Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Intelligent Software Engineering: Synergy between AI and Software Engineering

1,355 views

Published on

Keynote speech at the 2017 International Workshop on Software Mining (http://lamda.nju.edu.cn/conf/softwaremining17/)

Published in: Software
  • Be the first to comment

Intelligent Software Engineering: Synergy between AI and Software Engineering

  1. 1. Intelligent Software Engineering: Synergy between AI and Software Engineering Tao Xie University of Illinois at Urbana-Champaign taoxie@illinois.edu http://taoxie.cs.illinois.edu/
  2. 2. Artificial Intelligence  Software Engineering Artificial Intelligence Software Engineering Intelligent Software Engineering Intelligent Software Engineering
  3. 3. 1st International Workshop on Intelligent Software Engineering (WISE 2017) Tao Xie University of Illinois at Urbana-Champaign, USA Abhik Roychoudhury National University of Singapore, Singapore Organizing Committee Wolfram Schulte Facebook, USA Qianxiang Wang Huawei, China Sponsor: Co-Located with ASE 2017 https://isofteng.github.io/wise2017/
  4. 4. Workshop Program 8 invited speakers 1 panel discussion https://isofteng.github.io/wise2017/ International Workshop on Intelligent Software Engineering (WISE 2017)
  5. 5. Artificial Intelligence  Software Engineering Artificial Intelligence Software Engineering Intelligent Software Engineering Intelligent Software Engineering
  6. 6. Past: Automated Software Testing • 10 years of collaboration with Microsoft Research on Pex • .NET Test Generation Tool based on Dynamic Symbolic Execution • Example Challenges • Path explosion [DSN’09: Fitnex] • Method sequence explosion [OOPSLA’12: Seeker] • Shipped in Visual Studio 2015/2017 Enterprise Edition • As IntelliTest • Code Hunt w/ > 6 million (6,114,978) users after 3.5 years • Including registered users playing on www.codehunt.com, anonymous users and accounts that access api.codehunt.com directly via the documented REST APIs) http://api.codehunt.com/ http://taoxie.cs.illinois.edu/publications/ase14-pexexperiences.pdf
  7. 7. Past: Android App Testing • 2 years of collaboration with Tencent Inc. WeChat testing team • Guided Random Test Generation Tool improved over Google Monkey • Resulting tool deployed in daily WeChat testing practice • WeChat = WhatsApp + Facebook + Instagram + PayPal + Uber … • #monthly active users: 963 millions @2017 2ndQ • Daily#: dozens of billion messages sent, hundreds of million photos uploaded, hundreds of million payment transactions executed • First studies on testing industrial Android apps [FSE’16IN][ICSE’17SEIP] • Beyond open source Android apps focused by academia WeChat http://taoxie.cs.illinois.edu/publications/esecfse17industry-replay.pdf http://taoxie.cs.illinois.edu/publications/fse16industry-wechat.pdf
  8. 8. Next: Intelligent Software Testing(?) • Learning from others working on the same things • Our work on mining API usage method sequences to test the API [ESEC/FSE’09: MSeqGen] http://taoxie.cs.illinois.edu/publications/esecfse09-mseqgen.pdf • Visser et al. Green: Reducing, reusing and recycling constraints in program analysis. FSE’12. • Learning from others working on similar things • Jia et al. Enhancing reuse of constraint solutions to improve symbolic execution.. ISSTA’15. • Aquino et al. Heuristically Matching Solution Spaces of Arithmetic Formulas to Efficiently Reuse Solutions. ICSE’17. [Jia et al. ISSTA’15]
  9. 9. Mining and Understanding Software Enclaves (MUSE) http://materials.dagstuhl.de/files/15/15472/15472.SureshJagannathan1.Slides.pdf DARPA
  10. 10. Pliny: Mining Big Code to help programmers (Rice U., UT Austin, Wisconsin, Grammatech) http://pliny.rice.edu/ http://news.rice.edu/2014/11/05/next-for-darpa-autocomplete-for-programmers-2/ $11 million (4 years)
  11. 11. Program Synthesis: NSF Expeditions in Computing https://excape.cis.upenn.edu/https://www.sciencedaily.com/releases/2016/08/160815134941.htm 10 millions (5 years)
  12. 12. Software related data are pervasive Runtime traces Program logs System events Perf counters … Usage log User surveys Online forum posts Blog & Twitter … Source code Bug history Check-in history Test cases Keystrokes …
  13. 13. Software Analytics Vertical Horizontal Information Visualization Data Analysis Algorithms Large-scale Computing Software analytics is to enable software practitioners to perform data exploration and analysis in order to obtain insightful and actionable information for data-driven tasks around software and services. In Collaboration with Microsoft Research Asia http://taoxie.cs.illinois.edu/publications/malets11-analytics.pdf
  14. 14. Past: Software Analytics • StackMine [ICSE’12]: performance debugging in the large • Data Source: Performance call stack traces from Windows end users • Analytics Output: Ranked clusters of call stack traces based on shared patterns • Impact: Deployed/used in daily practice of Windows Performance Analysis team • XIAO [ACSAC’12]: code-clone detection and search • Data Source: Source code repos (+ given code segment optionally) • Analytics Output: Code clones • Impact: Shipped in Visual Studio 2012; deployed/used in daily practice of Microsoft Security Response Center In Collaboration with Microsoft Research Asia Internet
  15. 15. Past: Software Analytics • Service Analysis Studio [ASE’13]: service incident management • Data Source: Transaction logs, system metrics, past incident reports • Analytics Output: Healing suggestions/likely root causes of the given incident • Impact: Deployed and used by an important Microsoft service (hundreds of millions of users) for incident management In Collaboration with Microsoft Research Asia
  16. 16. Next: Intelligent Software Analytics(?) Microsoft Research Asia - Software Analytics Group - Smart Data Discovery IN4: INteractive, Intuitive, Instant, INsights Quick Insights -> Microsoft Power BI Gartner Magic Quadrant for Business Intelligence & Analytics Platforms
  17. 17. Microsoft Research Asia - Software Analytics Group https://www.hksilicon.com/articles/1213020
  18. 18. Open Topics in Intelligent Software Engineering (ISE) • How to determine whether a software engineering tool is indeed “intelligent”? • Turing test for such tool? • What sub-areas/problems in ISE shall the research community invest efforts on as high priority? • How to turn ISE research results into industrial/open source practice? • …
  19. 19. Artificial Intelligence  Software Engineering Artificial Intelligence Software Engineering Intelligent Software Engineering Intelligent Software Engineering
  20. 20. White-House-Sponsored Workshop (2016 June 28) http://www.cmu.edu/safartint/
  21. 21. Self-Driving Tesla Involved in Fatal Crash (2016 June 30) http://www.nytimes.com/2016/07/01/business/self-driving-tesla-fatal-crash-investigation.html “A Tesla car in autopilot crashed into a trailer because the autopilot system failed to recognize the trailer as an obstacle due to its “white color against a brightly lit sky” and the “high ride height” http://www.cs.columbia.edu/~suman/docs/deepxplore.pdf
  22. 22. Microsoft's Teen Chatbot Tay Turned into Genocidal Racist (2016 March 23/24) http://www.businessinsider.com/ai-expert-explains-why-microsofts-tay-chatbot-is-so-racist-2016-3 "There are a number of precautionary steps they [Microsft] could have taken. It wouldn't have been too hard to create a blacklist of terms; or narrow the scope of replies. They could also have simply manually moderated Tay for the first few days, even if that had meant slower responses." “businesses and other AI developers will need to give more thought to the protocols they design for testing and training AIs like Tay.”
  23. 23. • An example • Declaration from the PuTTY Website for Google’s search result change • This declaration suggests that Google’s search results for “putty” at some time may not be satisfactory and may cause confusions of the users. Motivation: Testing Search Engines 23
  24. 24. Our Past Work on Testing Search Engines • Given a large set of search results, find the failing tests from them • Test oracles: relevance judgments • Test selection • Inspect parts of search results for some queries by mining search results of all queries 24 Queries and Search Results Feature Vectors (Itemsets) Association Rules Detecting Violations http://taoxie.cs.illinois.edu/publications/ase11-searchenginetest.pdf [ASE 2011]
  25. 25. • Query items • Query words, query types, query length, etc. • Search result items • Domain, domain’s Alexa rank, etc. • Query-result matching items • Whether the domain name has the query, whether the title has the query, etc. • Search engine items • Search engine names Mining Output Rules 25
  26. 26. • SE:bing, Q:boston colleges, QW:boston, QW:colleges, TwoWords, CommonQ, top10:searchboston.com, top1:searchboston.com, top10:en.wikipedia.org, …, SOMEGE100K, SOMELE1K • SE:bing, Q:day after day, QW:day, QW:after, ManyWords, CommonQ, top10:en.wikipedia.org, top1:en.wikipedia.org, top10:dayafterday.org, …, SOMEGE100K, SOMELE1K Example Itemsets 26
  27. 27. Mining Association Rules • Mining frequent itemsets with length constraint • An itemset is frequent if its support is larger than the min_support {SE:bing, top10:en.wikipedia.org} • Generating rules with only one item in the right hand side • For each item xi in Y, generate a rule Y-xi => xi SE:bing=>top10:en.wikipedia.org 27
  28. 28. Experiments • Search results • Use the Web services of Google and Bing to collect the top 10 search results of each query from December 25, 2010 to April 21, 2011 • 390797 ranked lists of search results (each list contains the top 10 search results) 28
  29. 29. The Mined Rules • Mining from one search engine' results in one day • Google's search results on Dec. 25, 2010 • minsup = 20, minconf = 0.95, and maxL = 3 29
  30. 30. The Mined Rules • Mining from multiple search engines' results in one day • Google and Bing's search results on Dec. 25, 2010 • minsup = 20, minconf = 0.95, and maxL = 3 • Rules 9-12 show the different opinions of search engines to certain Websites 30
  31. 31. The Mined Rules • Mining from one search engine' results in multiple days • Google's search results from December 25, 2010 to March 31, 2011. • minsup = 20, minconf = 0.95, and maxL = 2 • Rules 13-18 show the rules about the top 1 results for the queries 31
  32. 32. Example Violations • Search results of Bing on April 1st, 2011 violate the following rule • The actual result of Bing http://www.jcu.edu/index.php points to the homepage of the John Carroll University, not easy to get the answer of the query 32
  33. 33. Our Recent Work: Multiple-Implementation Testing of Machine Learning Software
  34. 34. Faults Detected for kNN and Naive Bayes Implementations
  35. 35. Detected Fault Types
  36. 36. Other’s Work at Columbia/Lehigh U.: SOSP 2017 Best Paper Award http://www.cs.columbia.edu/~suman/docs/deepxplore.pdf
  37. 37. Our Most Recent Work: Malware Detection in Adversarial Settings: Exploiting Feature Evolutions and Confusions in Android Apps WeiYang, Deguang Kong,Tao Xie and Carl A. Gunter Annual Computer Security Applications Conference (ACSAC 2017) http://taoxie.cs.illinois.edu/publications/acsac17-malware.pdf http://web.engr.illinois.edu/~weiyang3/ Wei Yang on the job market!
  38. 38. Machine learning (ML) in adversarial settings 38 • The development of machine learning is algorithm-oriented, often neglecting real-world constraints: machine budget, time cost, security, privacy etc. • The real world is complex, uncertain and adversarial. • Our work • Robustness of ML algorithms in security systems (well-known). • Practicability of previous proposed attacks on ML-based security system (something new). • We use mobile malware detection systems as an example for study. • Complexity (handled by logic), uncertainty (handled by probability) and, adversarial (focus of our work).
  39. 39. Robustness of ML algorithms in security systems 39 A Dataset Training Data Test Data Training Feature Extraction Feature Vector ML Algorithm Trained Classifier Result • Conventionally, to train an ML model, researchers separate training data and test data from the same dataset. • Overly-strong assumption: training data shares similar stationary distribution as test data (i.e., training data is representative).
  40. 40. Robustness of machine learning algorithms in security systems 40 A Dataset Training Data Test Data Training Feature Extraction Feature Vector ML Algorithm Trained Classifier Result • In adversarial settings, such assumption often does not hold • Training data and test data may not come from the same data set. • Poisoning attack: engineering on features or labels of training data • Compromise the discriminant classifier. • Evade attack: engineering on features of testing data • Change discriminant feature values to indiscriminant ones.
  41. 41. Focus: Evasion Attack on Mobile Malware Detection • Goals: Understand classifier robustness; Build better classifiers (or give up) • Existing work: 43
  42. 42. Existing work • Purely based on mathematics. • Making slight perturbations (e.g., 𝛿𝑥2) to incur dramatic change on output (e.g., 𝑂𝑢𝑡𝑝𝑢𝑡(𝑥1, 𝑥2)). • Example: • Derivative-based approach 44 Malicious Benign https://arxiv.org/abs/1511.07528
  43. 43. What is missing: Practicability • These proposed attacks manipulate the feature vectors of a malware in theory w/o considering the feasibility/impact of the mutations on the malware’s code. • Mutations may cause compilation errors. • Mutations may crash the app, cause undesirable behaviors, or disable malicious functionalities. 45
  44. 44. Three practical constraints to craft a realistic attack • Preserving Malicious Behaviors • Maintain the original malicious purposes. • Maintaining the Robustness of Apps • Robust enough to be installed and executed in mobile devices. • Evading Malware Detectors w/ little or no knowledge of the model • Identify and change feature values to evade detection. 46
  45. 45. Malware Recomposition Variation (MRV) • Malware Evolution Attack (existing malware evolution histories ) • Reapply feature mutations in malware evolution to create new malware variants. • Malware Confusion Attack (existing evasive malware) • Mutate a malware’s feature values to mimic classified benign apps’ to create new malware variants. • Insight: feature mutations are less likely to break the apps when the mutations follow feature patterns of existing malware. 47
  46. 46. Intuition: Evolving space of malware • Malware Evolution: make a feature vector Vi that is correctly classified as malware to be misclassified as benign apps by mutating Vi’s feature values. Three options: • Mutating Vi to be exactly same as a benign feature vector (e.g., B1) • Mutating Vi to be exactly same as a malware feature vector that is misclassified as benign (e.g., M2) • Mutating Vi to a new feature vector Vn that can be potentially classified as benign (e.g., Mv)
  47. 47. Malware Detection • A feature vector space V, • The feature vectors of existing benign apps B, • The feature vectors of existing malware M, • The feature vectors detected by detection model D, • The feature vectors of all potential malware M’. 𝑉 𝐵 𝑀 𝐷 𝑀′
  48. 48. Malware Detection If the feature selection is perfect (i.e., all differences between benign apps and malware are captured by selected features), no malware and benign apps should be projected to the same feature vector. 𝑉 𝐵 𝑀 𝐷 𝑀′
  49. 49. Insight 1: Differentiability of Feature Set Malware detection usually performs poorly in differentiating the malware and benign apps with the same feature vector. We try to convert malware with unique malicious features (i.e., M n (B M)) to possess the feature vectors shared by benign apps (i.e., BM)). 𝑉 𝐵 𝑀 𝐷 𝑀′
  50. 50. Attack 1: Malware confusion attack Mutate a malware’s feature values existing in only malware to the values existing in both malware and benign apps E.g., M1 -> M2
  51. 51. Insight 2: Robustness of detection model • Detection models are limited to detect existing malware. • A robust malware detection model should aim to detect malware variants produced through known transformation. The robustness of a detection model can be represented by 𝑉 𝐵 𝑀 𝐷 𝑀′
  52. 52. Attack 2: Malware evolution attack • Mutate malware feature values to the “blind spots” of malware detection. • Mimic feature mutations in the actual evolution of malware in the real world. • We conduct a phylogenetic analysis for each family of malware • E.g., Apply the mutation M1->M2 (e.g., c->c’ in 𝐹3) on M3 to derive 𝑀𝑣
  53. 53. Why MRV works • Large feature set has numerous non-essential features • non-informative or even misleading features. • Observation 1: Malware detectors often confuse non-essential features in code clones as discriminative features. • Learning-based detectors may regard non-essential features (e.g., minor implementation detail) in code clones as major discriminant factor • Observation 2: Using a universal set of features for all malware families would result in a large number of non-essential features for each family  opportunities to be mutated to evade detection while preserving malicious purposes • Features essential to malicious behaviors are different for each malware family 55
  54. 54. Overview 56
  55. 55. Feature Model • A substitute model • Summarize essential features (i.e., security-sensitive resources) and contextual features (e.g., when, where, how the security-sensitive resources are obtained and used) commonly used in malware detection • Transferability property • Some adversarial samples generated based on the substitute model can also evade the original detectors 57
  56. 56. Approach • Mutation strategy synthesis • Phylogenetic analysis for evolution attack • Similarity metric for confusion attack • Program mutation • Program transplantation/ refactoring 58
  57. 57. Practicability of attacks • Checking the preserving of malicious behaviors • Our impact analysis is based on the insight that the component-based nature of Android constrains the impact of mutations within certain components. • Checking the robustness of mutated apps • Each mutated app is tested against 5,000 events randomly generated by Monkey to ensure that the app does not crash. 59
  58. 58. Results 60
  59. 59. Strengthening the robustness of detection • Adversarial Training • Randomly choose half of our generated malware variants into the training set to train the model • Variant Detector • Create a new classifier called variant detector to detect whether an app is a variants derived from existing malware. • Weight Bounding • Constrain the weight on a few dominant features to make feature weights more evenly distributed. 61
  60. 60. Results 62
  61. 61. Artificial Intelligence  Software Engineering Artificial Intelligence Software Engineering Intelligent Software Engineering Intelligent Software Engineering

×