DNNShield: Dynamic Randomized Model Sparsification, A Defense Against Adversarial Machine Learning

DNNs are known to be vulnerable to so-called adversarial attacks that manipulate inputs to cause incorrect results that can be beneficial to an attacker or damaging to the victim. Recent works have proposed approximate computation as a defense mechanism against machine learning attacks. We show that...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Samavatian, Mohammad Hossein, Majumdar, Saikat, Barber, Kristin, Teodorescu, Radu
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext bestellen
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
container_end_page
container_issue
container_start_page
container_title
container_volume
creator Samavatian, Mohammad Hossein
Majumdar, Saikat
Barber, Kristin
Teodorescu, Radu
description DNNs are known to be vulnerable to so-called adversarial attacks that manipulate inputs to cause incorrect results that can be beneficial to an attacker or damaging to the victim. Recent works have proposed approximate computation as a defense mechanism against machine learning attacks. We show that these approaches, while successful for a range of inputs, are insufficient to address stronger, high-confidence adversarial attacks. To address this, we propose DNNSHIELD, a hardware-accelerated defense that adapts the strength of the response to the confidence of the adversarial input. Our approach relies on dynamic and random sparsification of the DNN model to achieve inference approximation efficiently and with fine-grain control over the approximation error. DNNSHIELD uses the output distribution characteristics of sparsified inference compared to a dense reference to detect adversarial inputs. We show an adversarial detection rate of 86% when applied to VGG16 and 88% when applied to ResNet50, which exceeds the detection rate of the state of the art approaches, with a much lower overhead. We demonstrate a software/hardware-accelerated FPGA prototype, which reduces the performance impact of DNNSHIELD relative to software-only CPU and GPU implementations.
doi_str_mv 10.48550/arxiv.2208.00498
format Article
fullrecord <record><control><sourceid>arxiv_GOX</sourceid><recordid>TN_cdi_arxiv_primary_2208_00498</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>2208_00498</sourcerecordid><originalsourceid>FETCH-LOGICAL-a678-cf7e15c311c3b529287803616bd84452a5af36ce465c049e64dbcea87059b43e3</originalsourceid><addsrcrecordid>eNotz8tOwzAQhWFvWKDCA7DCD0CCE1_isIsaLpXaItHuo4k9aUdKnMqpKsrTA4XV2f06H2N3mUiV1Vo8QvykU5rnwqZCqNJes7Zerzd7wt4_8focYCDHPyD4caAv9Hw1euz55gBxoo4cHGkMD7ziNXYYJuTVDihMR175E8YJIkHPV-D2FJAvEWKgsLthVx30E97-74xtX56387dk-f66mFfLBExhE9cVmGkns8zJVudlbgsrpMlM661SOgcNnTQOldHu5zoa5VuHYAuhy1ZJlDN2_5e9IJtDpAHiufnFNhes_AbW708K</addsrcrecordid><sourcetype>Open Access Repository</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype></control><display><type>article</type><title>DNNShield: Dynamic Randomized Model Sparsification, A Defense Against Adversarial Machine Learning</title><source>arXiv.org</source><creator>Samavatian, Mohammad Hossein ; Majumdar, Saikat ; Barber, Kristin ; Teodorescu, Radu</creator><creatorcontrib>Samavatian, Mohammad Hossein ; Majumdar, Saikat ; Barber, Kristin ; Teodorescu, Radu</creatorcontrib><description>DNNs are known to be vulnerable to so-called adversarial attacks that manipulate inputs to cause incorrect results that can be beneficial to an attacker or damaging to the victim. Recent works have proposed approximate computation as a defense mechanism against machine learning attacks. We show that these approaches, while successful for a range of inputs, are insufficient to address stronger, high-confidence adversarial attacks. To address this, we propose DNNSHIELD, a hardware-accelerated defense that adapts the strength of the response to the confidence of the adversarial input. Our approach relies on dynamic and random sparsification of the DNN model to achieve inference approximation efficiently and with fine-grain control over the approximation error. DNNSHIELD uses the output distribution characteristics of sparsified inference compared to a dense reference to detect adversarial inputs. We show an adversarial detection rate of 86% when applied to VGG16 and 88% when applied to ResNet50, which exceeds the detection rate of the state of the art approaches, with a much lower overhead. We demonstrate a software/hardware-accelerated FPGA prototype, which reduces the performance impact of DNNSHIELD relative to software-only CPU and GPU implementations.</description><identifier>DOI: 10.48550/arxiv.2208.00498</identifier><language>eng</language><subject>Computer Science - Cryptography and Security ; Computer Science - Hardware Architecture ; Computer Science - Learning</subject><creationdate>2022-07</creationdate><rights>http://creativecommons.org/licenses/by/4.0</rights><oa>free_for_read</oa><woscitedreferencessubscribed>false</woscitedreferencessubscribed></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><link.rule.ids>228,230,780,885</link.rule.ids><linktorsrc>$$Uhttps://arxiv.org/abs/2208.00498$$EView_record_in_Cornell_University$$FView_record_in_$$GCornell_University$$Hfree_for_read</linktorsrc><backlink>$$Uhttps://doi.org/10.48550/arXiv.2208.00498$$DView paper in arXiv$$Hfree_for_read</backlink></links><search><creatorcontrib>Samavatian, Mohammad Hossein</creatorcontrib><creatorcontrib>Majumdar, Saikat</creatorcontrib><creatorcontrib>Barber, Kristin</creatorcontrib><creatorcontrib>Teodorescu, Radu</creatorcontrib><title>DNNShield: Dynamic Randomized Model Sparsification, A Defense Against Adversarial Machine Learning</title><description>DNNs are known to be vulnerable to so-called adversarial attacks that manipulate inputs to cause incorrect results that can be beneficial to an attacker or damaging to the victim. Recent works have proposed approximate computation as a defense mechanism against machine learning attacks. We show that these approaches, while successful for a range of inputs, are insufficient to address stronger, high-confidence adversarial attacks. To address this, we propose DNNSHIELD, a hardware-accelerated defense that adapts the strength of the response to the confidence of the adversarial input. Our approach relies on dynamic and random sparsification of the DNN model to achieve inference approximation efficiently and with fine-grain control over the approximation error. DNNSHIELD uses the output distribution characteristics of sparsified inference compared to a dense reference to detect adversarial inputs. We show an adversarial detection rate of 86% when applied to VGG16 and 88% when applied to ResNet50, which exceeds the detection rate of the state of the art approaches, with a much lower overhead. We demonstrate a software/hardware-accelerated FPGA prototype, which reduces the performance impact of DNNSHIELD relative to software-only CPU and GPU implementations.</description><subject>Computer Science - Cryptography and Security</subject><subject>Computer Science - Hardware Architecture</subject><subject>Computer Science - Learning</subject><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2022</creationdate><recordtype>article</recordtype><sourceid>GOX</sourceid><recordid>eNotz8tOwzAQhWFvWKDCA7DCD0CCE1_isIsaLpXaItHuo4k9aUdKnMqpKsrTA4XV2f06H2N3mUiV1Vo8QvykU5rnwqZCqNJes7Zerzd7wt4_8focYCDHPyD4caAv9Hw1euz55gBxoo4cHGkMD7ziNXYYJuTVDihMR175E8YJIkHPV-D2FJAvEWKgsLthVx30E97-74xtX56387dk-f66mFfLBExhE9cVmGkns8zJVudlbgsrpMlM661SOgcNnTQOldHu5zoa5VuHYAuhy1ZJlDN2_5e9IJtDpAHiufnFNhes_AbW708K</recordid><startdate>20220731</startdate><enddate>20220731</enddate><creator>Samavatian, Mohammad Hossein</creator><creator>Majumdar, Saikat</creator><creator>Barber, Kristin</creator><creator>Teodorescu, Radu</creator><scope>AKY</scope><scope>GOX</scope></search><sort><creationdate>20220731</creationdate><title>DNNShield: Dynamic Randomized Model Sparsification, A Defense Against Adversarial Machine Learning</title><author>Samavatian, Mohammad Hossein ; Majumdar, Saikat ; Barber, Kristin ; Teodorescu, Radu</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-LOGICAL-a678-cf7e15c311c3b529287803616bd84452a5af36ce465c049e64dbcea87059b43e3</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2022</creationdate><topic>Computer Science - Cryptography and Security</topic><topic>Computer Science - Hardware Architecture</topic><topic>Computer Science - Learning</topic><toplevel>online_resources</toplevel><creatorcontrib>Samavatian, Mohammad Hossein</creatorcontrib><creatorcontrib>Majumdar, Saikat</creatorcontrib><creatorcontrib>Barber, Kristin</creatorcontrib><creatorcontrib>Teodorescu, Radu</creatorcontrib><collection>arXiv Computer Science</collection><collection>arXiv.org</collection></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext_linktorsrc</fulltext></delivery><addata><au>Samavatian, Mohammad Hossein</au><au>Majumdar, Saikat</au><au>Barber, Kristin</au><au>Teodorescu, Radu</au><format>journal</format><genre>article</genre><ristype>JOUR</ristype><atitle>DNNShield: Dynamic Randomized Model Sparsification, A Defense Against Adversarial Machine Learning</atitle><date>2022-07-31</date><risdate>2022</risdate><abstract>DNNs are known to be vulnerable to so-called adversarial attacks that manipulate inputs to cause incorrect results that can be beneficial to an attacker or damaging to the victim. Recent works have proposed approximate computation as a defense mechanism against machine learning attacks. We show that these approaches, while successful for a range of inputs, are insufficient to address stronger, high-confidence adversarial attacks. To address this, we propose DNNSHIELD, a hardware-accelerated defense that adapts the strength of the response to the confidence of the adversarial input. Our approach relies on dynamic and random sparsification of the DNN model to achieve inference approximation efficiently and with fine-grain control over the approximation error. DNNSHIELD uses the output distribution characteristics of sparsified inference compared to a dense reference to detect adversarial inputs. We show an adversarial detection rate of 86% when applied to VGG16 and 88% when applied to ResNet50, which exceeds the detection rate of the state of the art approaches, with a much lower overhead. We demonstrate a software/hardware-accelerated FPGA prototype, which reduces the performance impact of DNNSHIELD relative to software-only CPU and GPU implementations.</abstract><doi>10.48550/arxiv.2208.00498</doi><oa>free_for_read</oa></addata></record>
fulltext fulltext_linktorsrc
identifier DOI: 10.48550/arxiv.2208.00498
ispartof
issn
language eng
recordid cdi_arxiv_primary_2208_00498
source arXiv.org
subjects Computer Science - Cryptography and Security
Computer Science - Hardware Architecture
Computer Science - Learning
title DNNShield: Dynamic Randomized Model Sparsification, A Defense Against Adversarial Machine Learning
url https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2025-01-01T00%3A36%3A21IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-arxiv_GOX&rft_val_fmt=info:ofi/fmt:kev:mtx:journal&rft.genre=article&rft.atitle=DNNShield:%20Dynamic%20Randomized%20Model%20Sparsification,%20A%20Defense%20Against%20Adversarial%20Machine%20Learning&rft.au=Samavatian,%20Mohammad%20Hossein&rft.date=2022-07-31&rft_id=info:doi/10.48550/arxiv.2208.00498&rft_dat=%3Carxiv_GOX%3E2208_00498%3C/arxiv_GOX%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_id=info:pmid/&rfr_iscdi=true