Measuring and Comparing Robustness of ML Algorithms Under Adversarial Attack

reportActive / Technical Report | Accession Number: AD1088314 | Open PDF

Abstract:

A machine learning algorithm can be evaluated for robustness against any number of different types of attacks. We consider attacks that seek to manipulate the training andor testing data inputs to a machine learning algorithm. Specifically, we do not consider physical attacks on machines hosting the algorithm.

Author(s):
Kanal, Eliezer
Author Organization(s):
CARNEGIE-MELLON UNIV PITTSBURGH PA PITTSBURGH United States
Descriptive Note:
Technical Report
Pagination:
0006

Security Markings

DOCUMENT & CONTEXTUAL SUMMARY

Distribution:
Approved For Public Release
Distribution Statement:
Approved For Public Release;

RECORD

Collection: TR
Identifying Numbers
Report Number(s):
DM17-0850
Contract/Grant Number(s):
FA8702-15-D-0002
 Subject Terms
Joint Capability Areas:
JCA_1.2.1_Training; JCA_1.2.3_Educating; JCA_4.6_Engineering; JCA_5_Command and Control; JCA_5.3_Planning; JCA_1.2.5_Lessons Learned
Modernization Areas:
Autonomy
Communities of Interest:
Autonomy
Descriptor(s):
algorithms, machine learning, cyberattacks, artificial intelligence, vulnerability, cybersecurity
Field(s)/Group(s):
Computer Programming and Software
Keyword(s):
robustness, performance, information leakage
Report Date:
2017 Aug 01
Creation Date:
2020 Jan 03