{"id":863118,"date":"2022-07-18T19:38:50","date_gmt":"2022-07-19T02:38:50","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/"},"modified":"2022-07-21T08:01:51","modified_gmt":"2022-07-21T15:01:51","slug":"active-data-pattern-extraction-attacks-on-generative-language-models","status":"publish","type":"msr-research-item","link":"https:\/\/www.microsoft.com\/en-us\/research\/publication\/active-data-pattern-extraction-attacks-on-generative-language-models\/","title":{"rendered":"Active Data Pattern Extraction Attacks on Generative Language Models"},"content":{"rendered":"<p><span dir=\"ltr\" role=\"presentation\">With the wide availability of large pre-trained language model checkpoints, such as GPT-2 <\/span><span dir=\"ltr\" role=\"presentation\">and BERT, the recent trend has been to fine-tune them on a downstream task to achieve the <\/span><span dir=\"ltr\" role=\"presentation\">state-of-the-art performance with a small computation overhead.<\/span> <span dir=\"ltr\" role=\"presentation\">One natural example is the <\/span><span dir=\"ltr\" role=\"presentation\">Smart Reply application where a pre-trained model is fine-tuned for suggesting a number of <\/span><span dir=\"ltr\" role=\"presentation\">responses given a query message. In this work, we set out to investigate potential information <\/span><span dir=\"ltr\" role=\"presentation\">leakage vulnerabilities in a typical Smart Reply pipeline and show that it is possible for an <\/span><span dir=\"ltr\" role=\"presentation\">adversary, having black-box or gray-box access to a Smart Reply model, to extract sensitive <\/span><span dir=\"ltr\" role=\"presentation\">user information present in the training data. We further analyse the privacy impact of specific <\/span><span dir=\"ltr\" role=\"presentation\">components, e.g.,<\/span> <span dir=\"ltr\" role=\"presentation\">the decoding strategy, pertained to this application through our attack set<\/span><span dir=\"ltr\" role=\"presentation\">tings. We explore potential mitigation strategies and demonstrate how differential privacy can <\/span><span dir=\"ltr\" role=\"presentation\">be a strong defense mechanism to such data extraction attacks.<\/span><\/p>\n","protected":false},"excerpt":{"rendered":"<p>With the wide availability of large pre-trained language model checkpoints, such as GPT-2 and BERT, the recent trend has been to fine-tune them on a downstream task to achieve the state-of-the-art performance with a small computation overhead. One natural example is the Smart Reply application where a pre-trained model is fine-tuned for suggesting a number [&hellip;]<\/p>\n","protected":false},"featured_media":0,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"footnotes":""},"msr-content-type":[3],"msr-research-highlight":[],"research-area":[13556,13558],"msr-publication-type":[193718],"msr-product-type":[],"msr-focus-area":[],"msr-platform":[],"msr-download-source":[],"msr-locale":[268875],"msr-field-of-study":[],"msr-conference":[],"msr-journal":[],"msr-impact-theme":[],"msr-pillar":[],"class_list":["post-863118","msr-research-item","type-msr-research-item","status-publish","hentry","msr-research-area-artificial-intelligence","msr-research-area-security-privacy-cryptography","msr-locale-en_us"],"msr_publishername":"","msr_edition":"","msr_affiliation":"","msr_published_date":"2022-7-18","msr_host":"","msr_duration":"","msr_version":"","msr_speaker":"","msr_other_contributors":"","msr_booktitle":"","msr_pages_string":"","msr_chapter":"","msr_isbn":"","msr_journal":"","msr_volume":"","msr_number":"MSR-TR-2022-23","msr_editors":"","msr_series":"","msr_issue":"","msr_organization":"Microsoft","msr_how_published":"","msr_notes":"","msr_highlight_text":"","msr_release_tracker_id":"","msr_original_fields_of_study":"","msr_download_urls":"","msr_external_url":"","msr_secondary_video_url":"","msr_longbiography":"","msr_microsoftintellectualproperty":1,"msr_main_download":"","msr_publicationurl":"","msr_doi":"","msr_publication_uploader":[{"type":"file","viewUrl":"https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2022\/07\/SmartReplyAttack-1.pdf","id":"863124","title":"smartreplyattack-1","label_id":"243109","label":0}],"msr_related_uploader":"","msr_attachments":[{"id":863124,"url":"https:\/\/www.microsoft.com\/en-us\/research\/uploads\/prod\/2022\/07\/SmartReplyAttack-1.pdf"}],"msr-author-ordering":[{"type":"user_nicename","value":"Esha Ghosh","user_id":37851,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Esha Ghosh"},{"type":"text","value":"Bargav Jayaraman","user_id":0,"rest_url":false},{"type":"text","value":"Sambuddha Roy","user_id":0,"rest_url":false},{"type":"user_nicename","value":"Wei Dai","user_id":38013,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Wei Dai"},{"type":"user_nicename","value":"Huseyin Atahan Inan","user_id":40426,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Huseyin Atahan Inan"},{"type":"user_nicename","value":"Melissa Chase","user_id":32878,"rest_url":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/microsoft-research\/v1\/researchers?person=Melissa Chase"}],"msr_impact_theme":[],"msr_research_lab":[],"msr_event":[],"msr_group":[1054512],"msr_project":[],"publication":[],"video":[],"download":[],"msr_publication_type":"techreport","_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/863118"}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-research-item"}],"version-history":[{"count":4,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/863118\/revisions"}],"predecessor-version":[{"id":863631,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-item\/863118\/revisions\/863631"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=863118"}],"wp:term":[{"taxonomy":"msr-content-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-content-type?post=863118"},{"taxonomy":"msr-research-highlight","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-research-highlight?post=863118"},{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=863118"},{"taxonomy":"msr-publication-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-publication-type?post=863118"},{"taxonomy":"msr-product-type","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-product-type?post=863118"},{"taxonomy":"msr-focus-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-focus-area?post=863118"},{"taxonomy":"msr-platform","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-platform?post=863118"},{"taxonomy":"msr-download-source","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-download-source?post=863118"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=863118"},{"taxonomy":"msr-field-of-study","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-field-of-study?post=863118"},{"taxonomy":"msr-conference","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-conference?post=863118"},{"taxonomy":"msr-journal","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-journal?post=863118"},{"taxonomy":"msr-impact-theme","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-impact-theme?post=863118"},{"taxonomy":"msr-pillar","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-pillar?post=863118"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}