國立虎尾科技大學 |

Towards Scaling Artificial Intelligence for Resilience and Robustness.

Record Type:	Language materials, manuscript : Monograph/item
Title/Author:	Towards Scaling Artificial Intelligence for Resilience and Robustness./
Author:	Sarmadi, Alireza.
Description:	1 online resource (230 pages)
Notes:	Source: Dissertations Abstracts International, Volume: 85-11, Section: B.
Contained By:	Dissertations Abstracts International85-11B.
Subject:	Computer science. -
Online resource:	click for full text (PQDT)
ISBN:	9798382762722

Towards Scaling Artificial Intelligence for Resilience and Robustness.
Sarmadi, Alireza.

Towards Scaling Artificial Intelligence for Resilience and Robustness. - 1 online resource (230 pages)

Source: Dissertations Abstracts International, Volume: 85-11, Section: B.

Thesis (Ph.D.)--New York University Tandon School of Engineering, 2024.

Includes bibliographical references

Integration of Artificial Intelligence (AI) into areas such as image classification, speech recognition, natural language processing, and autonomous systems illustrates its potential to revolutionize everyday life. However, this promise is tempered by significant vulnerabilities in AI technologies, including susceptibility to unpredictable behaviors and manipulations by adversaries. This dissertation focuses on the robustness and resilience of AI applications in domains such as high-dimensional controller tuning and biometrics identification. Moreover, this dissertation investigates robustness/security aspects of AI algorithms themselves, such as privacy-preserving cooperative learning between organizations, robustness against transistor aging effects on Deep Neural Networks (DNNs), mitigating backdoor attacks on DNNs, and explainability by mining human-readable decision rules from data.Data-driven control policies are utilized when an agent seeks to find a controller without any prior knowledge about the system dynamics by training an end-to-end controller (i.e., the input to the controller is the observed state of the system; the output is the actuation vector). Learning the controller typically requires large numbers of samples (e.g., tens of thousands) to achieve reasonable performance. However, collecting large amounts of data on real-world systems is challenging due to physical wear and tear, safety issues, and the time required. On the other hand, optimization-based control formulates the problem into an optimization problem. The performance of optimization-based methods is very sensitive to cost function constraints and system dynamics. These hyperparameters should be tuned carefully to achieve the desired behavior. A method is proposed to automatically and efficiently tune high-dimensional vectors of controller parameters.In the context of user identity verification, traditional methods like Personal Identification Numbers and passwords have proven vulnerable to security breaches, information leaks, and counterfeiting. Alternatives that rely on biometric features, including facial characteristics, fingerprints, and iris patterns, face risks from technological advancements that enable their duplication through methods like 3D printing and artificial iris contact lenses. In contrast, Surface Electromyography (sEMG) offers a more secure solution by leveraging the unique and complex neurophysiological responses of individuals, making impersonation and fraud significantly more challenging. An identification method is proposed that uses a CNN-based framework to identify subjects utilizing their sEMG data.AI algorithms suffer from weaknesses that make them vulnerable. Many organizations in different domains (e.g., enterprise networks and healthcare) are sharing data to increase the performance of their ML algorithms. However, sharing data puts the privacy of sample owners in danger. Privacy could be guaranteed by using encryption methodologies; however, these methods are computationally expensive for huge amounts of data. A new solution that seeks to address the computational bottleneck of privacy-preserving collaborative learning using feature extraction is proposed. The methods are Shared Feature Extractor Learning (SFE) and Locally Trained Feature Extractor Learning (LTFE), along with the baseline solution that is referred to as Cooperatively Trained Feature Extractor Learning (CTFE).Another risk in utilizing AI algorithms is when ML models are adopted from third-party providers by individuals or companies due to their own limitations in processing capabilities and extensive data availability. It opens up possibilities for malicious actors to inject backdoors in these models, which can be activated by specific inputs leading to incorrect classifications. A method is proposed to detect a backdoored neural network (Badnet) and remove the backdoor with minimal effect on the clean accuracy of the network.Moreover, the inference accuracy of DNNs deployed on real-world hardware can be compromised by physical phenomena like transistor aging, which slows down their switching speeds and causes system-level timing issues due to unreliable clock speeds. To counteract the aging effects on DNN's accuracy and maintain their performance, a strategy that incorporates adding gradient noise during training along with the deployment of a deep imitator network designed to understand DNN-specific aging effects is employed. This approach ensures the creation of a resilient DNN, even when employing aggressive (that is, smaller than necessary) guardbands.Another challenge arises in the quest for explainable models that can interpret large data for actionable insights, such as detecting malicious activity within an enterprise network. Although DNNs can be a solution, their decision-making processes are often opaque and not easily understood by humans. A methodology (Sentences in Feature Subsets, i.e., SiFS) is proposed to mine human-readable decision rules from empirical data sets. The results on synthetic and real-world benchmark datasets are presented to demonstrate the efficacy of SiFS in deriving human-readable decision rules.

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

Mode of access: World Wide Web

ISBN: 9798382762722Subjects--Topical Terms:

573171
Computer science.
Subjects--Index Terms:

AI technologiesIndex Terms--Genre/Form:

554714
Electronic books.

Towards Scaling Artificial Intelligence for Resilience and Robustness.
LDR:06583ntm a22003977 4500 001 1147881
005 20240916075421.5
006 m o d
007 cr bn ---uuuuu
008 250605s2024 xx obm 000 0 eng d
020 $a 9798382762722
035 $a (MiAaPQ)AAI31243799
035 $a AAI31243799
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Sarmadi, Alireza. $3 1473715
245 1 0 $a Towards Scaling Artificial Intelligence for Resilience and Robustness.
264 0 $c 2024
300 $a 1 online resource (230 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: 85-11, Section: B.
500 $a Advisor: Khorrami, Farshad.
502 $a Thesis (Ph.D.)--New York University Tandon School of Engineering, 2024.
504 $a Includes bibliographical references
520 $a Integration of Artificial Intelligence (AI) into areas such as image classification, speech recognition, natural language processing, and autonomous systems illustrates its potential to revolutionize everyday life. However, this promise is tempered by significant vulnerabilities in AI technologies, including susceptibility to unpredictable behaviors and manipulations by adversaries. This dissertation focuses on the robustness and resilience of AI applications in domains such as high-dimensional controller tuning and biometrics identification. Moreover, this dissertation investigates robustness/security aspects of AI algorithms themselves, such as privacy-preserving cooperative learning between organizations, robustness against transistor aging effects on Deep Neural Networks (DNNs), mitigating backdoor attacks on DNNs, and explainability by mining human-readable decision rules from data.Data-driven control policies are utilized when an agent seeks to find a controller without any prior knowledge about the system dynamics by training an end-to-end controller (i.e., the input to the controller is the observed state of the system; the output is the actuation vector). Learning the controller typically requires large numbers of samples (e.g., tens of thousands) to achieve reasonable performance. However, collecting large amounts of data on real-world systems is challenging due to physical wear and tear, safety issues, and the time required. On the other hand, optimization-based control formulates the problem into an optimization problem. The performance of optimization-based methods is very sensitive to cost function constraints and system dynamics. These hyperparameters should be tuned carefully to achieve the desired behavior. A method is proposed to automatically and efficiently tune high-dimensional vectors of controller parameters.In the context of user identity verification, traditional methods like Personal Identification Numbers and passwords have proven vulnerable to security breaches, information leaks, and counterfeiting. Alternatives that rely on biometric features, including facial characteristics, fingerprints, and iris patterns, face risks from technological advancements that enable their duplication through methods like 3D printing and artificial iris contact lenses. In contrast, Surface Electromyography (sEMG) offers a more secure solution by leveraging the unique and complex neurophysiological responses of individuals, making impersonation and fraud significantly more challenging. An identification method is proposed that uses a CNN-based framework to identify subjects utilizing their sEMG data.AI algorithms suffer from weaknesses that make them vulnerable. Many organizations in different domains (e.g., enterprise networks and healthcare) are sharing data to increase the performance of their ML algorithms. However, sharing data puts the privacy of sample owners in danger. Privacy could be guaranteed by using encryption methodologies; however, these methods are computationally expensive for huge amounts of data. A new solution that seeks to address the computational bottleneck of privacy-preserving collaborative learning using feature extraction is proposed. The methods are Shared Feature Extractor Learning (SFE) and Locally Trained Feature Extractor Learning (LTFE), along with the baseline solution that is referred to as Cooperatively Trained Feature Extractor Learning (CTFE).Another risk in utilizing AI algorithms is when ML models are adopted from third-party providers by individuals or companies due to their own limitations in processing capabilities and extensive data availability. It opens up possibilities for malicious actors to inject backdoors in these models, which can be activated by specific inputs leading to incorrect classifications. A method is proposed to detect a backdoored neural network (Badnet) and remove the backdoor with minimal effect on the clean accuracy of the network.Moreover, the inference accuracy of DNNs deployed on real-world hardware can be compromised by physical phenomena like transistor aging, which slows down their switching speeds and causes system-level timing issues due to unreliable clock speeds. To counteract the aging effects on DNN's accuracy and maintain their performance, a strategy that incorporates adding gradient noise during training along with the deployment of a deep imitator network designed to understand DNN-specific aging effects is employed. This approach ensures the creation of a resilient DNN, even when employing aggressive (that is, smaller than necessary) guardbands.Another challenge arises in the quest for explainable models that can interpret large data for actionable insights, such as detecting malicious activity within an enterprise network. Although DNNs can be a solution, their decision-making processes are often opaque and not easily understood by humans. A methodology (Sentences in Feature Subsets, i.e., SiFS) is proposed to mine human-readable decision rules from empirical data sets. The results on synthetic and real-world benchmark datasets are presented to demonstrate the efficacy of SiFS in deriving human-readable decision rules.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2024
538 $a Mode of access: World Wide Web
650 4 $a Computer science. $3 573171
650 4 $a Information technology. $3 559429
650 4 $a Electrical engineering. $3 596380
653 $a AI technologies
653 $a Autonomous systems
653 $a ML algorithms
653 $a Datasets
655 7 $a Electronic books. $2 local $3 554714
690 $a 0800
690 $a 0489
690 $a 0984
690 $a 0544
710 2 $a ProQuest Information and Learning Co. $3 1178819
710 2 $a New York University Tandon School of Engineering. $b Electrical & Computer Engineering. $3 1437750
773 0 $t Dissertations Abstracts International $g 85-11B.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31243799 $z click for full text (PQDT)