Learning when to stop Ileana Buhan @ileanabuhan Our story.. 2020 - PowerPoint PPT Presentation

Learning when to stop Ileana Buhan @ileanabuhan

Our story.. 2020 2017 2005 1997 1941 2

Its 1941 3

Its 1997… 4

Its 2005… 5

SCA (non-profiled) 𝑛 AES(k) closed sample 6

SCA (non-profiled) 𝑛 𝑛 AES(k) AES 𝑙 Leakage model Distinguisher 7

SCA (non-profiled) 𝑙 $ 𝑙 % 𝑙 ! 𝑙 " AES AES AES AES 𝑛 ! , 𝑛 " , 𝑛 $ . . 𝑛 # 𝑛 ! 𝑛 ! 𝑛 ! 𝑛 ! AES(k) AES AES AES AES 𝑛 " 𝑛 " 𝑛 " 𝑛 " AES AES AES AES 𝑛 # 𝑛 # 𝑛 # 𝑛 # 𝑙 & ≈ 𝑙 𝑒(𝑙 ! ) 𝑒(𝑙 " ) 𝑒(𝑙 # ) 𝑒(𝑙 $ ) 8

SCA (profiled) 𝑛 AES 𝑙 open sample 11

Its 2017 14

SCA (profiled) 𝑛 𝑛 AES AES(k) 𝑙 open sample closed sample ATTACK learning 15

Workflow Attack Acquisition Learning/Profiling TRACES BUILD the MACHINE USE the MACHINE TRAIN 16

Workflow Attack Acquisition Learning/Profiling TRACES BUILD the MACHINE USE the MACHINE TRAIN CONSTRUCT TRAIN 17

But.. 18

Workflow Attack Acquisition Learning/Profiling TRACES BUILD the MACHINE USE the MACHINE TRAIN CONSTRUCT TRAIN 19

But.. Perin G., Buhan I.R., Picek S., Learning when to stop:a mutual information approach to fight overfitting in profiled side-channel analysis a mutual information approach to fight overfitting in profiled side-channel analysis (Submitted to CHES 2020); 20

Leakage characterisation using deep networks labels 𝑌, 𝑍 data Input Layer Hidden Layers Output Layer Perin G., Buhan I.R., Picek S., Learning when to stop:a mutual information approach to fight overfitting in profiled side-channel analysis a mutual information approach to fight overfitting in profiled side-channel analysis (Submitted to CHES 2020);

Leakage characterisation using deep networks lets give a name to this layer 𝑌, 𝑍 𝑈 ! Input Layer Hidden Layers Output Layer

Leakage characterisation using deep networks 𝐽(𝑌; 𝑈 ! ), 𝐽(𝑈 ! ; 𝑍) the information this layer has about the data 𝑌, 𝑍 𝑈 ! Input Layer Hidden Layers Output Layer Perin G., Buhan I.R., Picek S., Learning when to stop:a mutual information approach to fight overfitting in profiled side-channel analysis a mutual information approach to fight overfitting in profiled side-channel analysis (Submitted to CHES 2020);

Leakage characterisation using deep networks 𝐽(𝑌; 𝑈 ! ), 𝐽(𝑈 ! ; 𝑍) the information layer T1 has about the labels 𝑌, 𝑍 𝑈 ! Input Layer Hidden Layers Output Layer Perin G., Buhan I.R., Picek S., Learning when to stop:a mutual information approach to fight overfitting in profiled side-channel analysis a mutual information approach to fight overfitting in profiled side-channel analysis (Submitted to CHES 2020);

Leakage characterisation using deep networks 𝐽(𝑌; 𝑈 " ), 𝐽(𝑈 " ; 𝑍) 𝑌, 𝑍 𝑈 𝑈 " ! Input Layer Hidden Layers Output Layer Perin G., Buhan I.R., Picek S., Learning when to stop:a mutual information approach to fight overfitting in profiled side-channel analysis a mutual information approach to fight overfitting in profiled side-channel analysis (Submitted to CHES 2020);

Leakage characterisation using deep networks 𝐽(𝑌; 𝑈 # ), 𝐽(𝑈 # ; 𝑍) 𝑌, 𝑍 𝑈 𝑈 𝑈 " ! # Input Layer Hidden Layers Output Layer Perin G., Buhan I.R., Picek S., Learning when to stop:a mutual information approach to fight overfitting in profiled side-channel analysis a mutual information approach to fight overfitting in profiled side-channel analysis (Submitted to CHES 2020);

Leakage characterisation using deep networks 𝐽(𝑌; 𝑇), 𝐽(𝑇; 𝑍) 𝑌, 𝑍 𝑈 𝑇 𝑈 𝑈 " ! # Input Layer Hidden Layers Output Layer Perin G., Buhan I.R., Picek S., Learning when to stop:a mutual information approach to fight overfitting in profiled side-channel analysis a mutual information approach to fight overfitting in profiled side-channel analysis (Submitted to CHES 2020);

How a network learns Perin G., Buhan I.R., Picek S., Learning when to stop:a mutual information approach to fight overfitting in profiled side-channel analysis a mutual information approach to fight overfitting in profiled side-channel analysis (Submitted to CHES 2020);

How the network learns Best epoch validation set

Leakage characterisation using deep networks GE AND Success Rate on ASCAD database Perin G., Buhan I.R., Picek S., Learning when to stop:a mutual information approach to fight overfitting in profiled side-channel analysis a mutual information approach to fight overfitting in profiled side-channel analysis (Submitted to CHES 2020);

Leakage characterisation using deep networks GE AND Success Rate on CHES AES database Perin G., Buhan I.R., Picek S., Learning when to stop:a mutual information approach to fight overfitting in profiled side-channel analysis a mutual information approach to fight overfitting in profiled side-channel analysis (Submitted to CHES 2020);

Message to my younger self never stop being the your best version never stop believing in yourself help Jane with security evaluations

Full-time teacher Guest Researcher (Senior) Research Scientist Radboud University Philips Research 2008 2019 Oct 2008 2011 2002 2003 Oct 2004 2010 Mathematics and consultant Computer Science Numerical Methods PhD, Product Manager Training Siemens University of Twente