{"title":"Session details: Security in societal computing","authors":"P. Laskov","doi":"10.1145/3249989","DOIUrl":"https://doi.org/10.1145/3249989","url":null,"abstract":"","PeriodicalId":422398,"journal":{"name":"Proceedings of the 2013 ACM workshop on Artificial intelligence and security","volume":"68 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2013-11-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126012649","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
A. Leontjeva, M. Goldszmidt, Yinglian Xie, Fang Yu, M. Abadi
We investigate possible improvements in online fraud detection based on information about users and their interactions. We develop, apply, and evaluate our methods in the context of Skype. Specifically, in Skype, we aim to provide tools that identify fraudsters that have eluded the first line of detection systems and have been active for months. Our approach to automation is based on machine learning methods. We rely on a variety of features present in the data, including static user profiles (e.g., age), dynamic product usage (e.g., time series of calls), local social behavior (addition/deletion of friends), and global social features (e.g., PageRank). We introduce new techniques for pre-processing the dynamic (time series) features and fusing them with social features. We provide a thorough analysis of the usefulness of the different categories of features and of the effectiveness of our new techniques.
{"title":"Early security classification of skype users via machine learning","authors":"A. Leontjeva, M. Goldszmidt, Yinglian Xie, Fang Yu, M. Abadi","doi":"10.1145/2517312.2517322","DOIUrl":"https://doi.org/10.1145/2517312.2517322","url":null,"abstract":"We investigate possible improvements in online fraud detection based on information about users and their interactions. We develop, apply, and evaluate our methods in the context of Skype. Specifically, in Skype, we aim to provide tools that identify fraudsters that have eluded the first line of detection systems and have been active for months. Our approach to automation is based on machine learning methods. We rely on a variety of features present in the data, including static user profiles (e.g., age), dynamic product usage (e.g., time series of calls), local social behavior (addition/deletion of friends), and global social features (e.g., PageRank). We introduce new techniques for pre-processing the dynamic (time series) features and fusing them with social features. We provide a thorough analysis of the usefulness of the different categories of features and of the effectiveness of our new techniques.","PeriodicalId":422398,"journal":{"name":"Proceedings of the 2013 ACM workshop on Artificial intelligence and security","volume":"125 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2013-11-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133265391","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Session details: Keynote address","authors":"B. Nelson","doi":"10.1145/3249988","DOIUrl":"https://doi.org/10.1145/3249988","url":null,"abstract":"","PeriodicalId":422398,"journal":{"name":"Proceedings of the 2013 ACM workshop on Artificial intelligence and security","volume":"1844 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2013-11-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129833186","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
B. Biggio, I. Pillai, S. R. Bulò, Davide Ariu, M. Pelillo, F. Roli
Clustering algorithms have been increasingly adopted in security applications to spot dangerous or illicit activities. However, they have not been originally devised to deal with deliberate attack attempts that may aim to subvert the clustering process itself. Whether clustering can be safely adopted in such settings remains thus questionable. In this work we propose a general framework that allows one to identify potential attacks against clustering algorithms, and to evaluate their impact, by making specific assumptions on the adversary's goal, knowledge of the attacked system, and capabilities of manipulating the input data. We show that an attacker may significantly poison the whole clustering process by adding a relatively small percentage of attack samples to the input data, and that some attack samples may be obfuscated to be hidden within some existing clusters. We present a case study on single-linkage hierarchical clustering, and report experiments on clustering of malware samples and handwritten digits.
{"title":"Is data clustering in adversarial settings secure?","authors":"B. Biggio, I. Pillai, S. R. Bulò, Davide Ariu, M. Pelillo, F. Roli","doi":"10.1145/2517312.2517321","DOIUrl":"https://doi.org/10.1145/2517312.2517321","url":null,"abstract":"Clustering algorithms have been increasingly adopted in security applications to spot dangerous or illicit activities. However, they have not been originally devised to deal with deliberate attack attempts that may aim to subvert the clustering process itself. Whether clustering can be safely adopted in such settings remains thus questionable. In this work we propose a general framework that allows one to identify potential attacks against clustering algorithms, and to evaluate their impact, by making specific assumptions on the adversary's goal, knowledge of the attacked system, and capabilities of manipulating the input data. We show that an attacker may significantly poison the whole clustering process by adding a relatively small percentage of attack samples to the input data, and that some attack samples may be obfuscated to be hidden within some existing clusters. We present a case study on single-linkage hierarchical clustering, and report experiments on clustering of malware samples and handwritten digits.","PeriodicalId":422398,"journal":{"name":"Proceedings of the 2013 ACM workshop on Artificial intelligence and security","volume":"8 1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2013-11-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116358869","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
As the amount of content users publish on social networking sites rises, so do the danger and costs of inadvertently sharing content with an unintended audience. Studies repeatedly show that users frequently misconfigure their policies or misunderstand the privacy features offered by social networks. A way to mitigate these problems is to develop automated tools to assist users in correctly setting their policy. This paper explores the viability of one such approach: we examine the extent to which machine learning can be used to deduce users' sharing preferences for content posted on Facebook. To generate data on which to evaluate our approach, we conduct an online survey of Facebook users, gathering their Facebook posts and associated policies, as well as their intended privacy policy for a subset of the posts. We use this data to test the efficacy of several algorithms at predicting policies, and the effects on prediction accuracy of varying the features on which they base their predictions. We find that Facebook's default behavior of assigning to a new post the privacy settings of the preceding one correctly assigns policies for only 67% of posts. The best of the prediction algorithms we tested outperforms this baseline for 80% of participants, with an average accuracy of 81%; this equates to a 45% reduction in the number of posts with misconfigured policies. Further, for those participants (66%) whose implemented policy usually matched their intended policy, our approach predicts the correct privacy settings for 94% of posts.
{"title":"What you want is not what you get: predicting sharing policies for text-based content on facebook","authors":"Arunesh Sinha, Yan Li, Lujo Bauer","doi":"10.1145/2517312.2517317","DOIUrl":"https://doi.org/10.1145/2517312.2517317","url":null,"abstract":"As the amount of content users publish on social networking sites rises, so do the danger and costs of inadvertently sharing content with an unintended audience. Studies repeatedly show that users frequently misconfigure their policies or misunderstand the privacy features offered by social networks. A way to mitigate these problems is to develop automated tools to assist users in correctly setting their policy. This paper explores the viability of one such approach: we examine the extent to which machine learning can be used to deduce users' sharing preferences for content posted on Facebook. To generate data on which to evaluate our approach, we conduct an online survey of Facebook users, gathering their Facebook posts and associated policies, as well as their intended privacy policy for a subset of the posts. We use this data to test the efficacy of several algorithms at predicting policies, and the effects on prediction accuracy of varying the features on which they base their predictions. We find that Facebook's default behavior of assigning to a new post the privacy settings of the preceding one correctly assigns policies for only 67% of posts. The best of the prediction algorithms we tested outperforms this baseline for 80% of participants, with an average accuracy of 81%; this equates to a 45% reduction in the number of posts with misconfigured policies. Further, for those participants (66%) whose implemented policy usually matched their intended policy, our approach predicts the correct privacy settings for 94% of posts.","PeriodicalId":422398,"journal":{"name":"Proceedings of the 2013 ACM workshop on Artificial intelligence and security","volume":"7 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2013-11-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121533138","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Many social networks are predicated on the assumption that a member's online information reflects his or her real identity. In such networks, members who fill their name fields with fictitious identities, company names, phone numbers, or just gibberish are violating the terms of service, polluting search results, and degrading the value of the site to real members. Finding and removing these accounts on the basis of their spammy names can both improve the site experience for real members and prevent further abusive activity. In this paper we describe a set of features that can be used by a Naive Bayes classifier to find accounts whose names do not represent real people. The model can detect both automated and human abusers and can be used at registration time, before other signals such as social graph or clickstream history are present. We use member data from LinkedIn to train and validate our model and to choose parameters. Our best-scoring model achieves AUC 0.85 on a sequestered test set. We ran the algorithm on live LinkedIn data for one month in parallel with our previous name scoring algorithm based on regular expressions. The false positive rate of our new algorithm (3.3%) was less than half that of the previous algorithm (7.0%). When the algorithm is run on email usernames as well as user-entered first and last names, it provides an effective way to catch not only bad human actors but also bots that have poor name and email generation algorithms.
{"title":"Using naive bayes to detect spammy names in social networks","authors":"D. Freeman","doi":"10.1145/2517312.2517314","DOIUrl":"https://doi.org/10.1145/2517312.2517314","url":null,"abstract":"Many social networks are predicated on the assumption that a member's online information reflects his or her real identity. In such networks, members who fill their name fields with fictitious identities, company names, phone numbers, or just gibberish are violating the terms of service, polluting search results, and degrading the value of the site to real members. Finding and removing these accounts on the basis of their spammy names can both improve the site experience for real members and prevent further abusive activity. In this paper we describe a set of features that can be used by a Naive Bayes classifier to find accounts whose names do not represent real people. The model can detect both automated and human abusers and can be used at registration time, before other signals such as social graph or clickstream history are present. We use member data from LinkedIn to train and validate our model and to choose parameters. Our best-scoring model achieves AUC 0.85 on a sequestered test set. We ran the algorithm on live LinkedIn data for one month in parallel with our previous name scoring algorithm based on regular expressions. The false positive rate of our new algorithm (3.3%) was less than half that of the previous algorithm (7.0%). When the algorithm is run on email usernames as well as user-entered first and last names, it provides an effective way to catch not only bad human actors but also bots that have poor name and email generation algorithms.","PeriodicalId":422398,"journal":{"name":"Proceedings of the 2013 ACM workshop on Artificial intelligence and security","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2013-11-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126974955","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
An increasing number of machine learning applications involve detecting the malicious behavior of an attacker who wishes to avoid detection. In such domains, attackers modify their behavior to evade the classifier while accomplishing their goals as efficiently as possible. The attackers typically do not know the exact classifier parameters, but they may be able to evade it by observing the classifier's behavior on test instances that they construct. For example, spammers may learn the most effective ways to modify their spams by sending test emails to accounts they control. This problem setting has been formally analyzed for linear classifiers with discrete features and convex-inducing classifiers with continuous features, but never for non-linear classifiers with discrete features. In this paper, we extend previous ACRE learning results to convex polytopes representing unions or intersections of linear classifiers. We prove that exponentially many queries are required in the worst case, but that when the features used by the component classifiers are disjoint, previous attacks on linear classifiers can be adapted to efficiently attack them. In experiments, we further analyze the cost and number of queries required to attack different types of classifiers. These results move us closer to a comprehensive understanding of the relative vulnerability of different types of classifiers to malicious adversaries.
{"title":"On the hardness of evading combinations of linear classifiers","authors":"David Stevens, Daniel Lowd","doi":"10.1145/2517312.2517318","DOIUrl":"https://doi.org/10.1145/2517312.2517318","url":null,"abstract":"An increasing number of machine learning applications involve detecting the malicious behavior of an attacker who wishes to avoid detection. In such domains, attackers modify their behavior to evade the classifier while accomplishing their goals as efficiently as possible. The attackers typically do not know the exact classifier parameters, but they may be able to evade it by observing the classifier's behavior on test instances that they construct. For example, spammers may learn the most effective ways to modify their spams by sending test emails to accounts they control. This problem setting has been formally analyzed for linear classifiers with discrete features and convex-inducing classifiers with continuous features, but never for non-linear classifiers with discrete features. In this paper, we extend previous ACRE learning results to convex polytopes representing unions or intersections of linear classifiers. We prove that exponentially many queries are required in the worst case, but that when the features used by the component classifiers are disjoint, previous attacks on linear classifiers can be adapted to efficiently attack them. In experiments, we further analyze the cost and number of queries required to attack different types of classifiers. These results move us closer to a comprehensive understanding of the relative vulnerability of different types of classifiers to malicious adversaries.","PeriodicalId":422398,"journal":{"name":"Proceedings of the 2013 ACM workshop on Artificial intelligence and security","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2013-11-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129308659","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Session details: Adversarial learning","authors":"Christos Dimitrakakis","doi":"10.1145/3249991","DOIUrl":"https://doi.org/10.1145/3249991","url":null,"abstract":"","PeriodicalId":422398,"journal":{"name":"Proceedings of the 2013 ACM workshop on Artificial intelligence and security","volume":"75 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2013-11-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121063291","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Hugo Gascon, Fabian Yamaguchi, Dan Arp, Konrad Rieck
The number of malicious applications targeting the Android system has literally exploded in recent years. While the security community, well aware of this fact, has proposed several methods for detection of Android malware, most of these are based on permission and API usage or the identification of expert features. Unfortunately, many of these approaches are susceptible to instruction level obfuscation techniques. Previous research on classic desktop malware has shown that some high level characteristics of the code, such as function call graphs, can be used to find similarities between samples while being more robust against certain obfuscation strategies. However, the identification of similarities in graphs is a non-trivial problem whose complexity hinders the use of these features for malware detection. In this paper, we explore how recent developments in machine learning classification of graphs can be efficiently applied to this problem. We propose a method for malware detection based on efficient embeddings of function call graphs with an explicit feature map inspired by a linear-time graph kernel. In an evaluation with 12,158 malware samples our method, purely based on structural features, outperforms several related approaches and detects 89% of the malware with few false alarms, while also allowing to pin-point malicious code structures within Android applications.
{"title":"Structural detection of android malware using embedded call graphs","authors":"Hugo Gascon, Fabian Yamaguchi, Dan Arp, Konrad Rieck","doi":"10.1145/2517312.2517315","DOIUrl":"https://doi.org/10.1145/2517312.2517315","url":null,"abstract":"The number of malicious applications targeting the Android system has literally exploded in recent years. While the security community, well aware of this fact, has proposed several methods for detection of Android malware, most of these are based on permission and API usage or the identification of expert features. Unfortunately, many of these approaches are susceptible to instruction level obfuscation techniques. Previous research on classic desktop malware has shown that some high level characteristics of the code, such as function call graphs, can be used to find similarities between samples while being more robust against certain obfuscation strategies. However, the identification of similarities in graphs is a non-trivial problem whose complexity hinders the use of these features for malware detection. In this paper, we explore how recent developments in machine learning classification of graphs can be efficiently applied to this problem. We propose a method for malware detection based on efficient embeddings of function call graphs with an explicit feature map inspired by a linear-time graph kernel. In an evaluation with 12,158 malware samples our method, purely based on structural features, outperforms several related approaches and detects 89% of the malware with few false alarms, while also allowing to pin-point malicious code structures within Android applications.","PeriodicalId":422398,"journal":{"name":"Proceedings of the 2013 ACM workshop on Artificial intelligence and security","volume":"26 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2013-11-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129461507","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Machine learning has been widely used for defensive security. Numerous approaches have been devised that make use of learning techniques for detecting attacks and malicious software. By contrast, only very few research has studied how machine learning can be applied for offensive security. In this talk, we will explore this research direction and show how learning methods can be used for discovering vulnerabilities in software, finding information leaks in protected data, or supporting network reconnaissance. We discuss advantages and challenges of learning for offensive security as well as identify directions for future research.
{"title":"Off the beaten path: machine learning for offensive security","authors":"Konrad Rieck","doi":"10.1145/2517312.2517313","DOIUrl":"https://doi.org/10.1145/2517312.2517313","url":null,"abstract":"Machine learning has been widely used for defensive security. Numerous approaches have been devised that make use of learning techniques for detecting attacks and malicious software. By contrast, only very few research has studied how machine learning can be applied for offensive security. In this talk, we will explore this research direction and show how learning methods can be used for discovering vulnerabilities in software, finding information leaks in protected data, or supporting network reconnaissance. We discuss advantages and challenges of learning for offensive security as well as identify directions for future research.","PeriodicalId":422398,"journal":{"name":"Proceedings of the 2013 ACM workshop on Artificial intelligence and security","volume":"11 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2013-11-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"114292449","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：