Description
Deep neural networks (DNNs) undergo lengthy and expensive training procedures whose outcome - the DNN weights - represents a significant intellectual property asset to protect.
Side-channel analysis (SCA) has recently appeared as an effective approach to recover this confidential asset of DNN implementations.
Ding et al. (HOST’25) introduced MACPRUNING, a novel SCA countermeasure based on pruning, a performance-oriented Approximate Computing technique: at inference time, the implementation randomly prunes (or skips) non-important weights (i.e., with low contribution to the DNN’s accuracy) of the first layer, exponentially increasing the side-channel resilience of the protected DNN implementation.
In this presentation, we describe a preprocessing methodology taking advantage of a control-flow dependency intrinsic to the countermeasure's design.
Through practical experiments, we demonstrate the effectiveness of our methodology, recovering up to 96% of the important weights of a MACPruning-protected Multi-Layer Perceptron.
Moreover, we show how microarchitectural leakage improves the effectiveness of our methodology, even allowing for the recovery of up to 100% of the targeted non-important weights.
Lastly, by adapting our methodology, we elaborate on how the pruning mechanism, which depends on the importance of the weights, enables the circumvention of a control-flow-free MACPRUNING implementation.
With this last point, we identify the pruning mechanism underlying MACPRUNING as the root of the countermeasure’s vulnerability.
Autre
Présentation en Anglais (slides en Anglais)
Talk in English (slides in English)
Practical infos
Next sessions
-
HDL simulation for Masked Software Verification
Speaker : Quentin Meunier - Sorbonne Univ. Lip6
Masking is a countermeasure against Side-Channel Attacks (SCA) that aims to ensure that intermediate computations in an algorithm have secret-independent distributions through the use of random variables. This theoretically prevents SCAs, as power consumption is directly linked to the values manipulated by the program or hardware device. Designing a masking scheme is often non-trivial, and a[…]-
SemSecuElec
-
Side-channel
-
-
Étude, caractérisation et détection de verrouillage d'anneaux oscillants utilisés dans les générateurs de nombres aléatoires.
Speaker : Eloise Delolme - LabHC
Les générateurs de nombres aléatoires matériels basés sur des oscillateurs en anneau (RO-TRNGs) exploitent le jitter d’horloge comme source d’aléa afin de produire des séquences de bits aléatoires. Parmi ces architectures, le MURO-TRNG repose sur un modèle stochastique complexe qui suppose notamment l’indépendance des oscillateurs. Toutefois, dans la pratique, les oscillateurs en anneau sont[…]-
SemSecuElec
-
TRNG
-
-
Chamois: Formally verified compilation for optimisation and security
Speaker : David MONNIAUX - CNRS - Verimag
Embedded programs (including those on smart cards) are often developed in C and then compiled for the embedded processor. Sometimes they are modified by hand to incorporate countermeasures (fault attacks, etc.), but care must be taken to ensure that this does not disrupt normal program execution and that the countermeasure is actually adequate for blocking the attacks.In the process, it is[…]-
SemSecuElec
-
Fault injection
-
Formal methods
-
-
Securing processor's microarchitecture against SCA in a post-quantum cryptography setting
Speaker : Vincent MIGLIORE - LAAS-CNRS
Hardware microarchitecture is a well-known source of side-channel leakages, providing a notable security reduction of standard cryptographic algorithms (e.g. AES) if not properly addressed by software or hardware. In this talk, we present new design approaches to harden processor's microarchitecture against power-based side-channel attacks, relying on configurable and cascadable building blocks[…]-
SemSecuElec
-
Side-channel
-
Micro-architectural vulnerabilities
-