞❧❛✳r✉♥✭♥❴❡♣♦❝❤s❂✸✵✮ ❞❞❧❛✳❢✐❣ ❖✉t❬✷❪✿ Sensitivity Analysis 6

Demonstration: Observe first layer gradient ■♥ ❬✸❪✿ ❢r♦♠ ♣❧♦ts ✐♠♣♦rt ♣❧♦t❴✇❡✐❣❤ts❴❛♥❞❴❣r❛❞❴✸❞ ♣❧♦t❴✇❡✐❣❤ts❴❛♥❞❴❣r❛❞❴✸❞✭❞❛t❛❴tr❛✐♥✐♥❣✮ ❖✉t❬✸❪✿ Observe first layer gradient during training ■♥ ❬✹❪✿ ❢r♦♠ ❞❡♠♦ ✐♠♣♦rt ❞❡♠♦❴❞❞❧❛❴❣r❛❞✐❡♥t ❞❞❧❛ ❂ ❞❡♠♦❴❞❞❧❛❴❣r❛❞✐❡♥t✭❞❛t❛❴tr❛✐♥✐♥❣✮ ❞❞❧❛✳r✉♥✭✶✺✮ ❞❞❧❛✳❢✐❣ ❖✉t❬✹❪✿ 7

■♥ ❬✺❪✿ ❢r♦♠ ♣❧♦ts ✐♠♣♦rt ♣❧♦t❴✇❡✐❣❤ts❴✷❞ ♣❧♦t❴✇❡✐❣❤ts❴✷❞✭❞❞❧❛✮ 8

❖✉t❬✺❪✿ Derivatives with regards to the inputs 9

Masked implementations Generate masked simulation traces ■♥ ❬✻❪✿ ❢r♦♠ ❞❡♠♦ ✐♠♣♦rt ❣❡♥❴s✐♠✉❴♠❛s❦❡❞❴❞❛t❛ ★ ●❡♥❡r❛t❡ ♠❛s❦❡❞ s✐♠✉❧❛t✐♦♥ tr❛❝❡s ❢♦r ❞❡♠♦♥str❛t✐♦♥ ❞❛t❛❴tr❛✐♥✐♥❣ ❂ ❣❡♥❴s✐♠✉❴♠❛s❦❡❞❴❞❛t❛✭✮ Demonstration: high-order DDLA ■♥ ❬✼❪✿ ❢r♦♠ ❞❡♠♦ ✐♠♣♦rt ❞❡♠♦❴❞❞❧❛❴❤✐❣❤❴♦r❞❡r ❞❞❧❛ ❂ ❞❡♠♦❴❞❞❧❛❴❤✐❣❤❴♦r❞❡r✭❞❛t❛❴tr❛✐♥✐♥❣✮ ❞❞❧❛✳r✉♥✭✷✵✮ ❞❞❧❛✳❢✐❣ ❖✉t❬✼❪✿ 10

■♥ ❬✽❪✿ ❢r♦♠ ♣❧♦ts ✐♠♣♦rt ♣❧♦t❴✇❡✐❣❤ts❴✷❞ ♣❧♦t❴✇❡✐❣❤ts❴✷❞✭❞❞❧❛✮ ❖✉t❬✽❪✿ 11

12

13

14