國立虎尾科技大學 |

Advanced Deep Learning-Assisted Side-Channel Attack Framework and Transfer Learning.

Record Type:	Language materials, manuscript : Monograph/item
Title/Author:	Advanced Deep Learning-Assisted Side-Channel Attack Framework and Transfer Learning./
Author:	Zhang, Ziyue.
Description:	1 online resource (132 pages)
Notes:	Source: Dissertations Abstracts International, Volume: 85-06, Section: B.
Contained By:	Dissertations Abstracts International85-06B.
Subject:	Mathematics. -
Online resource:	click for full text (PQDT)
ISBN:	9798381108569

Advanced Deep Learning-Assisted Side-Channel Attack Framework and Transfer Learning.
Zhang, Ziyue.

Advanced Deep Learning-Assisted Side-Channel Attack Framework and Transfer Learning. - 1 online resource (132 pages)

Source: Dissertations Abstracts International, Volume: 85-06, Section: B.

Thesis (Ph.D.)--Northeastern University, 2023.

Includes bibliographical references

Recently deep-learning (DL) techniques have been widely adopted in side-channel analysis and generally considered the state-of-art method. Specifically, when appropriately trained, DL-assisted Side-Channel Analysis (SCA) demonstrates unparalleled performance while breaking existing countermeasures and tolerating common issues like high-dimensional input and imbalanced labels, etc.In the meantime, the DL-assisted SCA (DL-SCA) also present new challenges. This thesis addresses three of such challenges. First, there needs a good quantitative evaluation for DL-SCA effectiveness. The existing metrics that quantify the effectiveness of an SCA, such as Guessing Entropy(GE) and Success Rate(SR), generally suffer from high computational complexity particularly for the full key case. For DL-SCA there is also the additional issue that no explicit leakage model exists which are required by many standard GE/SR evaluation techniques. Second, DL-assisted SCA generally consists of two phases: a deep neural network (DNN) training phase and a follow-on attack phase using the trained DNN, but the two phases may not align well. The DNN training phase solves a classification problem. But the traditional training metrics for DNN classifications were found to result in suboptimal or misleading follow-on attack in many scenarios. Thirdly, the DNN is training on a profiling device while the follow-on attack is applied on another target device in practice. While the DL-SCA is often evaluated with one device used as both the profiling device and the target device, those two devices will differ in realistic attack scenarios. To overcome the difference between the profile device and the target device, transfer learning is needed for DLSCA. Most previous transfer learning based DL-SCAs use supervised transfer learning that requires labeled leakage traces from the target device. However, in realistic attack scenarios, the adversary does not have much control of the target device and can only obtain leakage measurements in a passive way, with the key unknown, the leakage traces are unlabeled, and therefore unsupervised transfer learning is desired. The only previous unsupervised transfer learning DL-SCA, CDPA, assumes a domain-invariant DL model, and thus can only succeed when transferring between very similar devices.In this thesis, we propose solutions to the above three challenges. In the first chapter, we propose a new guessing entropy estimation algorithm (GEEA) utilizing a theoretical relationship between GE and pairwise success rates. This leads to an accurate and fast estimation of GE. In the second chapter, we propose a universal training framework for DL-SCA based on GEEA. This resolves the mismatch between the training phase and the follow-on attack phase of DL-SCA optimally in theory. In the third chapter, we propose a novel unsupervised transfer learning DL-SCA methodology to allow successful attacks across distinctively different devices and also different types of side channels. Our method only assumes a common domain-invariant classifier. We demonstrate for the first time practical successful transfer DL-SCAs between different side-channels, power vs. EM and cache side-channel vs. power/EM side channel, and also across different implementation platforms.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2024

Mode of access: World Wide Web

ISBN: 9798381108569Subjects--Topical Terms:

527692
Mathematics.
Subjects--Index Terms:

Deep learningIndex Terms--Genre/Form:

554714
Electronic books.

Advanced Deep Learning-Assisted Side-Channel Attack Framework and Transfer Learning.
LDR:04650ntm a22003737 4500 001 1148771
005 20240930100143.5
006 m o d
007 cr bn ---uuuuu
008 250605s2023 xx obm 000 0 eng d
020 $a 9798381108569
035 $a (MiAaPQ)AAI30815614
035 $a AAI30815614
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Zhang, Ziyue. $3 1474825
245 1 0 $a Advanced Deep Learning-Assisted Side-Channel Attack Framework and Transfer Learning.
264 0 $c 2023
300 $a 1 online resource (132 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 85-06, Section: B.
500 $a Advisor: Ding, Aidong Adam;Fei, Yunsi.
502 $a Thesis (Ph.D.)--Northeastern University, 2023.
504 $a Includes bibliographical references
520 $a Recently deep-learning (DL) techniques have been widely adopted in side-channel analysis and generally considered the state-of-art method. Specifically, when appropriately trained, DL-assisted Side-Channel Analysis (SCA) demonstrates unparalleled performance while breaking existing countermeasures and tolerating common issues like high-dimensional input and imbalanced labels, etc.In the meantime, the DL-assisted SCA (DL-SCA) also present new challenges. This thesis addresses three of such challenges. First, there needs a good quantitative evaluation for DL-SCA effectiveness. The existing metrics that quantify the effectiveness of an SCA, such as Guessing Entropy(GE) and Success Rate(SR), generally suffer from high computational complexity particularly for the full key case. For DL-SCA there is also the additional issue that no explicit leakage model exists which are required by many standard GE/SR evaluation techniques. Second, DL-assisted SCA generally consists of two phases: a deep neural network (DNN) training phase and a follow-on attack phase using the trained DNN, but the two phases may not align well. The DNN training phase solves a classification problem. But the traditional training metrics for DNN classifications were found to result in suboptimal or misleading follow-on attack in many scenarios. Thirdly, the DNN is training on a profiling device while the follow-on attack is applied on another target device in practice. While the DL-SCA is often evaluated with one device used as both the profiling device and the target device, those two devices will differ in realistic attack scenarios. To overcome the difference between the profile device and the target device, transfer learning is needed for DLSCA. Most previous transfer learning based DL-SCAs use supervised transfer learning that requires labeled leakage traces from the target device. However, in realistic attack scenarios, the adversary does not have much control of the target device and can only obtain leakage measurements in a passive way, with the key unknown, the leakage traces are unlabeled, and therefore unsupervised transfer learning is desired. The only previous unsupervised transfer learning DL-SCA, CDPA, assumes a domain-invariant DL model, and thus can only succeed when transferring between very similar devices.In this thesis, we propose solutions to the above three challenges. In the first chapter, we propose a new guessing entropy estimation algorithm (GEEA) utilizing a theoretical relationship between GE and pairwise success rates. This leads to an accurate and fast estimation of GE. In the second chapter, we propose a universal training framework for DL-SCA based on GEEA. This resolves the mismatch between the training phase and the follow-on attack phase of DL-SCA optimally in theory. In the third chapter, we propose a novel unsupervised transfer learning DL-SCA methodology to allow successful attacks across distinctively different devices and also different types of side channels. Our method only assumes a common domain-invariant classifier. We demonstrate for the first time practical successful transfer DL-SCAs between different side-channels, power vs. EM and cache side-channel vs. power/EM side channel, and also across different implementation platforms.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2024
538 $a Mode of access: World Wide Web
650 4 $a Mathematics. $3 527692
653 $a Deep learning
653 $a Side-channel attacks
653 $a Side-Channel Analysis
653 $a Deep neural network
655 7 $a Electronic books. $2 local $3 554714
690 $a 0405
690 $a 0800
710 2 $a ProQuest Information and Learning Co. $3 1178819
710 2 $a Northeastern University. $b Mathematics. $3 1474826
773 0 $t Dissertations Abstracts International $g 85-06B.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=30815614 $z click for full text (PQDT)