ãã®è¨äºã¯ãæ å ±ã»ãã¥ãªãã£ç³»è«æç´¹ä» Advent Calendar 2016ã14æ¥ç®ã®è¨äºã§ãã
è¿å¹´ããã£ã¼ãã©ã¼ãã³ã°ã¨å¼ã°ããæ©æ¢°å¦ç¿ææ³ã®é²å±ãããã¾ã£ã¦ããã£ã¼ãã©ã¼ãã³ã°ã§ã¯ãªãæ©æ¢°å¦ç¿ãããããã®æ³¨ç®ãéãã¦ããã ããã§ã¯ã2016å¹´ã«å ¬è¡¨ãããæ©æ¢°å¦ç¿ç³»ã®æ å ±ã»ãã¥ãªãã£è«æã«ã¤ãã¦ãæ°ã«ãªã£ããã®ãã¾ã¨ãã¦ã¿ãã
discovRE: Efficient Cross-Architecture Identification of Bugs in Binary Code (NDSS 2016)
å½ä»¤æ°çã®è¤æ°ã®æ°å¤ææ¨ãç¨ãã¦k-Nearest Neighborã«ãããã£ã«ã¿ãªã³ã°ãè¡ã£ãå¾ãMaximum common subgraphï¼MCSï¼ã«ããControl Flow Graphã®é¡ä¼¼åº¦æ¯è¼ãè¡ããã¨ã§ããã¤ããªï¼ãã¡ã¼ã ã¦ã§ã¢ï¼ããæ¢ç¥ã®èå¼±æ§ãå«ãé¢æ°ãã¯ãã¹ã¢ã¼ããã¯ãã£ã§åå®ããã æ¢åææ³ã®1000åãªã¼ãã¼ã®é度ãåºãã¨ã®ãã¨ã ããåå¦çãããç ©éãªã®ã¨ãèå¼±æ§ãä¿®æ£ãããåã¨å¾ã®éããè¦åãããããã«ã¤ãã¦ï¼éæ¥çã«ï¼æ¬¡ã®ããã«ãã触ãããã¦ããªãã®ãæ°ã«ãªãã
However, the vast majority of bugs can be pinpointed to one or a list of specific functions.
CFGã®é¡ä¼¼åº¦æ¯è¼ããã¨ã«ãã¦ãããããifæã®è¿½å çã«ããã°ã©ãã®éããæçµçãªçµæã«è¡¨ããã¯ãã ãããããåºå¥å¯è½ã§ãããã«ã¤ãã¦ã¯è§¦ãããã¦ããªãã ã¨ã¯ãããk-NNã§ã®ãã£ã«ã¿ãªã³ã°ã«ç¨ããææ¨ã®é¸æããã¤ãã¼ãã©ã¡ã¼ã¿ã®è¨å®ã«ã¤ãã¦è©³ç´°ãªåæãè¡ããã¦ãããåèã«ãªãã
Automatically Evading Classifiers: A Case Study on PDF Malware Classifiers (NDSS 2016)
ã»ãã¥ãªãã£åéã«ãããæ©æ¢°å¦ç¿ãä»åéã¨ç°ãªãç¹ã¨ãã¦æ»æè ã«ãããã¤ãã¹ãèæ ®ããªãã¨ãããªããã¨ãææããä¸ã§ãéºä¼çã¢ã«ã´ãªãºã ã«ããbenignãªPDFã®ç¹å¾´ãåãå ¥ãããã¨ã§æªæ§PDFåé¡å¨ããã¤ãã¹ããPDFãèªåçæããã çµæãPDFrate [ACSACâ12]ãHidost [NDSSâ13] ã100%ãã¤ãã¹ããPDFãèªåçæã§ããã ã¾ãã対HidostãªPDFã§ããç¨åº¦PDFrateããã¤ãã¹ã§ããããGmailã®åé¡å¨ã¯ãã¤ãã¹ã§ããªãã£ãã ããã§ãGmailã®åé¡å¨ããã¤ãã¹ããã«ã¼ãã³ã追å ãããã¨ã§ã47.1%ã®ç¢ºçã§ãã¤ãã¹ã§ããPDFãèªåçæã§ããã
æ»æè ãæ¬æ°ãåºãã¦ããã»ãã¥ãªãã£åéã§ã¯ãrobustãªåé¡å¨ãä½ãã®ãé£ãããã¨ãã話ã
Fast, Lean, and Accurate: Modeling Password Guessability Using Neural Networks (USENIX Security 2016)
ããããã¥ã¼ã©ã«ããããããããã¥ã¼ã©ã«ããã大好ãï¼åèï¼ã
Recurrent Neural Networkï¼RNNï¼ã§ãã¹ã¯ã¼ãã®æ¨æ¸¬ãããããè¨ç®ãããªã¢ã«ã¿ã¤ã ã§ã¦ã¼ã¶ã«ãã£ã¼ãããã¯ãããã¨ãèããã æµåºãããã¹ã¯ã¼ãã®ãã¼ã¿ã»ãããæ師ãã¼ã¿ãæ¢åç 究ã®ãã¹ã¯ã¼ããã¼ã¿ã»ããã¨000webhostããã®æµåºãã¹ã¯ã¼ãããã¹ããã¼ã¿ã¨ããã¢ã³ãã«ã«ãæ³ã使ã£ã¦è©ä¾¡ããã¨ããã次ã®ãããªãã¨ãããã£ãã
- 1ã¯ã©ã¹8æåã®ãããã¯ã¼ã¯â3ã¯ã©ã¹12æåã®ãããã¯ã¼ã¯ã®ãããªè»¢ç§»å¦ç¿ããããã¨ã§ç²¾åº¦ãä¸ãã
- èªç¶è¨èªãã¼ã¿ãå ãã¦ãããã»ã©ç²¾åº¦ã¯ä¸ãããªã
- ã¢ãã«ãµã¤ãºã大ãããããã¨ã«ãã精度åä¸ã¯æ¡ä»¶ã«ãã
ãã«ã³ãã¢ãã«ãPCFGãHashcatãJohn the Ripperã¨ã®æ¯è¼ã«ããã¦ãRNNã¯ã©ããããæ§è½ãããã ã¾ããå§ç¸®ãäºåè¨ç®ãè¡ããã¨ã§ã¢ãã«ãµã¤ãº850KBãè¨ç®æé17msç¨åº¦ã«ã§ãã精度ãè½ã¨ããã¨ãªããã©ã¦ã¶ä¸ã§ã®ãªã¢ã«ã¿ã¤ã ãã£ã¼ãããã¯ã«é©ç¨ã§ããã
ã³ã¼ããGithubã§å ¬éããã¦ãããããªãã¸ããªåããneural_network_crackingããªã®ã趣深ãã
Stealing Machine Learning Models via Prediction APIs (USENIX Security 2016)
BigMLãAmazon Machine Learningã®ãããªMachine Learning as a Service (MLaaS)ã¨ãããã®ãä¸ã®ä¸ã«åºã¦ãã¦ããããæåãµã¼ãã¹ã§ããããã«æçµçãªåé¡çµæ以å¤ã«ãconfidenceå¤ãªã©ã®ä»å æ å ±ãè¿ã£ã¦ããã ãããæªç¨ãããã¨ã§ãæå¡©ã«ããã¦å¦ç¿ãããã¢ãã«ãã©ã¡ã¼ã¿ã第ä¸è ã«ã³ãã¼ãããå¯è½æ§ãããã ãã®ãããªModel-Extraction Attacksã«é¢ããæ¢åç 究ã®ã¾ã¨ãã ããã«ãconfidenceå¤ãé ãã°ããã¨ãããã®ã§ããªããæ©æ¢°å¦ç¿ã§æ©æ¢°å¦ç¿ãµã¼ãã¹ã®ã¢ãã«ãã©ã¡ã¼ã¿ãæ¨å®ãããã¨èªä½ãå¯è½ã§ãããã¨ãææãã¦ããã
Black Hat USAã¿ãããªå 容ã ãUSENIX Securityãªã®ããããããã
Distillation as a Defense to Adversarial Perturbations against Deep Neural Networks (IEEE Security and Privacy 2016)
Deep Neural Networkï¼DNNï¼ã«å¯¾ãã¦ãæ»æè ãç¹°ãè¿ãå ¥åã調æ´ãããã¨ã«ããDNNã誤åé¡ããå ¥åãä½ãåºããã¨ãã§ããAdversarial Sample Craftingã¨ããåé¡ãç¥ããã¦ããã ãã®ãããªåé¡ã¸ã®é²å¾¡çã¨ãã¦ãDNNã®ã¢ãã«ãµã¤ãºã®åæ¸ã«ç¨ããããDistillationã¨ããææ³ãæå¹ã§ããã å ·ä½çã«ã¯ãåºå層ã®Softmaxã®ãã©ã¡ã¼ã¿Tï¼æ¸©åº¦ã¨å¼ã°ããï¼ã®å¤§ããªãããã¯ã¼ã¯ã§åé¡ããå¾ããã®çµæã追å æ å ±ã¨ãã¦T=1ã®DNNã§åé¡ããããã«ããã T=40ç¨åº¦ã§Distillationãããã¨ã§ãAdversarial Sampleã«ãã誤åé¡ã大ããæ¸ãããã¨ãã§ããã
æ¥æ¬ã§ãæ¨å¹´åå©å¥ç·å¤å®botã®èª¤èªèãã¿ã¼ã³ããããæ¦ãã¨ããã®ããã£ããããããã人å調æ´ã«ãæå¹ãªã®ããæ°ã«ãªãã
SandPrint: Fingerprinting Malware Sandboxes to Provide Intelligence for Sandbox Evasion (RAID 2016)
å®ã¦ã¼ã¶ã®ãã·ã³ã¨ãµã³ãããã¯ã¹ãèå¥ããæ±ç¨çãªåé¡å¨ã¯ä½ãããï¼ã¨ããåãã«å¯¾ãã¦ãä½ããã¨ãããã¨ã示ããè«æã ãµã³ãããã¯ã¹ã®ç°å¢æ å ±ï¼fingerprintï¼ãHTTPã§æãåãããã°ã©ã SandPrintãéçºããå®ã¦ã¼ã¶ã®ç°å¢æ å ±ã¨ããããåé¡ããåé¡å¨ãã¬ã¦ã·ã¢ã³ã«ã¼ãã«ãç¨ããéç·å½¢SVMã§ä½ãã çµæãã¡ã¢ãªã«é¢ããç¹å¾´éã ãã§ã98.06%ã®ç²¾åº¦ãåºãã ããã«ããã¹ã¦ã®ç¹å¾´éãç¨ããå ´åã¯false positiveãfalse negativeããªã100%ã®ç²¾åº¦ãéæã§ããã ã¾ããä¸ã¤ã®åç¨ãµã³ãããã¯ã¹ã«å¯¾ãã¦ãåæ§ã®ã¢ããã¼ããæå¹ã§ãããã¨ã確èªããã
ç°å¢æ å ±ããã¨ã«ãããµã³ãããã¯ã¹æ¤ç¥èªä½ã¯ãããªãã«ç¥ããã¦ããããå®éã«ããããåéãã¦æ±ç¨çã«ä½¿ããææ¨ãåæããã¨ããç¹ã§ãããããã
(Semi)-Supervised Machine Learning Approaches for Network Security in High-Dimensional Network Data (ACM CCS 2016 Poster)
é«æ¬¡å ã®ç¹å¾´éãæã¤ãããã¯ã¼ã¯ãã¼ã¿ããDDoSãHTTP flashcrowdsçã®ç°å¸¸ãã±ãããæ¤ç¥ããåæ師ããå¦ç¿ææ³ã¨ãã¦ãk-Meansã¯ã©ã¹ã¿ãªã³ã°ã¨k-Nearest Neighborã«ããå¤æ°æ±ºãçµã¿åãããk-CDA (k-means Clustering-based Detector of Attacks) ãææ¡ã k-Meansã®Kã¯è¨ç·´ãã¼ã¿ã®ãµã³ãã«ãµã¤ãºã®1000åã®1ã¨ããk-NNã®kã¯Kã®3åã®1ã¨ããã å®å ¨ãªæ師ããå¦ç¿ã§ããC4.5決å®æ¨ãRandom ForestãSVMããã¤ã¼ããã¤ãºãå¤å±¤ãã¼ã»ãããã³ã¨æ¯è¼ããçµæãã©ãã«ä»ãæ師ãã¼ã¿ã®5%ã®ã¿ããç¨ãã¦ããªãã«ãé¢ããããä½false positiveã®é åã§C4.5決å®æ¨ã«å£ããªã精度ãåºãã ãã¾ã精度ã®åºãªãã£ãDDoSãã±ããæ¤ç¥ã«ã¤ãã¦ã¯ãç¸é¢ãã¼ã¹ã®Best-first searchã§ç¹å¾´éã245åãã22åã«çµããã¨ã§ç²¾åº¦ãåä¸ããã
ã¸ã¼ãã¨ããæãã
Static ROP Chain Detection Based on Hidden Markov Model Considering ROP Chain Integrity (ACM CCS 2016 Poster)
é ããã«ã³ãã¢ãã«ã§ææ¸åãã«ã¦ã§ã¢ã«å«ã¾ããROP chainãæ¤ç¥ããææ³ã®ææ¡ã ROP chainããã¾ãç¹ãã£ã¦ãããããã§ãã¯ãããã¨ã§ãfalse positiveãæ¸ãã工夫ããã¦ããã è©ä¾¡ã®çµæãé«ã¹ã«ã¼ãããã§false negativeã¼ããä½false positiveãªåé¡ãã§ããã
èªåã¯ç 究è ã§ã¯ãªãã®ã§abstructããèªãã¦ããªãã
追è¨
ACM CCSã¨ãã®ã¯ã¼ã¯ã·ã§ããã®è«æã¯1å¹´ééå®ã§èªããã¨ãããã¨ãæãã¦ããã£ãã
32 bitç°å¢ã§ã®ROP chainã«ã¤ãã¦ããã¤ãåä½ã§ã®é·ç§»éç¨ã次ã®å³ã®ããã«ã¢ããªã³ã°ããã
ããã§ãDã¯ææ¸ãAã¯ã¢ãã¬ã¹ãCã¯å®æ°ãJã¯ROPä¸ã®junkã表ãã ã¾ããAã«ã¤ãã¦ã¯ã©ã¤ãã©ãªã«å«ã¾ããROP gadgetåè£ãå ¨åæãã¦ç¢ºçã¢ãã«ã«çµã¿è¾¼ãã ããã«ãROP chainããã¾ãç¹ãã£ã¦ãããï¼ROP Chain Integrityï¼ã®ãã§ãã¯ãè¡ãã å ·ä½çã«ã¯ãåã¢ãã¬ã¹åè£ã«é£ãã å ã§ã¹ã¿ãã¯ãã¤ã³ã¿ãä½ã¯ã¼ããããããã·ã³ããªãã¯å®è¡ãç¨ãã¦è¨ç®ãã¦ãããåã¢ãã¬ã¹ã«é£ãã å¾ã®ã¹ã¿ãã¯ãã¤ã³ã¿ãã¡ããã¨æ¬¡ã®ã¢ãã¬ã¹ï¼ä¸å³ã®A1ã«å¯¾å¿ããç®æï¼ãæãã¦ãããããã§ãã¯ããã çµæãfalse negativeã¼ããfalse positive 3%ã®ç²¾åº¦ã§æ¤ç¥ã§ããã ã¾ãã1ãã¡ã¤ã«2.5ç§ã®å¦çé度ãåºãã
ææ¸ããã¤ãåã¨ã¿ãªããé ããã«ã³ãã¢ãã«ã§ã®å°¤åº¦æ¯ãã¿ãã¨ããæ£çµ±æ´¾ã¢ããã¼ãã試ã¿ã¦ãã¦èå³æ·±ãã
AdversariaLib: An Open-source Library for the Security Evaluation of Machine Learning Algorithms Under Attack (arXiv preprint)
å¾é éä¸æ³ã§æ©æ¢°å¦ç¿ã誤åé¡ããå ¥åãçæããOSSã©ã¤ãã©ãªã®ç´¹ä»ã ããã¯ã¨ã³ãã«scikit-learnã¨FANNãç¨ãã¦ããã è«æä¸ã§ã¯ãããããªã3ãã®ç»åãã3ã¨åé¡ãããªãã3ãã®ãããªç»åãä½ãåºãä¾ãç´¹ä»ããã¦ããã ã¾ããåè¿°ã®è«æã«ããã£ãããã«ãå ã®åé¡å¨ã«è¿ãåé¡ãããæ°ããªåé¡å¨ãä½ãåºãããã¨ã«ã触ãããã¦ããã
å ã®æ©æ¢°å¦ç¿ã«å¯¾æããéã®æ©æ¢°å¦ç¿ã©ã¤ãã©ãªï¼æ¤è¨¼ç¨ï¼ã¨ãã£ãæãã
Applied Machine Learning for Data Exfil and Other Fun Topics (Black Hat USA 2016)
æ£ç¢ºã«ã¯è«æã§ã¯ãªãããæ©æ¢°å¦ç¿ã売ãã«ããã¨ã³ããã¤ã³ãã»ãã¥ãªãã£è£½åãéçºãã¦ããCylance社ã«ããããã¤ãã®ãã¼ã«ã®çºè¡¨ã
- NMAP Clustering
- 大éã®IPã«å¯¾ããNmapã®XMLåºåãk-Meansã§ã¯ã©ã¹ã¿ãªã³ã°ãWeb UIä¸ã§ã¤ã³ã¿ã©ã¯ãã£ãã«ã¯ã©ã¹ã¿ãåå²ã§ããã
- Botnet Panel Identification
- ããã¦ã§ããµã¤ãã«Botnet Panelãç½®ããã¦ãããã©ããã決å®æ¨ã®ã¢ã³ãµã³ãã«ã§èª¿ã¹ããChrome Extensionãç¡åã§å ¬éããã¦ããã
- Obfuscating Data with Markov Chains
- é©å½ãªæç« ãããã«ã³ããã§ã¼ã³ãä½ããæ°å¤ãé·ç§»ç¢ºçã®é ä½ã«å¯¾å¿ããã¦ã¨ã³ã³ã¼ããããã¨ã§ãã¼ã¿ãé£èªåããã
å ¨ä½çã«ãããã話ã§ã¯ãªãã®ã ããæå製åãã©ããã£ããã¨ããã£ã¦ããã®ãã®åèã«ã¯ãªãã
ææ
è³æã¸ã®ãªã³ã¯ããªãããåãä¸ããªãã£ãããAIã¨ã»ãã¥ãªãã£ã«é¢ããã¯ã¼ã¯ã·ã§ããã¨ãã¦AISec 2016ã¨ãããã®ãåå¨ããã ã¾ããDNNã¢ãã«ã®Adversarial Sample Craftingã«å¯¾ããrobustæ§ãæ¤è¨¼ããDeep-pwning (DEF CON 24)ã¨ãããã¬ã¼ã ã¯ã¼ã¯ãå ¬éããã¦ããã
é²å¾¡ææ³ã¸ã®å¿ç¨ã ãã§ã¯ãªãããã¤ãã¹ææ³ãæ©æ¢°å¦ç¿ãµã¼ãã¹ã«å¯¾ããæ»æã«ã¤ãã¦ãç 究ããã¦ããããããããã«ãã»ãã¥ãªãã£ãããã
é¢é£ãªã³ã¯
- Presentation of discovRE â woumn
- SandPrint: Fingerprinting Malware Sandboxes to Provide Intelligence for Sandbox Evasionã®ç´¹ä» - Twitterã«æ¸ããããªããã¨
- Adversarial machine learning - Wikipedia
- Adversarial exampleã«ã¤ã㦠- sotetsuk's tech blog
- Adversarial Examples ã試ãã¦ã¿ã - classic style
- é¡èªèã失æãããæè¡ã失æãããããã«ã½ã³ã«åå ãã - kivantiumæ´»åæ¥è¨
- TensorFlowã§é¡èå¥ã¢ãã«ã«æé©åããå ¥åç»åãçæãã - ãããã¼ãã¡ã¢
- SoK: Applying Machine Learning in Security - A Survey (arXiv preprint)