New li­brary to se­cure AI sys­tems

OpenSource For You - - Fossbytes -

IBM has re­leased a se­cu­rity li­brary into the open source com­mu­nity that is de­signed to help pro­tect ar­ti­fi­cial in­tel­li­gence (AI) sys­tems. IBM’s aim is to make this tool­box be­come a repos­i­tory and source of in­for­ma­tion on threats to cur­rent and fu­ture AI so­lu­tions.

Cer­tain weak­nesses in AI sys­tems expose them to ex­ploita­tion, such as un­de­tectable al­ter­ations in con­tent in­clud­ing images, video, and au­dio record­ings be­ing crafted by those with ma­li­cious in­tent, for which one does not need a deep knowl­edge of AI.

These changes can be small in size but re­sult in huge se­cu­rity breaches. They im­pact the per­for­mance of AI sys­tems like prompt­ing them to make a choice which would be deemed ma­li­cious.

Aimed to com­bat so called ‘Ad­ver­sar­ial AI’, the tool­box records threat data as well as as­sists de­vel­op­ers in cre­at­ing, bench­mark­ing and de­ploy­ing prac­ti­cal de­fence sys­tems for real-world AI. IBM shared that this re­search looks at the best ways to de­fend the AI sys­tems be­fore the bad guys at­tack.

By in­tro­duc­ing the tool­kit to the open source com­mu­nity, oth­ers may also be­come in­spired enough to cre­ate so­lu­tions be­fore Ad­ver­sar­ial AI be­comes a true threat. The tool­box also in­cludes a li­brary, in­ter­faces, and met­rics that will help de­vel­op­ers be­gin to cre­ate cy­ber se­cu­rity so­lu­tions for this emerg­ing field.

“Con­sid­er­ing tools didn’t pro­vide the de­fences needed to pro­tect AI sys­tems, this is the first and only AI li­brary that con­tains at­tacks, de­fences, and bench­marks to im­ple­ment im­proved se­cu­rity,” the com­pany sources have said.

Newspapers in English

Newspapers from India

© PressReader. All rights reserved.