BEGIN:VCALENDAR VERSION:2.0 PRODID:IEEE vTools.Events//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:America/Chicago BEGIN:DAYLIGHT DTSTART:20240310T030000 TZOFFSETFROM:-0600 TZOFFSETTO:-0500 RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3 TZNAME:CDT END:DAYLIGHT BEGIN:STANDARD DTSTART:20241103T010000 TZOFFSETFROM:-0500 TZOFFSETTO:-0600 RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11 TZNAME:CST END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20240328T212342Z UID:40B7E78C-D8F3-4017-9C5F-5C0B998AC9D6 DTSTART;TZID=America/Chicago:20240314T120000 DTEND;TZID=America/Chicago:20240314T130000 DESCRIPTION:Recent research has revealed that machine learning models are v ulnerable to adversarial attacks which seek to manipulate the model to ind uce undesired behavior or extract sensitive information. Such vulnerabilit ies are particularly concerning for the use of these methods in aerospace and defense applications where safety and security are paramount. This res earch evaluates an approach that combines several attack detection methods in tandem to produce an intrusion detection system (IDS) that ensures sec urity at each stage of the model’s lifecycle. The performance of the pip elined detection approach is compared to the performance of each individua l detector with the hypothesis that the Combined IDS will result in improv ed security. The goal of this research is to move toward a practice for se cure AI development and operation.\n\nCo-sponsored by: WIE\n\nSpeaker(s): Garrett\, \n\nBldg: 51\, 1100 Martin Goland Ave\, San Antonio\, Texas\, Un ited States\, 78238 LOCATION:Bldg: 51\, 1100 Martin Goland Ave\, San Antonio\, Texas\, United S tates\, 78238 ORGANIZER:sriram.nagaraj@swri.org SEQUENCE:6 SUMMARY:Secure Development of Machine Learning Against Poisoning Attacks URL;VALUE=URI:https://events.vtools.ieee.org/m/409446 X-ALT-DESC:Description: <br /><p><span class="NormalTextRun SCXW52503456 BC X9">Recent research has revealed that machine learning </span><span class= "NormalTextRun SCXW52503456 BCX9">models</span><span class="NormalTextRun SCXW52503456 BCX9"> are vulnerable to adversarial attacks which </span><sp an class="NormalTextRun SCXW52503456 BCX9">seek</span><span class="NormalT extRun SCXW52503456 BCX9"> to manipulate the model to induce undesired beh avior or extract sensitive information. Such vulnerabilities are particula rly concerning for the use of these </span><span class="NormalTextRun SCXW 52503456 BCX9">methods</span><span class="NormalTextRun SCXW52503456 BCX9" > in </span><span class="NormalTextRun SCXW52503456 BCX9">aerospace and </ span><span class="NormalTextRun SCXW52503456 BCX9">defense applications wh ere safety and security are paramount. </span><span class="NormalTextRun S CXW52503456 BCX9">This research evaluates an approach </span><span class=" NormalTextRun SCXW52503456 BCX9">that</span> <span class="NormalTextRun SC XW52503456 BCX9">combin</span><span class="NormalTextRun SCXW52503456 BCX9 ">e</span><span class="NormalTextRun SCXW52503456 BCX9">s</span><span clas s="NormalTextRun SCXW52503456 BCX9"> several</span><span class="NormalText Run SCXW52503456 BCX9"> attack</span><span class="NormalTextRun SCXW525034 56 BCX9"> detection methods in tandem to produce a</span><span class="Norm alTextRun SCXW52503456 BCX9">n </span><span class="NormalTextRun SCXW52503 456 BCX9">intrusion detection system (IDS) that ensures security at each s tage of the model&rsquo\;s lifecycle. The performance of the pipelined det ection approach is compared to the performance of each individual detector with the hypothesis that the Combined IDS will result in improved securit y.</span> <span class="NormalTextRun SCXW52503456 BCX9">T</span><span clas s="NormalTextRun SCXW52503456 BCX9">he goal of this research is to move to ward a practice for secure AI development and operation.</span></p> END:VEVENT END:VCALENDAR